DE3733674C2 - - Google Patents

Description

The invention relates to a speech analyzer. It is suitable for analyzing natural language and in particular in conjunction with an automatic Translation device can be used.

From the reference WICKERT, Klaus "Automatic Voice input and voice output ", Haar bei München, Verlag Mark & Technik 1983, pages 372-390, is already a Voice analyzer with a dictionary device known, which morphological data ent holds. In one embodiment of this known Voice analyzer arrives at a first analysis device Application to enter a morphological analysis of a a string (phrase) in a particular language perform. This first analysis device is a downstream of the second analysis device, in which a syntactic analysis of the entered string (word follow) on the basis of the analysis result of the first Analysis device is performed. In such a Speech analyzer need very large amounts of data in complex Data structures are transmitted and forwarded, so  that, among other things unwieldy big word trees arise and in addition also wrong decisions in the lower Ver work stages that can not occur afterwards can be corrected.

From the reference PAULUS, E .; WOLF, D; NTG Fachbe report 94 voice communication; VDE-Verlag GmbH, Berlin, Of fenbach; July 1986; Pages 209-214 (ISBN 3-8007-1465-5) a full synthesis system for the German language known in which a transcription program by incorporating a automatic morphological analysis can be realized can. This known system is based on a comprehensive Morpheme dictionary and a set of rules for structur description of German word forms. The resulting analysis The result of this is, on the one hand, the transcription in the narrower sense since knowing the morphological boundaries gives a direct, regular grapheme-phoneme conversion can be done and on the other hand the analysis result is the basis for a represents multiple syntactic and prosodic analysis. Dar In addition, word stress, speech rhythm and sentence me Thus, the quality of the product can be deduced decidedly improved synthetic language.

The object underlying the invention is to to provide a speech analyzer, where the amounts of data in conjunction with the word trees (possible word sequences with different probabilities for each branch) we can be considerably reduced, so that thereby the data Processing easier or to a lesser extent carried out can be led.

This object is achieved by the in claim 1 solved features.  

Particularly advantageous embodiments and developments The invention will become apparent from the dependent claims.

In the following the invention is based on execution Examples with reference to the drawing explained in more detail. Show it:  

Fig. 1 to 10 a first embodiment of a Sprachana analyzer with features of the invention, which is applied to an automatic translator English Japanese, showing in detail

Fig. 1 is a functional block diagram of an example of the detailed structure of a morphological analysis section;

Fig. 2 is a functional diagram of the entire structure;

Fig. 3 is an explanatory diagram showing an example of the structure of a dictionary file provided with a highest preference flag;

Fig. 4 is a flow chart of an example of a morphological analysis;

Fig. 5 is a flow chart of an example of the input processing in the morphological analysis;

Fig. 6 is an explanatory diagram of an example of forming an input character string;

Fig. 7 is an explanatory diagram showing an example of a dictionary retrieval;

Figs. 8A to 8D are flow charts, which will represent an example of how a contradiction at the highest Vorzugsflag in the morphological analysis eliminated;

Fig. 9 is an explanatory overview of an example of the contents of a retrieved dictionary information buffer after referencing the dictionary;

Fig. 10 is an explanatory overview of an example of the contents of the retrieved dictionary information buffer as a result of performing elimination of a conflict for the highest preference flag;

Fig. 11 dung a second embodiment having features to 16 after the OF INVENTION, in which show

Fig. 11 is a block diagram of the embodiment;

Fig. 12 is a table showing an example of data stored in a dictionary file;

Fig. 13 is a table showing an example of data stored in a dictionary file serving as a basic unit;

Fig. 14 is a table showing an example of data stored in a dictionary information preservation table;

Fig. 15 is a flow chart of the operation of this device;

Fig. 16 is a flow chart of the recognition unit;

Fig. 17 to 29 having features of the invention which is applied to an automatic transmission device English-Japanese, and show a third embodiment of a speech analyzer

Fig. 17 is a functional block diagram of an example of the detailed structure of a morphological analysis;

Fig. 18 is a functional block diagram of the entire structure;

Figs. 19A and 19B are flowcharts of an example of a morphological analysis;

Fig. 20 is a flow chart showing an example of a collective arrangement for a currency symbol and for the unit in the morphological analysis;

Figs. 21A and 21B are flowcharts of an example of hyphenated digits in the morphological analysis;

Figs. 22A and 22B are flowcharts of an example of processing successive numbers in the morphological analysis;

Figs. 23A and 23B are flowcharts of an example of a collective arrangement having a preceding numerical value in the morphological analysis;

Fig. 24 is an explanatory diagram showing an example of the structure of a dictionary file provided with a numerical flag;

Fig. 25 is an explanatory diagram of an example of an input character string;

Figs. 26A to 26D are explanatory diagrams showing the contents of the dictionary information preservation table found from the inputted character dictionary shown in Fig. 25 in response to the processing steps;

Fig. 27 is an explanatory diagram of another example of an input character string;

Fig. 28 is an explanatory diagram showing the contents of a currency symbol table of a position notation table and a decimal point table in the dictionary;

Figs. 29A to 29D are explanatory views showing an example of the dictionary information preservation table retrieved from the inputted character dictionary shown in Fig. 27 according to the processing steps;

Fig. 30 to 36 a fourth embodiment of a speech analyzer having features according to the invention, in which show

Fig. 30 is a block diagram of this embodiment;

Fig. 31 is a table showing an example of data stored in another dictionary;

Fig. 32 is a flowchart of the operation of the entire apparatus;

Fig. 33 is a flowchart of processing for the dictionary-registered proper name;

Fig. 34 is a flowchart of processing for a proper name not registered in the dictionary;

Fig. 35 is a flowchart of defective feature information processing;

Fig. 36 is a table showing an example in which data stored in the dictionary information preservation table is changed after the processing for the input sentence;

Figs. 37 to 46 show a fifth embodiment of a speech analyzer having features of the invention applied to an English-Japanese automatic translation device, wherein Figs

Fig. 37 is a functional block diagram of an example of a detailed construction of the morphological analysis section;

Fig. 38 is a functional block diagram of the entire construction;

Fig. 39 is an explanatory diagram of an embodiment for constructing a dictionary file;

Fig. 40 is a flow chart of an example of a morphological analysis for a proper name;

Fig. 41 is a flow chart of an example of the collective name system of proper names registered in a dictionary for morphological analysis;

Figs. 42 to 44 are flowcharts showing an example of processing depending on the position information in the morphological analysis of a proper name;

Fig. 45 is a flow chart of an example of the collective arrangement of proper names not registered in a dictionary in the morphological analysis;

Figs. 46A to 46F are explanatory diagrams showing the contents of the dictionary information preservation table referred to in the dictionary, for example, for an input character string according to the processing steps;

Figs. 47 to 52 show a sixth embodiment of a speech analyzer with features according to the invention, wherein show

Fig. 47 is a block diagram of this embodiment;

Fig. 48 is a table of, for example, data stored in a reference dictionary;

Fig. 49 is a flowchart of the operation of the entire apparatus;

Fig. 50 is a flowchart of the processing for the proper names stored in the dictionary;

Fig. 51 is a flow chart of processing for proper names not registered in the dictionary;

Fig. 52 is a table in which, for example, data stored in a dictionary information preservation table is changed after the processing of the input sentence;

Figs. 53 to 57 show a seventh embodiment of a speech analyzer having features of the invention applied to an English to Japanese automatic translation device, wherein Figs

Fig. 53 is a functional block diagram of one embodiment of the detailed structure of the morphological analysis section;

Fig. 54 is a functional block diagram of the entire structure;

FIGS . 55A and 55B are flowcharts showing an example of the morphological analysis;

Fig. 56 is an explanatory diagram showing an example of the content of the information table in an information processing section;

Fig. 57 is an explanatory overview of an example of the contents of a fitting table (7128);

Figs. 58 to 63 show an eighth embodiment of a speech analyzer having features according to the invention, wherein Figs

Fig. 58 is a block diagram for explaining the entire construction;

Fig. 59 is a block diagram for explaining an example of processing of derived words by means of a prefix;

Fig. 60 is a block diagram for explaining an example of processing of derived words by means of a suffix;

Fig. 61 is a block diagram of the entire details obtained by combining Figs. 58 to 60;

Fig. 62 is a block diagram showing further details of a complete unregistered word processing section in Fig. 61;

Fig. 63 is a block diagram for explaining an embodiment of an automatic translation device in which the invention is applied;

Figs. 64 to 90 show a ninth embodiment of a speech analyzer having features of the invention applied to an English-Japanese translator, wherein Figs

Fig. 64 is a functional block diagram of the entire construction;

Fig. 65 is a functional block diagram summarizing the function of recognizing the structural arrangement of an input English sentence as a block;

Fig. 66 is a flow chart of an example of a flow for collective arrangement of a block with respect to the input sentence;

Fig. 67 is a flowchart of details for word processing in the processing flow;

Fig. 68 is an explanatory overview of an example of dictionary information for English words or phrases stored in a word dictionary;

Fig. 69 is an explanatory overview of an example of the table data for the block initial state final state and for a purpose and roll evaluation state stored in an analysis file;

Fig. 70 is an explanatory diagram of an example of a collective arrangement of a structure;

Fig. 71 is an explanatory diagram of an example of a collective arrangement for a block;

Fig. 72 is an explanatory chart for an example of English information and word information collectively arranged in a block;

Fig. 73 is a flowchart showing an example of analysis processing performed in a corresponding analysis section;

Fig. 74 is an explanatory chart similar to that in Fig. 68, showing an example of access to a word phrase dictionary in the case that this embodiment has a function of an identical case score;

Fig. 75 is an explanatory chart similar to that in Fig. 69, showing an example of the start and end state of a block and a table of block preparation information in a case that this embodiment has a function of an identical case evaluation;

Fig. 76 is a functional block diagram similar to that in Fig. 64, illustrating the overall construction of a modification of this embodiment;

Fig. 77 is a functional block diagram, similar to that in Fig. 76, for summarizing the function of grammatical analysis of (let) information in view of the modified embodiment shown in Fig. 76;

Fig. 78 is an explanatory diagram of an example of dictionary information including the English words and phrases (let) information stored in the word memory in the modified embodiment;

Figs. 79 and 80 are explanatory diagrams similar to those in Fig. 72 and illustrating an example of the block and word information, wherein an English sentence containing (let) information is collectively arranged in a block;

Fig. 81 is a flowchart, similar to that in Fig. 73, illustrating an example of a flow for collective arrangement of the (let) information with respect to the input English sentence;

Fig. 82 is a flowchart, similar to that in Fig. 73, illustrating an example of the analysis processing including (let) information performed in the analysis section in the modified embodiment;

Figs. 83A and 83B are flowcharts showing an example of the flow for grammatically analyzing the (let) information for the input English sentence;

Fig. 84 is a flow chart of an example of the flow of the parsing grammatical analysis for the input English sentence;

Fig. 85 is an explanatory chart similar to that in Fig. 72, illustrating an example of the block and word information collectively arranged for the input English sentence containing a hyphenated word in a block;

Fig. 86 is a functional block diagram similar to that in Fig. 64 and showing the entire structure of another modified embodiment;

Fig. 87 is a functional block diagram similar to that in Fig. 65, in which a function of a morphological analysis of the supplementary question in the input English sentence is collectively arranged in the modified embodiment shown in Fig. 86;

Figs. 88 and 89 are explanatory views similar to those in Fig. 72 and illustrating an example of collectively arranged block and word information for an English sentence including a supplemental question in a block; and

Figs. 90A and 90B are flowcharts showing an example of a sequence of analysis of a supplementary question for the input English sentence.

Now, the first embodiment will be described with features according to the invention be. In Fig. 1, the entire structure of the first embodiment is shown, in which a speech analyzer is applied in an English-Japanese automatic translation device. Of course, the invention can be applied equally effectively not only to an automatic translator for translating English into Japanese, but also to any language analyzer in which the sentences of an input speech are mainly analyzed to translate one language into another.

The embodiment in Fig. 1 has an input section 1010 through which an English text 1012 to be translated into Japanese is input. The input section 1010 may comprise , for example, a keyboard with character keys such as alphanumeric keys or function keys, an optical character reader (OCR) for reading the English text recorded on paper, and / or a file memory to read the English text which is printed on a paper Storage medium as a magnetic disk is recorded. The English text input at the input section 1010 is read into a pre-editing section 1014 in which pretreatment for the translation is performed. This section mainly performs sentence recognition and unknown word processing. This then acts as part of a morphological analysis.

The pre-processed English data is transmitted to a first or morphological analysis section 1016 together with information contained in the pre-editing. The section 1016 divides the sentence by referring to a word dictionary 1018 , analyzes the morphemes of the English sentence, performs various types of arrangement such as unknown word processing, expression for time, expression for numbers, etc., and performs processing for the word Throughout the sentence as an additional survey and an identical case detection. The morphological analysis rule is stored in a rule file 1036 .

The English data after a morpheme analysis is transmitted to a second parsing section I 1020 together with the dictionary information obtained by the morphological analysis. (Parsing is understood here to mean a grammatical analysis or an automatic syntax analysis). Section I 1020 is a functional section that performs an analysis on the surface structure of an English sentence by applying a grammar rule to the English data, and then finds out all the structural possibilities.

The English data after the analysis in the section I 1020 are fed together with its analysis information to another analysis section II 1022 . In this further section, a solution is selected by applying a structure description based on the result of the analysis with respect to the surface structure by the section I. An acceptable "parse tree" for the English sentence is thus prepared to form its structure. These parsing rules are stored in a parsing rule file 1036 .

The English data is transmitted to a texture transform section 1034 as parsing tree data after analysis. In section 1034 , a corresponding Japanese sentence tree is prepared from a structural tree, ie an intermediate structure of the English sentence, and is translated into a Japanese underlying structure from which Japanese can then be easily translated.

The structural tree data showing the Japanese underlying structure which has undergone structural conversion in this manner is delivered to a translation section 1026 in which the translated sentence is formed. This is a functional section to create a Japanese sentence from the tree structure of the Japanese tree structure.

The Japanese sentence data translated in this way, ie, translated sentence data, is then output to a post-editing section 1030 . The section 1030 modifies the translated sentence data with reference to the dictionary 10128 , using information that has been used in the translation to complete a more natural Japanese sentence. The Japanese sentence data is then transmitted to an output section 1032 and then output as the translated Japanese sentence 1034 from the output section 1032 . The output section 1032 includes, for example, a printer, a display, and / or a file storage device such as a magnetic disk.

The flow for the subsequent translation operations is controlled by a control section 1038 which controls the entire facility. The word dictionary 1018 stores dictionary data for the words of English and Japanese languages, specifying not only the vocabulary but also various information such as a link relationship, ie, a concurrent relationship, meanings, plural and singular forms, language parts, etc. in this embodiment become. Further, the file 1036 stores the morphological and syntactic analysis rules data.

The control section 1038 is connected to an operation display section 1040 . The section 1040 has operation buttons which give various commands from an operator to the device, such as translation command keys or cursor keys, a display or display which visually displays an input English text, a Japanese sentence as a result of the translation, intermediate data such as a Display dictionary information, various commands to the operator, etc. Most of the operation display functions are arranged to be contained in a keyboard when it is disposed on the input section 110 or in a display, if provided at the output section 1032 .

FIG. 1 shows the detailed structure of the morphological analysis section 1016 . The section 1016 has an input unit 1100 , namely a keyboard for the input section 1010 and an input interface 1104 , which interfaces with the input template file 1102 . The input interface 1104 is provided with a buffer for inputted character strings inputted together with the English character string data in the form of code data such as ASCII from the input unit 1100 or the file 1102 , and temporarily stores the character string data. The input character string may be the one pre-edited in the portion 1014 .

As shown in FIG. 1, the first morphological analysis section 1016 includes a processing section 1106 , a dictionary reference section 1108 , a discard removal processing section 1110, and a control section 1112 . The processing section 1106 is a parsing function section for performing the morphological analysis and has a buffer for retrieved dictionary information, ie, a dictionary information preservation table 1110 (see FIG. 9). The morphological analysis is performed by properly instructing the dictionary retrieval from the upper end of the input character string corresponding to the interrogation key string and that the dictionary information obtained from the dictionary interrogation portion 1108 is written into the buffer 1120 is stored for retrieved dictionary information, and that the processing of a preferential degree is performed according to the highest preference flag, as will be described later.

The dictionary query section 1108 is a function section for taking out the dictionary information by querying the word dictionary 1018 based on the query key string, which has been commanded by the processing section 1106 and then transmitting it to the processing section 1106 .

The word dictionary 1018 stores grammar information such as a part of speech and a diffraction for access to each of the words as well as a highest preference flag as shown in Fig. 3 for the example of access information. The dictionary is intended as a dictionary file with a highest preference flag. "The highest preferred flag" is a flag indicating the amount of coupling between words contained in a compound word or phrase representing the dictionary access in which "0" indicates a weak coupling or no coupling, while "1" indicates "indicates a strong coupling. In this case, the linguistic usage is evaluated as a phrase for a compound word or phrase which, according to the judgment, has a strong coupling; On the other hand, the possibility of use in the form of individual words is also considered in parallel.

As illustrated in Figure 3, each of the entries into the word dictionary 1018 is possible for the compound word, phrase, and individual words that make up it, with no distinction made between the individual words and the compound word or phrase is. Furthermore, each diffraction form represents one access. If there are a number of diffraction forms, they are each registered as different inputs. The type of diffraction is displayed in the diffraction section. The situation is similar to the speech part in which the registration is permitted for a number of speech parts, and a speech part information is included for each of them. As further information, calculability or non-predictability for a noun, a transitive or intransitive verb or a translated word, etc. are registered.

For example, "get" is an infinitive form of a verb, and the highest preference flag is "0". The phrase "get up" is a phrase for an infinitive form, and its highest preference flag is "1". Further, a preposition group "up to" has the highest reference flag "1", but a phrase such as "white house" as the compound word has the highest preference flag of "0", and thus the latter shows that the degree of coupling between the words are low. In Fig. 3, the symbol "" indicates a space.

In this way, the dictionary information which has been retrieved in the query section 1108 contains the highest preference flag. If "1" is set for the highest preference flag for identical character strings or overlapping character series, such a contradiction must be eliminated. In the section 1110 , the objection elimination is performed and the subsequent processing, referring to the objection elimination rule, the highest preference flag stored in the file 1036 .

The Opposition Elimination Rule is in the present Embodiment in the following order (1) to (3), with a preferred selection performed becomes.

• (1) phrase or word whose part of speech is a verb;
• (2) composite word, phrase or word with many Word components;
• (3) compound word, phrase or word used in the front part located in the sentence.

The use for the thus-selected word, ie the parsing unit, is represented as the active information at the dictionary information retrieval buffer 1120 in the processing section 1016 . The active information indicates that the parsing unit is valid or effective if it is "1" while showing that its possibility is not needed if it is "0".

The control section 1112 is a function section for controlling and controlling the operation and the processing in each of the functional sections in the morphological analysis section 1016 . The section may be included in the control section 1038 from which the control for the entire device is performed. The result of the morphological analysis is transmitted via an output interface 1114 to the parsing section I 1020 . In the event that the result is not transmitted directly to the parsing section I 1020 , it is once stored in the parsing input file 1116 and in the parsing dictionary information file 1118 .

In this embodiment, while all the words, compound words and phrases starting from the branching position of the dictionary reference unit are taken out in a morphological analysis, dictionary information obtained for individual words containing the compound word or phrase form the phrase that has been judged to be a collective entity according to the highest preference flag. That is, the amount of coupling between the words in the sentence is judged while referring to the highest dictionary information reference flag obtained in the morphological analysis. The compound words or phrases that have been judged as having strong coupling are then judged to be used as a phrase in the sentence; if not, the possibility for use in the form of single words is considered in parallel. Such processing by the highest preference flag is performed by the sequence shown in FIG . Data for the input character strings are picked up by the input section 1010 ( 1200 ); the input character string is provided with a dictionary reference unit for retrieval in the dictionary file 1018 having the highest preference flag ( 1201 ); the dictionary 1018 is then interrogated again ( 1203 ) until the end position of the data string represented by the data for the input character string ( 1202 ); then the contradiction for the highest preference flag is eliminated ( 1204 ) and the result of the morphological analysis is given to the parsing section I ( 1205 ).

In the input processing ( 1200 ), the data is first read from the file 1102 or an input unit 1100 into the input character set buffer 1104 (see Fig. 5: 1210 ). For example, the data for the input character string is input in the form of ASCII; When data in the file is completely read out (for example, when the symbol EOF is read out), the processing section 1106 writes a NULL code in the buffer for the input character string as the ending position.

The processing section 1106 then reforms the input character string ( 1211 ). For example, if two or more characters belonging to a character corresponding to a space are continued, they will be corrected in a single space. The space-corresponding character includes spaces (represented by the symbol ""), tabbing, line-back (represented by the symbol ↙), etc. These space-corresponding characters between the upper and the first appearing characters except the space corresponding one Characters in the buffer for the input string are removed.

For example, the input string or arrangement
"I willgetup↙togotoawhite-
house ... ", as shown in Fig. 6, in
"Iwillgetuptogotoa whitehouse ...
(NULL) "The position of the" NULL "symbol indicates the end or end position of the memory.

The dictionary reference boundaries used for the removed dictionary processing unit processing 1201 are substituted for an alphabetic character, a numeric character, an apostrophe, and other characters except hyphen and paragraph, and an apostrophe followed by spaces, arranged. The processing section 1106 has an upper pointer for a dictionary reference, which is first set at the top of the buffer.

The query section 1108 queries the dictionary file 1018 provided with the highest preference flag, the character string being used from the character indicated by the upper pointer to the character preceding the next delimiter as the query key string. The dictionary access and the query key string are compared; if both are identical, the dictionary information is admitted ( 1203 ). The array is judged when the entire string of access matches with at least a part of the string starting from the top and when the part immediately after that part is a dictionary reference, an apostrophe or a paragraph. For example, as shown in Fig. 7, if the upper pointer indicates the upper character "g" in the query key string, "getupon" of the dictionary input is coincident thereto.

The retrieved dictionary information is stored in the buffer 1120 of the processing section 1106 . Along with the reading, the matching start position and end position of the character string are then stored. This specifies the position of the characters in the input buffer properly from the upper end. An accumulation area for the active information is arranged in the retrieved dictionary information buffer 1120 , which is information indicating whether the retrieved dictionary information is effective for the following processing or not, in which step all "1" is set.

Subsequently, the upper pointer is updated with respect to the respective dictionary reference, and is set to the character immediately after the delimitation which appears closest to the existing upper pointer indicating the character string from left to right. The dictionary reference is then performed. In the example mentioned above, the character at the top of the dictionary reference is displayed first as "I" for "I", then as "w" for "will" and then for "g" for "get". When the upper pointer goes through the NULL code, it is judged if this is the end position ( 1202 ).

Fig. 9 shows an example of the dictionary information which has thus been found for the example of the inputted English character string described above.

Now, in conjunction with FIGS. 8A to 8D, the contradiction elimination processing 1204 performed by the corresponding section 1110 will be described. The flow chart shown in Figs. 8A and 8B represents the processing in the case where the positions of the words where the highest reference flags are set overlap each other, while the flow chart in Figs. 8C and 8D shows the processing to parse - Unit, ie elements with the highest preference flag to remove, ie a processing to make the active information to "0". In these flowcharts, the notation "≦" represents a substitution, the character "→" represents an indication, and "P → x" represents the content of x taken by the input of the pointer p.

First, a set of words each having the highest reference flag "1" and the positions in the sentence overlapping each other are detected (steps 1220 to 1223 ). Then, the elimination rule regarding the highest reference flag is applied to each of the detected sentences, and those in which it is effective are selected (steps 1224 to 1235 ).

In the embodiment described above, the highest preference flag "1" is at "start up" at the start position "8" and at the end position "13" and at "up to" at the initial position "12" and the end position "16" for the string
"getupto" as shown in Fig. 9, and the positions for the characters then overlap each other. Then, the rule (1) mentioned above is first applied, and it is judged whether it is a verb or not, referring to the speech part of the preservation pointer psave and the speech part of the pointer p ( 1224 ). Since it corresponds to a verb in this example, the combination "get up" is selected.

If rule (1) is not satisfied, rule (2) is applied ( 1228 ) and the length for the character array, which refers to the input of the preservation pointer psave and the length for the character array, which depends on the input of the pointer p refers are compared. If the rule (2) is also not enough, then the rule (3) is applied ( 1229 ) and the position start referring to the start position for the preservation pointer psave and the position start referring to the start position of the pointer p, are compared. Then, if one of the rules (1) to (3) is satisfied when applied in this order, the active information is made "not satisfied", ie, a non-effective input is made "NULL" ( 1232 ). While otherwise, ie the active information is different, ie an effective input is left at "1" ( 1231 ). Such an application of the contradiction-eliminating rule is performed one by one, progressing with the pointer p stepwise ( 1234, 1235 ) to the end position for each of the entries, and the active information is made only for the effective input "1". The state of the aforementioned example is shown in FIG . For example, for the entry "up to" their active information is made "0".

The overlapping, even partially overlapping positions with the combination in which both the active information and the highest preference flag are "1" are determined ( 1236 to 1241 ) and their active information is made "0" ( 1242, 1249 ) , The application of such a contradiction eliminating rule is properly performed for each of the entries, with the pointer p advancing ( 1243, 1248 ) to the end position and the active information of a non-effective entry being made "0". Consequently, the active information for "get" and for "up" is made "0", for example for the entry "get up" ( Figure 10). For the entry "white", "white house" and "house", since all have the highest preference flag "0", even if the positions overlap, their active information is kept at "1".

In this way, when the processing has been performed until just before the end position (NULL), the content of the input buffer of the input interface 1104 and the retrieved dictionary information buffer 1120 from the output interface through interface 1114 to the parsing section I 1016 delivered. The content of the buffer 1120 is delivered only for the entry for which "1" is displayed as the active information. For example, the contents of the input buffer may be written to the parsing input file 1116 , while the content of the information buffer 1120 may be written to the parsing dictionary information file 1118 . In this case, since both the active information and the highest preferred flag are output, the structure of the information file 1118 is identical to that of the information buffer 1120 . However, it may also be so agreed that the active information and the highest preferred flag are not delivered.

The invention will now be described with reference to a second embodiment of a speech analyzer shown in Fig. 11 which is applied to an English-Japanese automatic translation device. This embodiment has an input section 2014 to which data is input from an input device 2010 or a template file 2012 . For example, the input device 2010 has a keyboard with character keys such as alpha / numeric or function keys and an optical character reader to read an English-written text recorded on paper. The template file 2012 is a storage device in which the English text is recorded on a storage medium such as a magnetic disk.

The input section 2014 includes a buffer 2014 a is an input character string and stores the inputted English sentence, which from the input device 2010 or the template file 2012 has been entered into the buffer 2014 a. The input section 2014 reads out the input sentence stored in the buffer 2014 a and delivers it to a processing section 2016 .

The processing section 2016 is a function section that performs the morphological analysis for the input sentence that has been issued from the input section 2014 by querying a dictionary file. The processing section 2016 includes a dictionary information preservation table 2016 a and stores the information which has been retrieving a dictionary file 2022 or a basic unit dictionary file 2026, which will be described later is obtained, in the table 2016 a.

The processing section 2016 retrieves a query key string as a unit in case that the dictionary is retrieved from the character string representing the input sentence input from the input section 2014 . The query string is properly retrieved from the first character of the string representing the input sentence according to a predetermined query rule. For example, the input sentence from the upper end of the sentence is properly divided by boundaries such as spaces, commas, etc., and the divided character series are used as the query key string. In this case, character strings expressing units such as m, k, m / s are formed as interrogation key character strings. The processing section 216 sends the interrogation key string which has been interrogated from the character string representing the input sentence to the dictionary interrogation section 2020 .

The dictionary inquiry section 2020 retrieves a dictionary file 2022 on the basis of the query character strings which have been output from the processing section 2016 . In the dictionary file 2022 , entry and grammatical information such as a voice part is stored as shown in FIG . If there is an entry in the dictionary file 2022 , the dictionary query section 2020 reads out the speech part information, etc. of that entry and delivers it to the processing section 2016 . If there is no entry in the dictionary file 2022 as a result of the query of the file 2022 , the dictionary query section 2022 returns this situation to the processing section 2016 .

The processing section 2016 stores the language part information, etc., which has been requested by the section 2020 into a dictionary information preservation table 216 a. If there is no entry for the query string in the file 2022 , the processing section 2016 outputs the query string to a unit recognition section 2024 .

The section 2024 retrieves a base unit dictionary file 2026 on the basis of the query character string which has been output from the processing section 2016 . The base unit entries are stored in file 2026 , as shown in FIG . If the primitive entry is present in the file 2026 , the unit detection section 2024 reads out the primitive entry. If there is no entry in the file 2026 , the interrogation key string is divided into a number of character strings, which will be described later, and the file 2026 is interrogated a plurality of times. If primitive entries occur on multiple queries of file 2026 , a number of unit information is obtained from the primitive entries. If no primitive entry occurs on any of the many queries, information indicating that it is not registered in the dictionary is obtained.

The unit recognizing section 2024 outputs the basic unit entry, composite unit information and information indicating that the word is not registered in the memory to the processing section 2016 . This ( 2016 ) stores this information input from the section 2024 in the table 2016 a. The table 2016 a stores the input for the query key character strings and grammar information, such as a part of speech in order to conserve this which the retrieval key character row has been obtained with respect to by querying the file 2022 or the file 2026th After the data has been stored in the table 2016 a, the processing section 2016 analyzes this data together with the input set to the output interface 2018th The interface 2018 outputs the input sentence and the morphological analysis data output from the processing section 2016 to an output unit 2030 such as a printer or a display unit or a memory file 2032 such as a magnetic disk.

On the other hand, it is also possible to input the input sentence and the morphological analysis data which have been output from the processing section 2016 directly into a parsing means (not shown) for parsing the input sentence in the parsing means and further to prepare a translated sentence based on the syntax analysis. The control section 2028 serves to control the operation of each of the functional sections in the illustrated device and may advantageously be implemented in the form of a microprocessor.

The operation of the illustrated device will now be explained with reference to the flowchart shown in FIG . First, the input English sentence is read from the input unit 2010 or from a template file 2012 into the input section 2014 ( 2100 ). The read-in the input section 2014 set is stored in the buffer 2014 for a an input character string. The sentence in the buffer 2014 a is then read out and delivered to the processing section 2016 . When the sentence is entered into the processing section 2016 , the dictionary query unit is excluded ( 2102 ). That is, the string (word string) representing the input sentence is sequenced according to a predetermined rule into a query key string as the units for retrieving the dictionary file 2022 or the constituent dictionary dictionary file 2026 , starting from the upper end of the character string divided up. Then, it is judged whether the split query string series exists or not ( 2104 ); if present, the query string is submitted to the dictionary query section 2020 .

When the query string is submitted to section 2020 , it ( 2020 ) queries file 2022 for the query key string ( 2106 ). It is then judged whether or not this character string is present in the entry of the dictionary file 2022 as shown in Fig. 12; if the entry is present, the grammar information such as a speech item that is stored in the file 2022 is read, and the read data is sent to the processing section 2016 and stored in the table 2016 a (2110). Then, step 2102 is returned, and the dictionary retrieval unit is turned off again.

If there is no entry in the file 2022 , the section 2020 returns the query key string to the processing section 2016 . The section 2016 then outputs the interrogation key string to the recognition section 2024 in which the unit recognition is performed ( 2112 ).

In the event that the query string from the section 2020 has ordinary words such as nouns and verbs, since they are mostly the entries in the dictionary file 2022 , grammar information, such as a speech part, is read from the file 2022 , and the data will be submitted to processing section 2016 and recorded in Table 2016 a. As described above, entries for ordinary words such as main words and verbs are formed, but no entries for the character string expressing the units in the file 2022 are formed. Thus, in the event that the query key string is a character string expressing a unit such as kg or m / s, since this is not an entry in file 2022 , in the process, step 2112 for one unit is proceeded.

The unit recognition operation in step 2112 will be explained with reference to FIG . When the interrogation key string for which there is no entry in the file 2022 in the interrogation is delivered from the processing section 2016 to the section 2024 , the pointer P becomes the character at the upper end of the interrogation key string in the section 2024 set ( 2200 ).

The section 2024 then queries the base unit dictionary file 2026 for the character string starting from the character at which the pointer P is set ( 2201 ).

In this polling, it is then judged whether the basic unit for which the entry exists in the file 2026 appears as a complete character string in the character string starting from the character at which the pointer P is set, and whether it is started from the character at which the pointer P is set or not. Namely, it is inquired whether or not a character string having one or a number of characters, starting at the character at which the pointer is set, coincides with any one of the constituent units for which the item is present in the file 2026 , For example, in the case where the characters in which the pointer P is set are k, m, s, etc., there are entries in the file 2026 for these individual characters beginning at the character at which the pointer P is set, as shown in Fig. 13.

The unit recognizing section 2024 judges whether the entries in the file 2026 are available as a result of the query to the file in 2026 or not (2204). If the entry exists, the pointer P advances by the length of the recognized primitive ( 2208 ). Consequently, in the case where the basic unit is k, m, s, etc., the pointer is advanced by one character and then set at the next character in the inquiry key string.

The section 2024 then judges whether or not the character string starting from the character at which the pointer P is set exists ( 2208 ). In the event that such a character string still exists, it returns to step 2202 where the file 2026 is again retrieved for the character string starting from the character to which the pointer P is set. Then, the section judges whether or not the entry in the basic unit exists due to the retrieval of the file 2026 ; if the entry is present, the pointer P advances by the length of the detected basic unit.

If, at step 2208, the character string starting from the character at which the pointer P is set is no longer present, the fetch to the file 2026 has been completed, ie, the recognition for the composite entity has been successful.

For example, in the case where the fetch key string outputted to the section 2024 is km / s, which is a unit, the entry in the file 2026 is absent because km / s is a complicated unit per se , The pointer P is then first set to k ( 2200 ), and k is retrieved from the base unit dictionary file 2026 to confirm the presence of the entry ( 2202 ). Then the pointer P is set to m ( 2206 ) and m is retrieved from the file 2026 ( 2202 ) to confirm the presence of the entry in the same way. Since the unit recognition section 2024 considers a slash (/), a solid circle (○), etc. as a part of a unit, the pointer P is set to s next, skipping "/" in km / s ( 2206 ) , Then s is retrieved from the file 2026 to confirm the presence of the entry in the same manner ( 2202 ). Since this was present for each of the units k, m, and s in retrieving in the file 2026 , it is now judged that km / s is a character string expressing a unit. In this way, in the event that entries are present in the file 2026 for all the characters forming the fetch key string, or in the event that entries in the file 2026 are for all the characters except the symbols such as slash and solid Circle, etc., which are considered to be part of the unit anyway, now judges that the fetch key string is a string expressing the unit.

When the unit recognizing section 2024 finishes retrieving the file 2026 and has reached the destination of recognizing the composite unit, it sends the unit information thus obtained to the processing section 2016 , which is then stored in the table 2016 a ( 2210 ). The unit detection is then terminated.

If, in step 2204, there is no entry in file 2026 as a result of querying this file 2026 for the string starting from the character at which pointer P is set, it means that the string is not a primitive or a composite unit can be recognized. Therefore, the section 2024 outputs the information indicating that the character string is a word that is not registered in the dictionary, that is, information indicating that the word does not express the unit is sent back to the processing section 2016 included in the table 2016 a of the processing section 2016 is conserved, whereby the unit detection is then completed.

When the unit recognition ( 2112 ) has been completed, the flowchart of FIG. 15 is returned to step 2101 , and the exclusion of the dictionary reference unit is controlled again by the processing section 2016 . After disabling the reference unit, the section 2016 judges whether or not the removed unit is still present ( 2104 ). When the cut-out unit, that is, the retrieval key string is no longer present, they are in the table 2016 a stored information by means of the output interface 2018 from the output unit (2114). The syntax analysis or so-called parsing for the input sentence is therefore completed. As described above with respect to this embodiment, the input English sentence is divided into retrieval key strings and retrieved first from an ordinary dictionary file 2022 ; if there is no entry in the file 2022 , unit discovery is performed. In the unit recognition, the fetch key string is divided and indicated by the pointer P, and the constituent dictionary file 2026 is fetched at the respective divided character string. That which is recorded in the file 2026 , or which is composed of a series of series recorded in the file 2026 , is judged whether there are character strings expressing units.

Therefore, since it is possible to perform unit recognition even for the string expressing a complicated unit by combining basic units stored in the file 2026 , so-called parsing corresponding to versatile expressions of units can be performed. Moreover, since it is only necessary that the file 2026 has basic units such as k, m, s,. , , etc., and complicated units composed of them, such as km, ○ km / s, etc. need not be stored, the capacity of the dictionary file can be reduced.

Referring to Fig. 18, the overall structure of a third embodiment in which the speech analyzer is applied to an English-Japanese automatic translation apparatus will be described.

This embodiment has an input section 3010 through which an English text 3012 to be translated into Japanese is input. The input section 3010 may comprise , for example, a keyboard with character keys such as alphanumeric or function keys, an optical character reader (OCR) for reading the English text recorded on paper, and / or a file memory device to read the English text which is printed on the Storage medium as a magnetic disk is recorded. The English text input from the section 3010 is read into the pre-editing section 3014 in which pretreatment for the translation is performed. In this case, recognition is mainly performed for the sentence and the processing of unknown words. This then acts as part of a morphological analysis. The English data is transmitted to a morpheme analysis section 3016 after the pre-editing together with the information obtained in the pre-editing. Section 3016 analyzes the morphemes of the English sentence, which sentence is subdivided by retrieving a word dictionary 3018 , performs various arrangements such as unknown word processing, noun, time-out, numbers, etc., and performs editing for the whole sentence. The morphological analysis rules are stored in a corresponding rule file 3036 .

The English data is stored in a parsing section I 3020 after the morphological analysis together with the dictionary information obtained in the morphological analysis. Section I 3020 is a functional section that analyzes the surface layer structure for the sentence by applying a grammar rule to the English data, and finds out all the possibilities in terms of structure.

The English data which has been subjected to the morphological analysis in section I 3020 is transmitted to a parsing section II 3022 together with the morphologically analyzed information. In this section, a solution is selected by applying a structural description based on the result of the morphological analysis with respect to the surface-layer structure by syntactic analysis I, thereby preparing a plausible parsing tree for the English sentence to its structure to shape. These parsing or analysis rules are stored in the rule data 3036 .

The English data which has been subjected to the corresponding analysis is transmitted as the data for the so-called parsing tree to a texture conversion section 3024 . In section 3024 , a structure tree of a corresponding Japanese sentence is prepared from the structure tree, which is an intermediate structure of the English sentence, to convert the Japanese sentence into a Japanese underlying structure, from which then a Japanese sentence can be easily translated. The structure tree data indicating the Japanese underlying structure which has been transformed in this way is delivered to a translation forming section 3026 in which the translated sentence is then formed. This is a functional section to form a Japanese sentence from the structure of the Japanese sentence structure tree.

The Japanese sentence data prepared as a translated sentence, that is, translated data is delivered to a pre-editing section 3030 . In section 3030 , the translation data retrieved from a dictionary 3018 using information that has been used in the translation to complete a more natural Japanese sentence is modified. The data for the Japanese sentence is transferred to and outputted from an output section 3032 as a translated Japanese sentence 3034 . The output section 3032 includes, for example, a printer, a display, and / or a file storage device, such as a magnetic disk.

The flow of a series of translation operations is controlled by a control section 3038 which controls the control for the entire device. The word dictionary 3018 in this embodiment stores dictionary data for English and Japanese words describing various information such as a connecting relationship, ie, concurrent relationships, meanings, plural or singular forms, language parts, etc., in addition to the vocabulary , Further, the file 3036 stores morphological and syntactic analysis rules data.

The control section 3038 is connected to an operation display section 3040 . For example, the section 3014 has operation buttons such as a translation display button or a cursor button to give various information from the operator to the corresponding device, has a display or display that visually displays the input English text, the Japanese text as a result a translation, intermediate data such as dictionary information, etc. as well as various indications for the operator. Many of these operation-indicating functions may be formed to be included in a keyboard if located in the input section 310 or to be displayed when located on the output section 3032 .

FIG. 17 shows an example of a detailed structure for processing numbers in the morphological analysis section 3016 . Of course, the portion 3016 also has another functional portion, but the parts which are immediately required for the understanding of the invention are also shown here. The morphological analysis is performed by commanding dictionary retrieval from the upper end of the input character string successively according to the retrieval key string, and processing for the dictionary information obtained from the dictionary retrieval portion 3104 according to a numerical flag becomes what will be described later.

The section 3016 has an input processing section 3100 for receiving and processing the data for the character string input from the preprocessing section 3014 . The section 3100 is provided with an input string buffer which receives the input of English string data in the form of code data of, for example, ASCII data, and temporarily stores the string data.

The input character string data temporarily stored in the section 3100 is output to a unit dividing section 3102 which divides the inputted character string data into dictionary query units such as words. Section 3102 is a functional section for distinguishing the dictionary reference units which represent a fetch key string in retrieving dictionary 3018 in section 3104 . The dictionary reference boundaries that have been used to divide the dictionary reference unit are located at a particular location of an English character, a numeric character, an apostrophe, a character other than a hyphen and a paragraph, and an apostrophe that points to a Space follows. This is stored in a bounding table 3108 and will be referred to in dividing the dictionary reference unit in section 3102 . In particular, word dictionary 3018 contains information for querying the split units. As illustrated by the example of the entry information in Fig. 24, for the entry, each of the dictionary reference units, e.g. For example, words, grammar information, such as a clause, and a discrimination indication are stored indicating that a word represents a number, ie, a numeric flag and numerical value information indicating the numerical value for a word indicating the number.

As shown in Fig. 17, both singular and plural forms are described along with each of the entries in the word dictionary 3018 , and each of them represents an entry. The numeric flag indicates that a word means a number when for this "1" is set. As further information, for example, countability and non-countability are registered for a noun, identification for transitive or intransitive verbs, translated words, etc. Since, for example, referring to "thousand", this is a noun representing a number this is the numeric flag "1", and the numeric value is "1000". Since, referring to "thread", this is a noun but not a noun indicating a number, ie, not a number, the numeric flag is registered as "0".

The recognition of the number is therefore performed by the numerical flag in the case that it is a word registered in the dictionary 3018 as, for example, "one" or "thousand". Even the unregistered words, such as a string of digits, such as "123", are two sets of digits with a "dot" in between, namely a small number, such as "10.2", and also the sequence of numeric characters with a comma in between , such as "1,000,000" are also recognized as numbers. In the present specification, the term "numeric character" usually refers not only to Arabic numerals but also to an out-of-order numeric expression such as a numeric expression. "Thirteen".

As shown in Fig. 28, the dictionary 1318 has a currency symbol table 3018a in which various currency symbols are registered, a notation symbol table 3018b in which notation symbols ","".""" (Space) "are registered, and a decimal point table 3018c in which decimal points "."",", etc. are registered. The tables for the notation symbols or the decimal points are used so that "," is used for the notation symbol, or "." for the decimal point in Japanese or English, while a space or "dot" is used mainly for the notation symbol and "," for the decimal point in other European languages such as German or French; consequently, the use of symbols between languages to be processed is different.

The dictionary retrieving section 3104 is a functional section which extracts the dictionary information by retrieving the word dictionary 3018 based on the retrieval string input from the input splitting section 3102 and transmitting them to the processing sections 3110, 3112 and 3116 .

A compilation for the succession of numerical Character is through the following two processes carried out. When words are recognized as a number, like described above, when referring to the next dictionary Reference unit is referred to and also as a Number is recognized, they are first arranged together to them in a single number. The operation will repeated as long as digits follow. For example "30 thousand" transformed into "30000", and "1.5 million" becomes transformed into "1500000". If then the numeric expression is continued, with "and" interposed if all the digits are left of "zero", which is the one Numerals of the numerical values correspond to what is indicated by the Pointer right from "and" is displayed, with regard to Meaning of the numeric expression "0", they are in  a number summarized. For example, "one hundred and thirty" in "130", while "30 thousand and two hundred "in" 30200 "is summarized.

After recognition as a number then becomes the necessary local Analysis, the so-called parsing continued. at This processing is a series of so-called Parsing units, which by the morpheme actuation information for each of the parsing units has been actuated are arranged together in a single parsing unit, which is based on a local parse rule. For example are a currency symbol and a numeric value like "Y1,000" arranged together as "a thousand yen", and a numerical value and a unit such as "1.5 km" arranged.

These arrangements are performed in the processing sections 3110 to 3122 . The processing section 3110 is a functional section for collectively arranging a number together with a currency symbol or a unit. The processing section 3112 is a function section for performing the numbering for the number. Further, the processing section 3114 is a function section for processing numbers connected by a hyphen. The processing section 3116 is a functional section for processing consecutive numeric characters.

With regard to the number after being composed with a currency symbol or a unit, the composition between the currency symbol and the numerical number is made into a single noun in the processing section 3118 . Further, the composition between the unit and the numerical value is made a proper name in the processing section 3120 . Further, in the case of a number which has been subjected to processing for numbering, in case of a hyphenated number and for a consecutive number, processing for compilation with the preceding numerical value is performed in the processing portion 3122 . The dictionary information for the inputted character string completed by such processing is stored in the ordered dictionary information buffer, ie, the dictionary information preservation table 3124 . The results of the morphological analysis are transferred from Table 3124 to parsing section I 3020 . Processing by the numerical flag is performed sequentially, as shown in Figs. 19A and 10B. The data for the input character string is received at the input processing section 3100 where input processing is performed ( 3200 ). Then, the unit dividing section 3102 divides the input character string into dictionary reference units for retrieving the dictionary 3018 ( 3201 ). The dictionary retrieval section 3104 searches the dictionary 3018 accordingly ( 3203 ) and, if there is a dictionary entry ( 3204 ), checks the numeric flag ( 3205 ). If the numeric flag is not set because the word is not a number, the dictionary information is collected in the table 3124 . If "1" is set for the numerical flag, the number is numbered ( 3206 ) in the processing section 3112, and the collective map 3207 having the previous numerical value is performed in the section 3122 . When these processings are performed for the terminal of a sentence indicated by the entered character row data ( 3202 ), a collective composition ( 3209 ) is performed between the currency symbol or the unit in the sections 3118 and 3120 , the result of the morphological analysis is then written to the syntactic analysis section I ( 3020 ) is issued ( 3210 ).

If, as a result of the dictionary reference, there is no entry at step 3204 , and the element is appended by means of a hyphen, processing for a hyphenated number ( 3213 ) is performed in section 3114 . If the first character is not a hyphenated character but a currency symbol ( 3214 ), the currency symbol is preserved ( 3216 ) alone in the table 3124 , and the currency symbol is deleted from the dictionary reference unit ( 3217 ). If the first character is not a currency symbol ( 3214 ), the processing for the subsequent numeric characters ( 3215 ) is performed in the processing section 3116 . The operation is performed to the terminal ( 3202 ).

The collective composition ( 3209 ) with the currency symbol and the unit is performed in the processing section 3110 by the processing flow shown in FIG . First, in the initial processing ( 3220 ), the upper pointer for processing is first set to the top of the buffer. If the item indicated by the pointer is not a numeric value ( 3221 ), the pointer is shifted stepwise ( 3226 ). In the event that the character is a numeric value but has no previous currency symbol and no preceding unit, the pointer is also incrementally moved ( 3222, 3224 ). The processing is carried out to the terminal of the dictionary reference unit ( 3227 ).

If the character is a numeric value ( 3222 ), the currency symbol and the numerical value are collectively assembled into a single noun ( 3223 ). For example, if the preceding symbol is not a currency symbol and the subsequent symbol is a unit, the numerical value and the unit are collectively assembled into a single noun ( 3225 ) numerical character and a unit "1.5 km" are collectively assembled into a single main word, and processing is performed up to the terminal of the dictionary retrieving unit ( 3227 ).

The processing of the hyphenated number is performed in the processing section 3114 by the flowchart shown in Figs. 21A and 21B. First, the hyphenated dictionary reference unit is stored in the buffer at the initial processing ( 3230 ). Further, the numerical value "0" is preserved, and the hyphen in the original dictionary reference unit is changed to a space. Then, the dictionary retrieving unit is divided ( 3231 ) to perform dictionary retrieval ( 3235 ). If there is no entry as a result of the dictionary retrieval, ie if the word is not registered in the dictionary ( 3236 ), the whole hyphenated dictionary reference unit is stored as a non-dictionary registered word in the table 3124 ( 3237 ).

When an entry is obtained as a result of the dictionary retrieval ( 3236 ), it is checked whether or not the numerical flag for this is "1". If the numerical flag is not a "1", it means that the character is not a numeric character, and the whole hyphenated unit of reference is kept as a non-dictionary registered word in the table 3124 ( 3237 ).

When "1" is set at the numerical flag for the dictionary input, the section 3012 numbers the number based on the input data ( 3239 ). Then, the numbered numerical value is added to a numerical value stored at that time ( 3240 ), and the addition result is conserved ( 3241 ). Thus, for example, "two" is added in "twenty-two" at "3020" corresponding to "3022". The processing is performed up to the terminal of the dictionary fetch unit ( 3232 ).

As the end position is progressively advanced, the flow is transferred to the processing 3233 at step 3232 , and the stored numerical value is made a numerical value for the entire hyphenated dictionary reference unit. Then, a collective composition 3207 for the numerical value is performed together with the previous numerical value. Next, the following numeric character processing 3215 will be explained with reference to FIGS. 22A and 22B, which is performed in the processing section 3116 . In these flowcharts, the symbol "≦" means substitution. First, an initialization 3250 is performed in which the stored numerical value val-save "0" is set, the parameter "i" is set to "1" and the pointer p is set to the upper end of the character string of the dictionary reference unit.

It is then checked if the character * p indicated by the pointer p is a numeric character ( 3251 ), a notation character ( 3252 ) or a decimal point ( 3253 ). If it is none of them, the entire character string is stored as the non-dictionary registered word in the table 3124 ( 3255 ). If it is a decimal point ( 3253 ), the parameter (i) is multiplied by 10 ( 3254 ) and step 3258 is performed. At step 3258 , the numerical value num (* p) for the character * p is added to the stored val-save numerical value to prepare a new stored numerical value. The numerical value num (* p) is a value that considers the character (* p) to be a numeric value.

If at step 3251 or 3252 the character is a numeric or a notation character, step 3257 is performed. At step 3257 , the stored value val-save is multiplied by 10, to which the numerical value num (* p) for the character * p is added to prepare a new stored numerical value.

After these processings, the pointer is incremented ( 3259 ) and the processing is repeated until the dictionary reference unit's end location ( 3260 ). When the end position for the character string is reached, the numerical value for the entire character string is formed as a conserved numerical value ( 3261 ), and a collective compilation 3207 having the previous numerical value is performed in the portion 3122 . By processing the following numeric characters, z. For example, "1.000.5" is parsed as a numeric value "1000.5".

The collective composition 3207 having the previous numerical value is performed in the section 3122 as follows. First, the dictionary table pointer is set to a previous location of the dictionary reference unit ( 3270 ). If there is nothing at this point, it means that the first digit in the preservation table indicates the numerical value, and that the numerical value for the current dictionary reference unit is recorded in the table 3124 ( 3284 ). The recorded location is the location closest to the location indicated by the pointer P.

If at step 3271 there is a word at the previous location, if the entry indicated by the pointer p is not "and" ( 3272 ), and the pointer p does not indicate the numerical value ( 3273 ), the numerical value for the current dictionary Reference unit is recorded in the table 3124 which is closest to the current position indicated by the pointer p ( 3284 ). In the example "To him two ...."("to him two ..." is "two" recently recorded as a numerical value "2".

If, in step 3273, the pointer p indicates a numerical value, the numerical value p → v for the entry indicated by the pointer p is multiplied by the numerical value v-now for the current dictionary reference unit to obtain a new numerical value p → v for the current dictionary reference unit form the entry indicated by the pointer p ( 3274 ). For example, in the case of "two thousand", "2 × 1000 = 2000" is performed to compose the entire "two thousand" into one expression. Then, the end position for the current dictionary reference unit is set as the end position for the entry of the pointer p, the p-end position ( 3282 ).

If, at step 3272, the entry indicated by the pointer p is "and", the pointer p is previously transferred to the dictionary fetch unit ( 3275 ). If it is not at the terminal ( 3276 ) and if it is a numerical value ( 3277 ), the numerical value v-now is transmitted to the current dictionary reference unit and rounded off at the most significant digit, which is then set as a value vl. For example, if the numerical value v-now of the current reference unit is "8", "8.1", "98" or "11", the value vl is "10", "10", "100" and "100", respectively.

Then, it is checked whether or not the excess obtained by dividing the numerical value pv for the entry indicated by the pointer p by vl, ie, mod (pv, vl) is "0". If it is not "0", the pointer p is incremented ( 3283 ) and the numerical value for the current reference unit is recorded in table 3124 at a location which is closest to the position indicated by the current pointer p ( 3284 ). For example, in the case of "I and two", "two" is recently recorded as a numerical value "2".

If the excess at step 3279 is "0", the numerical value v-now for the current reference unit is added to the numerical value p → v for the entry indicated by the pointer p to obtain a new numerical value p → v for the current value make the pointer p displayed ( 3280 ). In the case of "two thousand and two", for example, "two thousand" are already composed at "2000" in this stage. Then it is added with "2" of "two" by the addition 3200 , so collectively composing the whole thing to "2002". Then, the information "and" indicated by the pointer p + 1 is deleted from the table 3124 ( 3281 ), and it goes to step 3282 . Now the retrieval is explained with an example. For example, if a dictionary retrieves the input string "To him two thousand and twenty-two ...." As shown in Fig. 25, the dictionary input information is written in the table 3124 as shown in Fig. 26A. for example, the start position is "4", the end position is "6", and the speech part is a pronoun For numerical processing, it is first judged for "two" that the numerical flag is "1" ( 3205 ) and the numerical value Since the previous character of "2" in this character string is not a numerical value, it is immediately stored in the table 3124 ( 3206, 3208, 3284 ).

Then the pointer is incremented to go to the processing of "thousand". The numeric flag is "1" and the numerical value is "thousand" ( 3205, 3206 ). In addition, since the numerical value of the previous character is "2" ( 3207, 3273 ), the multiplication: 2 × 1000 is performed ( 3274 ), and its result is stored in the table 3124 (see Fig. 26B). For the next "and", then the dictionary information is temporarily stored as it is in the table 3124 (see Fig. 26C).

The pointer is then advanced to process "twenty-two". Since the words are a provided hyphenated word that is not found in the dictionary (3212), is "20 + 2 = 22" in the processing 3213 for the provided hyphenated numbers performed (3237, 3239-3241). Since the previous word "and" is ( 3272 ) and the previous numerical value "2000" is ( 3277 ), the numerical value "22" at the most significant digit is rounded down to "100" ( 3278 ) and a division operation 3279 is performed. Since the excess is "0", an addition "3280" is made between "2000" and "22". The information for "and" is eliminated from the table 3124 ( 3282 ), and the result of the addition ( 2022 ) is kept as a numerical value in the table 3124 , whereby "two thousand and twenty-two" is recognized as "2022". Consequently, a collective composition having the previous numerical value has been performed ( 3207 ). Now another example is presented. As shown in Fig. 27, so-called parsing for the input string is "$ 1,000.5 thousand. (You said $ 1,000,5 thousand was ...). "$ 1,000.5" was not registered in the 3018 dictionary. The first character is the currency symbol "$", which can be recognized as the currency symbol from the dictionary input. This is recorded independently in the table 3124 ( 3214, 3216, Fig. 29A).

Then "1.0005." is formed into a numerical value "1000.5" by the subsequent numerical character processing 3215 . Since the previous character is the symbol "$" and not a numerical value, the numerical value is recorded per se ( 3280 to 3273, 3284, Fig. 29B).

The next word "thousand" is a number and its numerical value is "1000". Since the previous character is a numerical value ( 3272, 3273 ), a calculation is made: "1000.5 × 1000 = 1000500" ( 3274 , Fig. 29C). After the dictionary retrieval has been completed, the content stored in the table 3174 is checked one by one. Since the currency symbol "$" exists immediately before the numerical value "1000500", both are collectively assembled and "$ 1000500" is formed as a single word input ( 3209, 3221 to 3223 , Fig. 29D).

Now, the fourth embodiment will be explained. Fig. 30 shows the fourth embodiment of the speech analyzer having features of the invention applied to an English-Japanese automatic translator.

This embodiment has an input processing section 4014 into which data is input from an input device 4012 . The input device 4012 includes, for example, a keyboard with character keys such as alphanumeric and function keys, an optical character recorder for reading an English text recorded on paper, and a reader such as a magnetic disk.

The input processing section 4014 has a buffer 4014 for a an input character string, and stores the input from the device 4012 English sentence in the buffer 4014 a. The section 4014 reads an input sentence stored in the buffer 4014 a and delivers it to the unit partitioning section 4016 . The section 4016 is a function section which divides the dictionary reference unit of the sentence input from the section 4014 . In a delimitation table 4018 delimitations, such as space and comma, are stored. The section 4016 stores boundaries from the delimitation table 4018 and divides the input sentence from the section 4014 into character strings as the units in the case of fetching a reference memory 4020 by dividing the sentence at the places where the separations exist. The divided character strings are input to the dictionary retrieval section 4022 .

The section 4022 retrieves the reference dictionary 4020 for the input sentence, which is divided into reference units that have been submitted from the section 4016 . For example, the reference dictionary 4020 stores entries for the English character strings, language parts thereof, feature information, etc. as shown in FIG . The dictionary 4020 additionally stores, for example, proper names shown in the figure, character strings, for another part of speech, verbs and adjectives. The record for the proper name as the speech part in the figure means that it is applied in the processing for the registered proper name, which will be described later, but that no ordinary common grammatical names are expressed. Further, the feature information indicates what is represented by the proper name, which may not always be limited to one, since, as will be described later, a proper name represents a number of features depending on the use case.

The section 4022 retrieves the dictionary 4022 for the reference-unit-divided string, and when the character string is a proper name, it is delivered to the proper name processing section 4024 where processing for the proper name is performed, which will be described later. If it is not a proper name, it is delivered to the processing section 4036 and stored in its table 4036 a.

The proper name processing section 4024 has a preceding sentence end processing section, a preceding proper name processing section 4028, a proper name processing section 4030 itself, a previous proper name processing section 4032 and proper name, and a default setting section 4034, respectively. or standard feature information. The section 4026 judges whether or not the string preceding the string retrieved by the section 4022 and input from the section 4022 is at the end of the sentence. When the previous string is at the end of the sentence, it capitalizes the capital letter at the beginning of the string to be processed, passes it to dictionary retrieval section 4022 , and causes section 4022 to retrieve the reference dictionary 4020 . The string which has not been recalled even by the second fetching operation is judged to be a non-registered proper name, which is then sent to the processing section 4036 and then stored in the table 4036a . Further, if the preceding character string is not at the end of the sentence, it is delivered in the processing section 4036 as a proper name whose feature information is unknown and is registered in the table 4036 a, as will be described later.

The section 4028 dissects the feature information for the previous character string from the section 4026, and outputs the result to the proper name processing section 4030 itself. The section 4030 checks the feature information of the proper name to be decomposed, as described later. If the feature information is not registered for either the proper name or the previous proper name, it collectively dissects the proper name and the previous proper name using the registered information for the other of them and stores it in the table 4036a in the section 4036 .

As such, the previous proper name processing section 4032 and proper name itself checks the part common to the feature information for the proper name and the preceding proper name to be dissected, dissects these proper names with respect to the common part, outputs the result to the processing section 4036 , and stores him in his table 4036 a. The section 4036 provides feature information on a proper name after having been read from the table 4036 a, sent over the section 4016 to the section 4022 , and found out as the result of querying the dictionary 4020 in the section 4022 that no Feature information is present. Since it is effective that a proper name has different kinds of features depending on the use cases, all feature information is provided, which are necessarily to be considered. For example, "person, place, group, and the like." intended. After the proper name is provided with the feature information, the section 4034 sends the data to the processing section 4036 and stores it in its table 4036a .

The section 4036 having the table 4036 a stores the data from the previous proper name and proper name section 4032 , the section 4034, and the fetch section 4022 into the dictionary information preservation table 4036 a, and thereafter reads and outputs the stored data parsing section 4038 . The section 4038 performs the analysis for the input sentence after being subjected to morpheme analysis and read out from the table 4036 a.

The operation of this device will now be explained with reference to the flowchart in FIG . First, an English sentence input from the device 4012 is read into the input processing section 4014 ( 4100 ). The read section 4014 in the set is loaded in the buffer 4014 a, 4014 a buffer in the loaded, entered sentence is read in the unit Aufteilabschnitt 4016th When the sentence is entered, the section 4016 reads from the delimitation table 4018 boundaries to partition the dictionary reference units ( 4102 ). That is, characters following the input sentence are successively divided from the top of the character strings into the fetching character strings as the fetching reference dictionary unit 4020 by dividing them into the parts where demarcations such as there is a gap or a colon. The section judges whether or not the split dictionary reference unit, ie, fetch strings are finished ( 4104 ), and if there are still (unfinished) fetch strings, returns the fetch string to the section 4022 .

When a fetch string is submitted to section 4022 , section 4022 retrieves fetch string reference dictionary 4022 ( 4106 ). The section judges whether the retrieval character string in the entry of the reference dictionary 4020 is present or not, as shown in Fig. 31 (4108), and if there is an entry, reads out the information stored in the dictionary 4020 part of speech information and assessed whether the fetch string is a proper name or not ( 4110 ). If the fetch string is not a proper name, the portion 4022 outputs the data read from the dictionary 4020 to the processing portion 4036 and records it in its table 4036a ( 4112 ). When the data is stored in the table 4036a , the entry indicating that the data is stored in the section 4016 and the data for the just-retrieved retrieval string are inputted from the processing section 4036 . Then, step 4102 is returned and the dictionary reference unit is again divided into the section 4016 .

At step 4110, if the fetch string is a proper name, the portion 4022 gives the data for the proper name read from the dictionary 4020 , which is hereinafter referred to as a proper name, to the proper name processing portion 4024 along with the data for the preceding one Fetch data stream entered from the table 4036a in the section 4036 to the section 4022 via the section 4016 , and processing for the dictionary-registered proper name is performed in the section 4024 ( 4124 ).

Now, the processing for the proper name registered in the dictionary will be explained with reference to the flowchart in FIG . The data output from the section 4022 to the section 4024 is input to the previous proper name processing section 4028 by the preceding sentence end processing section 4026 . When processing the proper name registered in the dictionary, section 4026 has no function. The processing section 4028 judges whether or not the input preceded by the proper name is a proper name not registered in the dictionary 4020 , that is, whether or not it is used in processing for the unregistered proper name, which will be described later ( 2100 ). If it is an unregistered proper name, the processing section judges the whole part of the proper name and the previous unregistered proper name as a proper name having the feature information for the proper name ( 4202 ) outputs the data to the processing section 4036 and stores it in its table 4036 a ( 4218 ).

If the section 4028 judges in step 4200 that the input preceding the proper name is not a registered proper name, it then judges whether the entry preceding the proper name is a proper name registered in the dictionary 4020 or not ( 4204 ). If the entry just preceded by the proper name is a registered proper name, it is judged whether or not the feature information for the previous proper name is known, that is, whether it is registered in the dictionary 4022 ( 4206 ).

If the feature information for the previous proper name is unknown, step 4202 is proceeded to judge ( 4202 ) the entire portion of the proper name per se and the immediately preceding proper name as a proper name with feature information ; the previous proper name processing section 4028 then outputs the data to the processing section 4036 , in the table 4036 a of which the output data is recorded ( 4218 ).

When the processing section 4028 judges in step 4206 that the feature information for the previous proper name is not unknown, ie, that it is registered in the dictionary 4020 , the data is output from the section 4028 to the processing section 4030 . The section 4030 then judges whether the feature information of the proper name is unknown or not ( 4208 ). When the feature information for the proper name is unknown, the section 4030 judges the whole portion of the own name per se and the immediately preceding proper name as a proper name having feature information of the proper name ( 4210 ) and outputs the data to the processing portion 4036 , Table 4036 a records the data.

When the processing section 4030 determines that the feature information of the self-name per se is not unknown, that is, that it is registered in the dictionary 4020 , the section 4030 gives the data to the preceding proper name processing section 4032 and the proper name itself. The processing section 3032 then judges whether any common feature exists in the feature information between the proper name itself and the proper preceding name ( 4212 ). If there is a common feature, it judges the whole share for the proper name per se and for the immediately preceding proper name as a proper name having the feature information of the common part ( 4214 ), and outputs the data to the processing section 4036 in which Table 4036a , the data is then recorded ( 4218 ).

If there is no common feature information characteristic between the proper name itself and the proper name immediately preceding it, the processing section judges that the proper name is a proper name with the feature information retrieved from the memory 4020 , which is different from the immediately preceding one Proper names discriminate ( 4216 ), and outputs the data to the processing section 4036 in whose table 4036 a the data is then recorded ( 4218 ). In FIG. 32, if there is no retrieval string in the entry of the reference dictionary 4020 in step 4108 , it is judged whether or not the first character of the retrieval string is a capitalized letter, and if it is not a capitalized letter, the portion 4022 regards The retrieval string as an unregistered word is sent to the processing section 4036 to be recorded in its table 4036 a ( 4118 ).

If the first character is a capitalized letter, the data for the fetch string, along with the data for the preceding fetch string, is passed from the portion 4022 to the proper name portion where the processing for the non-dictionary registered proper name is performed ( 4120 ).

Now, the processing of non-dictionary registered proper names will be described with reference to FIG . The data for the fetch string is output to the processing section 4026 along with the data for the previous fetch string, which then judges whether the end of the previous input is a candidate for the end of the sentence or not ( 4300 ). The judgment as to whether or not it is a candidate for the end of the sentence is made by judging whether the end of the previous input is a candidate for the end of the sentence, such as a separate point (.), Etc. or not ,

If the end of the previous input is a candidate for the end of the sentence, data is output from processing section 4026 to processing section 4028 , which then considers the previous input as the end of the sentence ( 4302 ), the first character in the fetch string converted to a small letter and sent to section 4022 .

The section 4022 retrieves the dictionary 4020 for the retrieval string that has been converted into the lowercase characters (4304), and judges whether or not there is an entry in the reference dictionary 4020 is not (4306). If there is an entry, the section 4022 outputs the data retrieved from the dictionary 4020 to the processing section 4036 and stores it in its table 4036a ( 4308 ). If there is no entry, the section 4022 restores the first character in the fetch string to the uppercase character, issues it as a non-registered proper name to the processing section 4036 , thereby recording it in its table 4036a ( 4310 ).

If, at step 4300, the previous sentence end processing intersection 4026 judges that the end of the previous entry is not a candidate for the end of the sentence, the data from section 4026 is presented to section 4038 , which then judges the previous entry such that he is not the end of the sentence ( 4312 ). The data from section 4028 is passed to section 4030 , which then considers the fetch string as a proper name whose feature information is unknown ( 4314 ).

The section 4030 then returns the data to the processing section 4028 , which then performs the processing for the dictionary-registered proper name ( 4316 ). The processing for the proper name registered in the dictionary is the same as that shown in FIG .

Now, in FIG. 32, when the reference unit is at step 4104 , the section 4022 outputs a signal indicating this to section 4034 , which then reads out the information recorded in table 4036a in section 4036 and the proper name with the default Feature information ( 4122 ).

Now, with reference to Fig. 35, the operation will be described to provide the proper name with the default feature information. In the dedicated section 4034 , a pointer is first set to the top of the data in the table 4036 a ( 4400 ). That is, the pointer is placed at the top of the input sentence, which is divided into entries each provided with information by the retrieval in the reference dictionary 4020 . It is then judged whether the entry indicated by the pointer is a proper name or not ( 4402 ); if it is a proper name, it is judged whether the feature information of the proper name is unknown or not ( 4404 ). If it is not a proper name, the flowchart proceeds to step ( 4408 ) and the pointer advances to the next entry.

If the feature information for the proper name is unknown at step 4404 , the default feature information is provided ( 4406 ). The default feature information is provided at a proper name whose feature information is unknown, as shown in the lower part of FIG . There, then, for example, the proper name "Johnson" whose feature information is unknown is provided with all kinds of feature information, ie, "person, place, group and others". By providing the proper name whose feature information is unknown with all kinds of feature information, it is possible to leave a space so that the proper name can be decomposed into one of the plurality of features in the subsequent texture analysis.

If the feature information for the proper name is not known at step 4404 , the flowchart advances to step 4408 and the pointer is advanced to another entry. It is then judged whether the entry indicated by the pointer is at the end or not ( 4408 ), and if it is not at the end, the flow returns to step 4402 , and then the next entry is checked to see if it is Proper name is or not. If the entry indicated by the pointer is at the input, the provision of the default information is terminated. After the provision of the default feature information on the proper name is completed, the data recorded in the table 4036 a is output from the section 4036 to the syntactic analysis section 4036 ( 4124 ), whereby the morphemic analysis in this Execution is finished.

Now, the operation of the above-described device will be explained, for example, on an input sentence. The explanation will be made in the case where a phrase "met at Tokyo Station Mr. Walter Johnson" is entered. First, an input processing ( 4100 ) is performed to read the input sentence into the processing section 4014 . Then, the dictionary reference unit is divided ( 4102 ), and the input sentence is divided into spaces by spaces. The reference dictionary 4020 is first fetched for "on"("In") ( 4106 ). Since there is no entry for "Up" located in the reference dictionary 4020 , the step first transfers to the processing for the proper name not registered in the dictionary. However, since the previous part is at the top of the input string, it is considered to be a candidate for the end of the sentence which has been retrieved as "on" by adding I and A in a and i, respectively, from the reference dictionary 4020 is implemented. Since there is no proper name ( 4110 ), the data retrieved from the dictionary 4020 is recorded in the table 4036a ( 4112 ). Then, the dictionary 4020 is retrieved for "Tokyo" ( 4106 ). Since there is no entry in dictionary 4020 for "Tokyo" ( 4108 ) and the first character is a capital letter ( 4116 ), processing is performed for the non-dictionary registered proper name ( 4120 ). Then, in the flowchart, FIG. 34 is advanced. Since the previous part is "up or in", this is not a candidate for the end of the sentence ( 4300 ); "On" is not judged as the end of the sentence ( 4312 ) "Tokyo" is recognized as a proper name whose feature information is unknown ( 4314 ), and the processing for the proper name registered in the dictionary is performed ( 4316 ) ,

In the flowchart, Fig. 33 is advanced. Since "Up or In" in the preceding part is neither a non-registered ( 4200 ) nor a registered proper name ( 4204 ), it is recorded as a proper name having inherent feature information per se, ie, as a proper name whose feature information is unknown is ( 4216 ). In Fig. 32, retrieval is then performed for dictionary 4020 for "station" ( 4108 ). Since there is an entry in the dictionary 4020 for "station" ( 4108 ), and since it is a proper name ( 4110 ), processing for the proper name registered in the dictionary is performed ( 4114 ). Turning now to the flowchart of FIG . Since "Tokyo" in the preceding part is an unregistered proper name ( 4200 ), "Tokyo Station"("TokoyStation") as a whole is recorded as a proper name containing the feature information "Location and Group" in the form of the term "Station". (or "station") has ( 4202 ). Then, the reference dictionary 4020 in Fig. 32 for "Mr." retrieved ( 4106 ) because "Mr." an entry in the dictionary 420 and a proper name is ( 4110 ), processing for the proper name registered in the dictionary is performed ( 4114 ). Then the flow advances to Fig. 33. "Station" in the previous part is not a non-registered proper name ( 4200 ) but a registered proper name ( 4204 ) and the feature information (place, group) is not unknown ( 4206 ). Since "Mr." is a feature information "person" and is unknown ( 4208 ), it is judged whether any common part in the feature information "station" in the preceding part and "Mr." is present ( 4212 ). Since "station" means "place, group" while "Mr.""Person" means and there is no common part between them, "Mr." alone registered as a proper name with the feature information "person". ( 4216 ). Returning to the flowchart of Fig. 32, the reference dictionary 4020 is fetched for "Walter" ( 4106 ). Since there is an entry for "Walter" in the dictionary 4020 ( 4108 ) and this is a proper name ( 4110 ), processing for the registered proper name is performed ( 4114 ). It is then transferred back to the flowchart of FIG. 33. Since "Mr." in the preceding part is not a non-registered proper name ( 4200 ) but a registered proper name ( 4204 ), and since the feature information "person" is not unknown ( 4206 ) and furthermore the feature information for "Walter""person, place, group" also does not is unknown, the common part for the feature information is judged ( 4212 ). Since there is a separate part ("person") present for the feature information, "Mr. Walter" is recorded collectively as a common noun with the feature information "person" ( 4214 ).

Then "hit" is retrieved from the dictionary 420 . And since there is an entry ( 4108 ) and it is not a proper name ( 4110 ), the data recorded from the dictionary 4020 is recorded in the table 4036a ( 4112 ). Further, "Johnson" is retrieved from dictionary 4020 ( 4106 ). Since there is no entry for "Johnson" ( 4108 ) and the first character is a capitalized character ( 4116 ), processing is performed for the proper name not registered in the dictionary ( 4120 ). Then, the flowchart of Fig. 34 is gone. Since "hit" in the preceding part has no candidate for the end of the sentence ( 4300 ), it is judged whether "hit" is not the end of the sentence ( 4312 ), "Johnson" is considered a proper name with unknown feature information ( 4314 ) and processing for the proper name registered in the dictionary is performed ( 4316 ). Then, again, the flowchart of Fig. 33 is gone. Since "hit" is present in the previous part as neither a non-registered ( 4200 ) nor a registered proper name ( 4204 ), "Johnson" alone is recorded as a proper name whose feature information is unknown.

After the foregoing processing steps, default characteristic information is provided for the proper name, as shown in FIG . The pointer is set to "up"("In") as the top of the dictionary reference unit ( 4400 ). Since it is not a proper name ( 4402 ), the pointer is advanced ( 4408 ) and set to "Tokyo Station". Since "Tokyo Station" is a proper name ( 4402 ) and the feature information is not unknown, since the total part "Tokyo Station" has been recognized as a place, group in the previous processing for the registered proper name ( 4404 ), the pointer is advanced ( 4408 ) and set to "Mr. Walter".

Since "Mr. Walter" is also a proper name ( 4402 ) and the feature information is not unknown ( 4404 ), the pointer is advanced ( 4408 ). Since "hit" is not a proper name ( 4402 ), the pointer is advanced ( 4408 ). Since "Johnson" is a proper name ( 4402 ) whose feature information is unknown, default feature information is provided ( 4406 ) and "Johnson" is provided with feature information "person, place, group, and the like." provided as shown in Fig. 36.

As described above, in this embodiment, a German (English) input sentence in fetch divided strings, which are then retrieved from the reference dictionary 4020, and if there is an entry as a proper name in the dictionary 4020, a processing for the registered Proper names performed. In the processing of the registered nickname, the previous fetching string is taken into account, and if it is a proper name, the feature information for the previous fetching string and the proper name are checked as the object. If it has no feature information, the other feature information is provided, while if any feature information exists for both, the common part is considered as the feature information of those proper names. Consequently, it is possible to suitably provide a proper name having no feature information with feature information and limit the provided feature information to more appropriate feature information. This allows a more efficient analysis in the subsequent structural analysis and in performing a corresponding translation.

Further, for the character string not registered in the dictionary 4020 , if the first character is a capitalized character and it is determined that the previous character string is the end of the sentence, since the capitalized character is converted to a lowercase character, and When dictionary 4020 is retrieved again, it is possible to retrieve the string at the top of the sentence in dictionary 4020 as well . Further, when a character string beginning with a capitalized character appears in a part other than the upper part of the sentence, it is judged to be a proper name, and the feature information for the proper name is provided with registered feature information by means of a proper name, if they are before and after exists. Thus, a proper name not registered in the dictionary 4020 may be parsed to some extent grammatically.

Furthermore, since a proper name, not with a feature information  is provided with all the necessary feature information and the non-required feature information removed during processing for the word is, it is possible, a proper name, whose feature information is not known or an unregistered proper name analyze.

Further, because a number of feature information is at a whole certain proper names are provided and appropriate feature information depending on the feature information of the proper name before and after, Is it possible to provide corresponding feature information for select the case of having a proper name with a number of feature information in relation to others before and is then parsed grammatically to thereby a to enable effective parsing of the entered sentence.

A fifth embodiment will now be described, in which Fig. 38 shows the entire structure of the fifth embodiment of a speech analyzer having features of the invention applied to a German (English) Japanese automatic translation device. This embodiment has an input section 5010 , and a German / English text 5020 to be translated into Japanese is entered through the section. The input section 5010 may be, for example, a keypad with character keys such as alphanumeric or function keys, an optical character recorder (OCR) to read the English / German text recorded on paper, and / or a file memory device for reading the German / English text, which is recorded on a storage medium, such as a magnetic disk.

The German / English text input from the input section 5010 is read to a pre-editing section 5014 where pretreatment for the translation is performed. In this case, sentence recognition and the processing of unknown words are mainly performed. These functions are part of the morphological analysis. The pre-processed German / English data, together with the information obtained in the pre-editing, is transferred to a section 5016 for morphological analysis. The section 5016 divides the data into sentences while retrieving a word / dictionary 5018 , dissects German / English morphemes, performs processing for unknown words, proper names, various compositions such as a time expression, and performs processing for the entire sentence, like a supplementary question and an apposition recognition. The rules for the morphological analysis are stored in a rule file 5036 .

The German / English data are transferred to a section I 5020 for syntactic analysis after morphological analysis together with the dictionary information obtained in the morphological analysis. Section I 5020 is a functional section that parses the surface structure for the sentence grammatically by applying grammatical rules to the German / English data, and finds out every structural possibility.

The German / English data which has been subjected to the syntactic analysis in the section I 5020 , together with the analyzed information, is delivered to a section II 5022 for syntactic analysis, whereby a solution from the result of the syntactic analysis with respect to the surface structure by the syntactic analysis I is selected by applying the structure description. A plausible parsing tree for the German / English sentence is thus prepared and its structure formed. The rules for a syntactic analysis are also stored in the rules file 5036 .

The German / English data is transmitted to a structure conversion section 5024 as the data of the so-called parse tree after performing the syntactic analysis. In the section 5024 , a structure tree of the corresponding Japanese sentence is prepared from the structure tree, which is an intermediate structure of the German / English sentence, and is converted into a Japanese underlying structure from which the Japanese sentence can be easily translated. The data for the structure tree indicating the Japanese underlying structure that has undergone the texture conversion is delivered to a translation section 5026 in which a translated sentence is generated. This is a function section that generates a Japanese sentence from the tree structure of the Japanese structure tree.

Japanese data as a result of the translation, that is, data for the translated sentence is then given to a post-editing section 5030 . The section 5030 modifies the translated data by retrieving a dictionary 5018 using information used in the translation to perfect a more natural Japanese sentence. The data for the Japanese sentence is output to an output section 5032 and as a translated Japanese sentence 5034 from the output section 5032 . The dispensing section 5032 may include, for example, a printer, a display, and a file storage device, such as a magnetic disk. A flow of such a series of translation operations is controlled by a control section 5038 which controls the control of the entire device. The word dictionary 5018 stores dictionary data for German / English and Japanese words, and various information is set therein, such as a connecting relationship, ie, an existing relationship, as well as meanings, a singular or plural form, a speech part, etc. and all in addition to that Vocabulary. Further, a file 5036 stores rule data for morphological and syntactic analysis.

The control section 5038 is connected to an operation display section 5040 having operation buttons for providing various displays from an operator to the illustrated device, such as a translation display button, a cursor button, etc., and a display or a visual display entered German / English text, a Japanese text as a result of translation, intermediate data, such as dictionary information, various displays for the operator, etc. represents. It may also be designed so that most of the operation display functions are provided in a keypad when it is disposed on the input section 5010 or provided in a display when it is disposed on the output section 5032 .

In Fig. 37, detailed structures concerning the processing of proper names in the morphological analysis section 5016 are shown. Section 5016 is illustrated for the part having direct relationships to the understanding of the invention, although of course there are other functional sections concerning the morphological analysis. The morphological analysis is performed by commanding the dictionary retrieval from the top of the input character string sequentially according to the retrieval key string and performing the processing for the thus obtained dictionary information from the dictionary retrieving portion 5104 according to the position information of the proper name, which will be described later.

The section 5016 has an input processing section 5100 for receiving the data for the character string input from the section 5014 and performing the input processing. The section 5100 is provided with a buffer for the input character string supplied with the German / English character string data in the form of code data such as ASCII, and temporarily stores the character string data.

The data for an input character string temporarily stored in the portion 5100 is delivered to a portion 5102 which divides the data into dictionary reference units such as words. The section 5102 is a functional section which discriminates the dictionary reference units representing the retrieval key strings to successively retrieve the dictionary 5018 in the section 5104 . The dictionary reference boundaries used in dividing the dictionary reference units are arranged at the respective location of the German / English character as a number, an apostrophe, a character other than a hyphen and dot, and an apostrophe followed by a space. They are stored in a delimitation table 5108 and are referred to when dividing the dictionary reference units in section 5102 .

The reference dictionary 5018 stores, in particular, the information for retrieving the apportionment unit. For example, as shown in Fig. 38 for the example of the input information, grammar information such as a sentence part for each of the dictionary reference units, for example, an entry for a word, is included. The language part information contains an indication for the noun if it is a general noun or a proper noun. For the proper name, a discrimination notice indicating the way to limit the position in the sentence, that is, position information for the proper name is stored. This will be described later in detail. Also, as other information, for example, a countability or non-countability of a main word, a distinction such as intransitive or transitive verb, their translation, etc. are registered.

There are four kinds of position information for the proper name, d. H. in the present embodiment, "0" to 3". The pattern "0" shows a proper name without positional restriction, for example "city / city" or one Person name "Walter". The pattern "1" indicates that it's a proper name, for example "Mr." is the one at the top End of a single proper name is arranged, or one Sequence of a number of proper names, d. H. a proper name, which is arranged in a single group of proper names. The pattern "2" indicates that it is a proper name, for example "Station / Station" or "Bay / Bay" at the end of a single proper name, or that it is a Is proper name in one word as a group of proper names is arranged, or which is different than the pattern "3" described below. The pattern "3" shows assume that it is a proper name, for example, "River" in "the Sumida River," which is the same as the pattern "2," however, from a particular article "the / the" at the beginning of a Nickname in a proper noun as a group of Proper name is arranged.

The dictionary retrieval section 5104 is a functional section which issues information by retrieving the word dictionary 5013 based on the retrieval key string which has been input to the section 5102 , and transmits the same to the dictionary information preservation table 5124 . to the position information processing section 5110 and to the section 5111 which judges the previous sentence end. The processing based on the patterns "0" to "3" corresponding to the position information for the proper names retrieved from the dictionary 5018 are performed by proper name processing sections 5114, 5116, and 5118 . Processes for the proper names are performed by the pattern "1" in the section 5114 , by the patterns "2" and "3" in the section 5116, and by the pattern "0" in the sections 5118 and 5114, respectively.

In this embodiment, proper names are collectively arranged by, for example, using as a key a word representing a part of a group of proper names grouped in a single proper name and subjecting them to a positional restriction when arranged as a proper name. Even if a number of proper names are continuously present, they can be appropriately arranged together with the context without such erroneous composition simply by always being arranged as a single group of proper names. The processing for this purpose is performed in Sections 5114, 5116 and 5118 .

Proper names with a certain scope are registered in the reference dictionary 5018 . Such proper names registered in the dictionary are subjected to a grammatical analysis in the section 5110 and the sections 5114, 5116 and 5118 . They form a processing section for proper names registered in the dictionary. The proper names not registered in the dictionary 5018 are grammatically parsed in the preceding sentence end judging section 5112 and in the proper name processing section 5118, respectively, for the pattern "0". These form the processing section for proper names not registered in the dictionary.

Processing of a proper name is performed by the following two steps. First, a proper name is recognized in the input string. In the case of a word registered in the dictionary 5018 , this is done by displaying the proper name in its morphemic operation information. In the case of a word not registered in the dictionary 5018 , it is done so that a character at the top is a German / English uppercase character, for example, "John" or "US", etc.

Then a group of proper names is arranged collectively, to make the whole part a single proper name. If recognized as a proper name from the dictionary information and if the next dictionary reference unit is also a proper name, the whole part becomes collective composed of a proper name. For example, "M. Weber "analyzed as a whole as a proper name The analysis forms a candidate to the ideological Expression including proper names in the local group grammatical dissection or analysis.

Then the necessary local analysis is performed. In In this case, a sequence of so-called parsing Units indicated by the morpheme actuation information for each of the parsing units have been operated collectively arranged in a parsing unit, which is on a local Parsing rule is based. For example, "Mr. Brown" arranged in "Brown shi". Furthermore, words that have a Part of a district name also represent collectively. For example, "Lake Biwa" is put together in "Biwako". In the same way words become a part of a group name, also collectively arranged. For example, "Yale University" is called "Yale Daigaku" analyzed.

In the case of the proper name "Mr. ...." and "Lake ...." There is always an end to the context in front of it. Consequently, when "Tom Brown" is collectively grouped into a single proper name, an error is made in the following analysis. For example, a proper name "University / University" is always immediately followed by an end. For example, in an English sentence "At Yale University Tom is ...." it is recognized that there is an interruption between "University" and "Tom". In this embodiment, information for the position at which respective proper names in the sequence of proper names undergo a positional restriction is stored in the dictionary 5018 as the above-described position information, that is, as patterns "0" to "3". The collective arrangement with this position information is performed in the processing sections 5110, 5112, 5114, 5116 and 5118 . The dictionary information for the input character string after the completion of these operations is stored in the retrieved dictionary information buffer, that is, in the dictionary information preservation table 5124 .

The result of the morphological analysis is transferred from the table 5124 to the section I 5020 for morphological analysis. The processing by the proper name position information is performed by means of the sequence shown in FIG . Input processing is performed by taking the data for the input character string into the input processing section 5100 ( 5200 ). Then, the section 5102 divides the entered character string into dictionary reference units for retrieving the dictionary 5018 ( 5201 ). Section 5104 retrieves dictionary 5018 accordingly ( 5203 ), and if there is a dictionary entry ( 5204 ), checks its dictionary ( 5205 ). When the speech part is not a proper name, processing for the proper name is not performed in this embodiment because the dictionary information is stored in the table 5124 ( 5206 ). If it is a proper name, the processing for the dictionary registered proper name 5207 in the section 5110 and the sections 5114, 516, and 5118 is performed. When these processings are performed at the end position of the sentence indicated by the data of the input character string ( 5202 ), the result of the morphological analysis is given to the designated section I 5020 ( 5210 ).

If, as a result of the dictionary reference, there is no input at step 5204 , and the element starts from a capitalized character ( 5212 ), it is recognized as a proper name not registered to the dictionary and a processing for one not in the dictionary registered proper name 5213 is performed in the preceding part judging section 5112 and in the proper name processing section 5118 . If the beginning character is not a capitalized character, since this is a word not registered in the dictionary 5018 , it is kept as a non-registered word in the table 5124 ( 5214 ). The processing is then performed at the end position ( 5202 ).

Processing for the dictionary-registered proper name 5207 is performed in the processing sections 5110, 5114, 5116, and 5118 in the flowchart shown in FIG . First, the obtained position information included in the dictionary information is referenced ( 52209 ). The proper name processing 5221 for the pattern "0" is performed when displaying the pattern "0"; the proper name processing 5220 is performed for the pattern "1" when displaying the pattern "1", and the proper name processing 5223 is performed for the pattern 2, 3 when displaying patterns "2" and "3", respectively , The processing 5221 for the pattern "0" is performed in the section 5114 . Processing is applied to a proper name that has no positional constraint. First, when a part which precedes the dictionary reference unit concerned is a non-registered proper name ( 5230 ), the whole part is collectively arranged in a proper name, the position information being "1", and is stored in the table 5124 ( 5233 ) , If the previous part is a proper name with position information "1" 852319 , the processing is performed in the same way.

If the preceding part is a proper name having position information "0" ( 85231 ), the whole part is collectively assembled into a proper name having position information "0" and stored in the table 5124 ( 5235 ). Further, if the previous part is not a proper name having position information of "0", the whole part is stored alone as a proper name with position information "0" in the table 5124 ( 5134 ).

The proper name processing 5222 for the pattern 1 is performed as described below. The processing is applied to a proper name, such as "Mr.", which is located at the beginning of a single nickname or at the beginning of a nickname organized in a group as a sequence of a number of nicknames. First, if the part preceding the candidate reference unit is a non-registered proper name ( 4240 ), the word is converted to unregistered ( 5241 ). If it is not a non-registered word, it is stored alone as a proper name with the position information pos "1" in the table 5124 ( 5242 ).

Now, the proper name processing 5523 for the patterns 2, 3 will be described with reference to FIG. 44. The processing is applied, for example, to a proper name such as "station" or "river" arranged at the end of a single proper name, or a proper name belonging to a group as a sequence of a number is arranged by proper name. First, if the part which precedes the candidate reference unit is an unregistered proper name ( 5250 ), that part collectively collects together with the previous word as a proper name with its own position information pos-self as the proper name position information pos and in the Table 5124 is stored ( 5225 ). Further, the processing is performed in the same manner when the previous part is a proper name with the position information "1". ( 5251 ).

If the preceding part is not a proper name with position information "0", it is ( 5252 ) arranged in a proper name with a single position information pos-self as the proper name position information and stored in the table 5124 ( 5257 ).

If, at step 5252, the previous part is a proper name of the position information "0", its own position information is checked pos-self ( 5253 ), and the processing 5255 is performed if it is the pattern "2". If the own position information pos-self is the pattern "3", it is further checked whether the item before a reference unit is "the" or not. If it is not the particular article "the", processing 5255 is performed. If it is "the", the group of "the" with the own item is collectively assembled as a proper name with the position information "3" and stored in the table 5124 ( 5256 ).

For a word that begins with a capitalized letter and is recognized as an unregistered word for which there is no entry in the dictionary 1518 as a result of the retrieval operation 5203 , the flow is switched to the processing 5213 by means of steps 5204 and 5212 and the processing 5213 is performed in the previous sentence end judging section 5112 . First, when the part which precedes the candidate reference unit is not a candidate for the sentence end, the processing 5221 for the pattern 0 is performed in the processing section 5118 as described above.

The previous part may be a candidate for the end of the sentence in the following four cases. The first is a case where a separate dot "." is available. The next is a case where the previous entry the point is at the last place, and the position information  for the proper name is not "21". This case for example, includes an abbreviation such as "U.S.A." on. There is also the case of a colon ":", a semicolon ";", a sequence of a dot and an apostrophe Write "." and a sequence of one dot and one Quotation marks "." ". The last case is that it is at the top of the buffer for an input String is located.

If the previous part is one of the four cases described above, the previous candidate for the sentence end is recognized as the sentence end ( 5261 ), and the dictionary was retrieved after the uppercase letter of the word has been converted to a lower case letter ( 5262 ). If a dictionary entry is obtained as a result of retrieval ( 5263 ), it is recorded in table 5124 ( 5264 ). If not, it is recorded as a non-registered proper name in the table 5264 with the upper character left unchanged as the capitalized letter ( 5265 ).

Now, the explanation of another example will be given. For example, if a dictionary entered an input string "Along the Sumida River, Paul and Mr. Gold Smith. ("Along the Sumida River Paul and Mr. Gold Smith went ..."), the dictionary input information is first written to the table 5124 as shown in Fig. 64A. For example, for "the" or "the", the starting position in the sentence is "7" and the ending position is "9", and the speech part is a particular article. No entry can be obtained for the word "Along" at the top of the input character string at dictionary retrieval 5203 , and it is determined to be unregistered. However, since the previous part can satisfy the condition of the candidate for the sentence end, since it is present at the upper end of the input buffer ( 5260 ), the large initial letter "E" or "A" is initially converted to a small letter and the Dictionary fetch 5262 is performed as "along""along".

Then the pointer is incremented to proceed to processing for "Sumida". This word is not registered in the dictionary 5018 in the present embodiment. Since the previous part is not a candidate for the sentence end, the flow is transferred to the proper name processing 5221 for the pattern 0. As shown in Fig. 46A, a proper name for the speech part and "0" for the proper name position information are retrieved.

The next reference unit "River" is a proper name with the position information "3". The previous part is a proper name with the position information "0", and the other preceding part is "the". In view of this, the term "the Sumida River" is collectively assembled as a single proper name through steps 5250 to 5254 and stored as the position information "3" in the table 5124 ( Fig. 46B). Then, the next reference unit "P" is an unregistered proper name having the positioning information "0" at which processing 5213 is performed. Although the previous word is a proper name because the position information therefor is "3", nothing is collected collectively, but the dictionary information is stored as it is in the table 5124 ( Fig. 46C). Ordinary processing is applied to the subsequent conjunction "and".

The next word "Mr." is a proper name with the position information "1" stored in the table 5124 as it is ( FIG. 46D). Even if the previous part is a proper name "Paul", since punctuation between words may be directly from "Mr." is present as it is in table 5124 .

Further, the word "gold" is a proper name that is not registered in the dictionary and in which the processing 5213 is applied. Because the previous word "Mr." has position information "1", both are collectively collected, and the whole part is formed in a proper name with the position information "1" ( Fig. 46E). Then, the processing at the next word "Smith" is similar ( Fig. 46F). The subsequent word "went" is a past tense of a verb and a common grammatical analysis is performed below.

As explained above with reference to the explained Embodiment is described, proper names thereby arranged that a word is used as the key represents a part of a group of proper names, which collectively grouped into a single proper name, and that undergoes a positional restriction when it does Collectively summarized in a proper name. Even if a number of proper names are constantly present, it is In this way possible, a proper collective compilation along with the context without a flawed one perform collective compilation simply by be copied in a group of proper names. In the aforementioned Example will be the words "Sumida River" ("the Sumida river") analyzed as a group of proper names, separated from the subsequent word "Paul" are. Further, Mr. Gold Smith also turns out to be a group Analyzed proper names.

Referring now to Fig. 47, a sixth embodiment of the speech analyzer having features of the present invention applied to an English / German-Japanese automatic translation device will be described. This embodiment has an input processing section 6014 , and data is input to the section 6014 from an input unit 6012 . The input unit 6012 includes, for example, a keypad with character keys such as alphanumeric and function keys, an optical character reader for reading the English / German text recorded on paper, and a magnetic disk reader.

The input processing section 6014 has a buffer 6014 for a an input character string (word string), and stores the input from the device 6012 English / English sentence in the buffer 6014 a. The section 6014 reads out the sentence stored in the buffer 6014 a and delivers it to a unit partitioning section 6016 . The section 6016 is a functional section that divides the dictionary reference unit from the input sentence of the section 6014 by retrieving a delimitation table 6018 . Table 6018 contains boundaries such as spaces, commas, etc.

The section 6016 reads out the delimitations from the table 6018 and divides the sentence input from the section 6014 into strings as the units for retrieving a reference dictionary 6020 by dividing the sentence into parts where the boundaries exist. The divided strings are input to a dictionary retrieval section 6022 .

The section 6022 retrieves the input sentence reference dictionary 6020 which has been submitted from the section 6016 divided into reference units. For example, as shown in Fig. 48, the reference dictionary 6020 includes entries for the character strings, their language part, feature information, etc. of the English sentence. The reference dictionary 6020 includes, in addition to the proper names shown in the figure , the character strings for another part of speech, for example, verbs, adjectives, etc. The proper name as the part of speech in this figure means that they are used in conjunction with the registered proper-name distribution will be described later, but does not express a common grammatical proper name. Further, the feature information indicates what the proper proper name expresses and is not limited to only one.

The section 6022 retrieves the reference dictionary 6020 for the character string which is divided into the reference unit, and if the character string is a proper name, it delivers it to a section 6024 to perform the proper name processing, which will be described later. Further, if it is not a proper name, it is delivered to a processing section 6036 and is stored in the table 6036 a in the processing section 6036 . The section 6024 has a preceding sentence end processing section 6026 , a preceding proper name processing section 6028, and a self-name processing section 6030 .

In the section 6026 , it is judged whether or not a string preceding the input string to be fetched from the section 6022 is at the end of the sentence, and if the preceding string is at the sentence end, it is delivered to the section 6022 after the capitalized letter at the beginning of the string to be processed is converted to a small letter and the retrieval section is caused to retrieve the reference dictionary 6020 again. The string not fetched at the second fetch is classified as an unregistered proper name, is delivered to the processing section 6036 and stored in its table 6036a . If the string preceding the input string is not at the end of the sentence, it is delivered to the processing section 6036 as a proper name whose feature information is unknown, and is registered in the table 6036a as will be described later.

Section 6028 parses the feature information for the previous string output from section 6026 and returns the result to section 6030 . The self-name processing section 6030 checks the feature information for the proper name to be parsed and, if the feature information is not registered as either the proper name or the preceding proper name, analyzes the own name and the previous proper name using the registered feature information of the other proper name and stores this result in the table 6036a in the processing section 6036 .

The processing section 6036 has the dictionary information preservation table 6036 a, stores the output from the section 6028 or 6032 data in the table 6036 a, and then reads out the thus stored data and outputs it to a syntactic analysis performing portion 6038 from performing the structural analysis for the input sentence, reading it from the table 6036a and subjecting it to morphological analysis.

The operation of the device will be explained with reference to the flowchart shown in FIG . First, a German / English input sentence is read from the input device 6012 into the input processing section 6014 ( 6100 ). The input sentence read in the section 6014 is stored in the buffer 6014 a, and the input sentence stored in the buffer 6014 a is read out to the unit dividing section 6016 .

When the input sentence is read, section 6016 reads out boundaries from the corresponding table 6018 to perform a division for the dictionary reference units ( 6102 ). That is, the strings representing the input sentence input are successively divided into fetch key strings from the beginning thereof as the units retrieving the reference dictionary 6020 by dividing them at the locations where the boundaries such as spaces and colons, available. It is judged whether or not the divided reference units, that is, the fetch key strings, have been completed ( 6104 ), and if there is still a (unfinished) fetch string, it is delivered to the dictionary retrieving section 6022 .

When the fetch string is submitted to section 6022 , it retrieves ( 6022 ) the fetch string reference dictionary 6020 ( 6106 ). It is then judged whether or not the retrieval string is present in the entry of the reference dictionary 6020 as shown in Fig. 48 ( 6108 ), and if there is an entry, a part of speech information stored in the reference dictionary 6020 becomes is read out, and it is judged whether the fetch string is a proper name or not ( 6110 ).

If the fetch string is not a proper name, the portion 6022 outputs the data read from the reference dictionary 6020 to the processing section 6036 and stores it in the table 6036a ( 6112 ). When the data is stored in the table 6036a , an entry indicating that the data is stored and the data for the fetch string stored immediately before are input from the processing section 6036 to the section 6016 . As a result, step 6012 is returned to, and the dictionary reference unit dividing is performed in section 6016 .

At step 6110, if the retrieval string is a proper name, the portion 6022 outputs the proper name read from the reference dictionary 6020 to the dictionary retrieval section together with the data of the previous retrieval string input from the table 6036a by means of the section 6016 6022 and the section 6024 where the processing for the proper name registered in the dictionary is performed ( 6114 ).

Now, with reference to the flowchart shown in Fig. 50, the processing for the proper name registered in the dictionary is performed. The data output from the section 6022 to the section 6024 is output to the previous proper name processing section 6028 by the preceding sentence end processing section 6026 . When processing the proper name registered in the dictionary, section 6026 has no function.

In the section 6028 , it is then judged whether or not the retrieval string precedes the proper name, a proper name not registered in the reference dictionary 6020 , that is, a proper name subjected to the processing for the proper name not registered in the dictionary ( 6200 ), or not, as will be described later. If it is an unregistered proper name, the entire part of the proper name and the previous unregistered proper name are judged as a proper name having the feature information of a proper name ( 6202 ); the data is then output to processing section 6036 and stored in its table 6036a ( 6214 ).

In the processing section 6028 , when the retrieval string preceding the proper name is judged to be a non-registered proper name, it is judged at step 6200 whether or not the retrieval string preceding the proper name is a proper name registered in the reference dictionary 6020 ( 62049 ). If the retrieval string preceding the proper name is a registered proper name, it is judged whether or not the feature information of the previous proper name is unknown, ie, whether it is registered in the reference dictionary 6020 or not ( 6206 ).

If the feature information of the previous proper name is unknown, the flow advances to step 6202 where the entire part of the proper name and the previous proper name are considered as a single proper name having the feature name of the proper name ( 6202 ); the processing section 6028 then outputs data from the processing section 6036 where it is stored in its table 6036 a ( 6214 ).

In the processing section 6028 , if the feature information of the previous proper name is not judged to be unknown, that is, if judged to be registered in the reference dictionary 6020 , the data is output from the processing section 6028 to the section 6030 . In the section 6030 , it is then judged whether the feature information of the proper name is known or not ( 6208 ). In the case where the feature information for the proper name is unknown, the section 6030 judges the whole part of the proper name and the previous proper name as a proper name having the feature information of the previous proper name ( 6210 ) and outputs the data to the section 6036 , in whose table 6036 a they are recorded ( 6214 ).

When it is determined in the section 6030 that the feature information of the proper name is not unknown, that is, registered in the reference dictionary 6020 , the processing section 6030 judges the proper name as a proper name with feature information that is independent of the reference dictionary 6020 preceding proper name ( 6212 ) and outputs the data to the processing section 6036 in whose table 6036a the data is then recorded ( 6214 ). Returning now to Fig. 49, if no fetch string is present at the input of the reference dictionary 6020 at step 6108 , it is judged whether or not the first character in the string is a capitalized letter ( 6116 ); if it is not a capitalized letter, section 6022 assesses the fetch string as an unregistered word, submits it to section 6036 , and stores it in its table 6036a ( 6118 ).

If the first character is a capitalized letter, the data for the fetch string, along with the data for the previous fetch string, is delivered from the portion 6022 to the portion 6024 where the unregistered proper name processing is performed ( 6120 ).

Now, with reference to Fig. 51, the processing for a proper name not registered in the dictionary will be described. The data for the retrieval string is output to the previous sentence end processing section 6026 together with the data for the previous entry recorded in the table, when it is judged whether the end of the previous entry recorded in the table a candidate for the end of the sentence is or is not ( 6300 ). This sentence end candidate judgment is made to judge whether the end of the previous entry recorded in the table is a candidate for the end of the sentence, for example, a separate point (.), Etc. or not.

When the end of the previous entry recorded in the table is a candidate for the end of sentence, data is output from the section 6026 to the previous proper name processing section 6028 , and the section 6028 then judges the previous entry recorded in the table as the End of sentence ( 6302 ) and returns it to the dictionary retrieval section 6022 after the capitalized letter at the beginning of the retrieval string has been changed to a small letter.

The section 6022 retrieves the retrieval string reference dictionary 6020 that has been reconstituted for the small letter ( 6304 ) and judges whether it is an entry in the reference dictionary 6020 ( 3603 ). If there is an entry, the section 6022 outputs the data retrieved from the reference dictionary 6020 to the processing section 6036 and stores it in the table 6036a ( 6308 ). If there is no entry, section 6022 returns to the first character in the fetch string uppercase, returns it as a non-registered proper name to processing section 6036 , and stores it in table 6036a ( 6310 ). If at step 6300, the processing section 6026 judges that the end of the previous data recorded in the table entry is not a candidate for the end of the block, the data from the section 6026 to the processing section 6028 to be dispensed, and the portion 6028 then judges that the in Previous entry not recorded in the table is the end of the sentence ( 6312 ). The data is then passed from section 6028 to section 6030 , and this section 6030 then judges the retrieval string as a proper name whose feature information is unknown ( 6314 ).

The section 6030 then returns to the data in the section 6028 and the processing for the dictionary-registered proper name is performed in the section 6028 ( 6316 ). The processing for the proper name registered in the dictionary is the same as that shown in FIG .

If the dictionary reference units are completed in step 6104 is now in Fig. 49, in the table 6036 a recorded data from the section 6036 to the section 6038 are submitted (6122), whereby the morphological analysis is completed according to this embodiment. The operation of this embodiment, which has been described above, will now be explained with reference to an input sentence.

In the explanation, reference will be made to Fig. 52 where, for example, an introductory sentence "At the Tokyo railway station Mr. Walter. ("At Tokyo Station Mr. Walter ...") is entered. First, an input processing 1100 is performed in which the input sentence is read into the processing section 6014 . Then, the dictionary dividing unit is made ( 6102 ) by dividing the input sentence by intervals into corresponding words. First, the reference dictionary 6020 for "In" is retrieved ( 6106 ). There is no entry for "Im or In" in the reference dictionary 6020 . If the step has advanced to processing for a non-dictionary registered proper name, since the previous portion is recognized as the beginning of the sentence (top of the file), "In or In" is converted to "in". Since "in" has an entry in the reference dictionary 6020 and is not a proper name ( 6110 ), the data retrieved from the reference dictionary 6020 is recorded in the table 6036a ( 6112 ).

Then, the reference dictionary 6020 is retrieved for "Tokyo" ( 6106 ). Since there is no entry for "Tokyo" in the reference dictionary 6020 ( 6108 ) and the first character is a capital letter ( 6116 ), processing for a proper name not registered in the dictionary is performed ( 6120 ). Then, advance to FIG. 51. Since the previous part is "In" and is not a candidate for the end of sentence ( 6300 ), "Im or In" is recognized to be not the end of sentence ( 6312 ); "Tokyo" is recognized as a proper name whose feature information is unknown ( 6314 ), and processing for a dictionary-registered proper name is performed ( 6316 ). Then, advance to FIG. 50. Since the previous "Im or In" is neither a registered proper name nor an unregistered proper name ( 6204 ), "Tokyo" alone is recorded as a proper name with its own information feature, ie, a proper name whose feature information is unknown ( 6216 ), then the next step to Fig. 49 decreased, and the reference dictionary 6020 is retrieved for "station or station" (6106). Since there is an entry for "station" in the reference dictionary 6020 ( 6108 ), and this is a proper name ( 6110 ), processing for a dictionary-registered proper name is performed ( 6114 ). In the next step, Fig. 50 is advanced. Since the previous "Tokyo" is a non-registered proper name ( 6200 ), the entire part for "Tokyo Station" is recorded as a proper name having the feature information "place, in the form of" station "( 6206 ).

Then, the reference dictionary 6020 in Fig. 49 for "Mr." retrieved ( 6016 ). Since there is an entry for "Mr." in the reference dictionary 6020 and it is a proper name ( 6110 ), processing for a dictionary registered proper name is performed ( 6114 ). Then, in the next step, FIG. 50 is advanced. The previous term "station" is an unregistered proper name ( 6200 ), but a registered proper name ( 6200 ) and the feature information "location" is not unknown ( 6206 ). Since "Mr." the feature information "not unknown" ( 6208 ), becomes "Mr." alone registered as a proper name with the feature information "Person" ( 6212 ).

Now, again in Fig. 49, the reference dictionary 6020 for "Walter" is retrieved ( 6016 ). Since "Walter" has an entry in the reference dictionary 6020 ( 6108 ) and it is a proper name ( 6110 ), processing for a dictionary-registered proper name is performed ( 6114 ). Then the next step advances to point 50 . Since the previous "Mr." an unregistered proper name is ( 6200 ), but a registered proper name ( 6204 ) having the feature information for "Person" is not unknown ( 6202 ), while the feature information for "Walter" is unknown ( 6208 ), "Mr. Walter" is composed and recorded as a proper name with the feature information "person" ( 6210 ).

As described above, in this embodiment, the German / English input sentence is divided into fetching strings for which the reference dictionary 6020 is fetched first. If there is an entry in the form of a proper name in the reference dictionary 6020 , a registered proper name processing is performed taking the previous entry recorded in the table into consideration.

If the previous entry recorded in the table is a proper name, the feature information for the previous entry recorded in the table and checked for the current proper name to process. If one of these is missing in the feature information, the feature information of the other of them is provided while if both have feature information, They are individually as proper names with associated feature information recognized.

Consequently, it is possible to have a proper name that does not contain feature information has, appropriately, with feature information to provide, and the intended feature information to limit appropriately. this makes possible a more effective analysis in the subsequent structural analysis and a corresponding translation.

Further, for the character string which is not registered in the reference dictionary 6020 , if the first character is a capital letter and the previous character string is judged to be the sentence end, the reference dictionary 6020 is retrieved after the capital letter has been changed to a lower case, and hence It is also possible to retrieve the reference dictionary 6020 for the string at the beginning of the sentence. Further, when a character string starting with a capital letter appears at a position other than the beginning of the sentence, it is judged to be a proper name, and the feature information of the proper name is provided with a proper name information having registered feature information existing before or after it is. Thus, a proper name not registered in the reference dictionary 6020 can be parsed to some extent grammatically.

Now, a seventh embodiment will be described. Here, Fig. 54 shows the entire structure of the seventh embodiment of a speech analyzer having features of the invention used in an English / Japanese automatic translator. This embodiment has an input slot 7010 through which an English-to-translate English text 7012 is input. The input section 7010 may comprise , for example, a keypad with character keys such as alphanumeric or function keys, an optical character reader (OCR) which reads the English text recorded on paper, and / or a file memory means for reading English text onto a storage medium such as a magnetic disk is recorded.

The input from the input section 7010, English text is read in a Vorredigierabschnitt 7010, in which a translation pretreatment is carried out. In this case, sentence recognition and treatment of unknown words are mainly performed.

This then acts as part of a morphological analysis. The pre-processed English data is transmitted together with the information obtained in the pre-editing in a section 7010 for a morphological analysis. The section 7010 dissects the morphemes of the English sentence by dividing them by retrieving a word dictionary 7018 , performs various kinds of arrangements or compilations by the processing of unknown words, expression for a proper name, time, number, etc., and carries one Processing for the whole sentence, such as a supplemental question and apposition recognition by. The morphological analysis rules for this are contained in a rule file 7036 .

The English data which has been subjected to the morphological analysis, together with the dictionary information obtained by the morphological analysis, is transferred to a dedicated section I 7020 . Section 7020 is a functional section that parses the surface layer structure for the sentence, applying grammatical rules to English data, and finds all structural possibilities. The English data analyzed in section I 7020 is then sent along with the analysis information to syntactic analysis section II 7020 where a solution is applied to the result of the previous analysis corresponding to the surface layer structure in section I a structural description is selected. Consequently, a plausible parsing or analysis tree of the English sentence is prepared, and the structure is then made. The analysis rules are stored in rules file 7036 .

The English data is transmitted to a texture conversion section 7024 after performing the syntactic analysis as the data for the parsing tree . This ( 7024 ) prepares a corresponding Japanese structure tree from the structure tree , which is an English intermediate structure, and converts it into a Japanese underlying structure, from which a Japanese sentence can be easily translated.

The structure tree data showing the Japanese underlying converted structure is delivered to a translation section 7062 where a translated sentence is generated. This is a functional section for generating a Japanese sentence from the Japanese structure tree. The data for the Japanese, which are formed as a translated sentence, that is, the translated sentence information is outputted to a Nachre 87040 00070 552 00004 001000280000000200012000285918692900040 0002003733674 86921digierabschnitt 7030th This section 7030 modifies the translation sentence data relative to the dictionary 7018 using information used in the translation to perfect a more natural Japanese sentence. The data for the Japanese sentence is transferred to an output section 7032 and then output therefrom as a translated Japanese sentence 7034 . The dispensing section 7032 may include, for example, a printer, a display, and / or a memory file device, such as a magnetic disk. The flow of a series of translation processes is controlled by a control section 7038 , which controls the control of the entire device.

In the word dictionary 7018 , in this embodiment, the dictionary data for English and Japanese words are stored. A vocabulary is defined, as well as a connecting relationship, ie a co-existing relationship or various information, such as meanings of a singular or plural form, a speech part, etc. Further, rule data for morphological and syntactic analysis is stored in rules file 7036 . The control section 7038 is connected to the operation display section 7040 , which in turn has operation buttons for issuing various commands from an operator to the device, such as a translation command key and a cursor key, as well as a display or display visually representing the input Japanese Sentence text, the Japanese sentence as a result of a translation, intermediate data such as a dictionary information indicating various information to the operator.

Here, most of the operation display functions may be included in a keyboard when located at the input section 7010 or in a display when disposed at the output section 7032 .

In Fig. 53, detailed explanations for the morphological analysis performing section 7016 are shown with respect to the processing of numbers. These parts, which directly contribute to the understanding of the invention, are presented, although, of course, the morphological analysis section 7016 also has other functional sections for morphological analysis. The section 1017 has an input processing section 7100 for receiving and processing the input string data input from a pre-editing section 7014 . The section 7100 is provided with a buffer into which data for the English character string has been input in the form of code data such as ASCII, and temporarily stores the data for the character string.

The input character string data temporarily stored in the processing section 7100 is delivered to a section 7102 which divides the data into dictionary reference units such as words. The section 7102 is a functional section that distinguishes a dictionary reference unit that successively displays the string in retrieving the dictionary 7016 in the section 7106 . Dictionary reference delineations used in the dictionary reference unit dividing processing are arranged in the place of an English character, a numeric character, an apostrophe, non-dash characters, and an apostrophe following a space , They are stored in a delimitation table 7104 and are referred to in dividing the dictionary reference unit in section 7108 .

Dictionary 7018 contains information particularly for retrieving the split units. Further, the dictionary 7018 stores morpheme processing information such as the name of a month, a day of the week, a cardinal number representing only a numerical value, an ordinal number, a unit for expressing grams, and the like. ä. time, the, of, comma (,). (.) etc.

The dictionary retrieval section 7106 is a function section that retrieves the dictionary 7018 to extract the dictionary information based on the character string input from the section 7102 , and transmits it to a morpheme processing information creating section 7108 . Section 7108 contains Morpheme processing information (see Figure 56) indicating that a string having a morphemic feature has a time meaning, such as hour, year, month, etc., providing further specific information in the recognized string is to obtain a cardinal number or a time meaning in the dictionary retrieval section 7016 . For example, such information is provided, which means a numerical figure, ie a number "yes". The string provided with information in section 7108 is also applied in necessary local grammatical analysis.

In this case, a unit group of the dictionary Reference unit, such as a word, which by the Morpheme actuation information has been actuated with Help the local grammatical dissection or Parsing rules are collectively composed. For example "Month name", "numeric expression" are displayed in "month name + Numeric Expression ", i.e." Oct. "and "18" are grouped to "Oct. 18". There will also be one made collective compilation, such as "November the 2nd" for "month name + the + numeric Expression "," 22 March "for" numeric expression + month name ", "the 23rd May" for "the + numeric expression + Month name "," the 11th of June "for" the + numeric expression + of + month name "," '86, Jan. 27th month "for" year +, plus month and day +, + weekday ", Sunday 26, Jan., 1986 "for weekday +, + month and day +, + year", "11: 30 a.m." for "number: number + a m (or p m)" or "Month name + year" "month name + of + year" etc.

The local analysis processing is performed in an initial value setting section 7110 , an adaptation retrieval section 7112 , a unit dividing section 7114 , a morpheme processing information creating section 7111 , interrogation sections 7116 and 7120, and processing sections 7122 and 7124, and a fitting table 7128 includes a morpheme processing display table which is a discrimination reference table for discriminating that a unit sequence indicating a number and time factors as shown in Fig. 57 is a composite unit of time factors according to specific rules. The section 7110 sets the initial value of a counter n, which counts the number of dictionary reference units, to match, on retrieval, a sequence of dictionary reference units as a unit group described above in the adaptation query section 7112 .

Section 7112 calls the customization table 7128 for each of the dictionary reference units to make an adjustment. The unit partitioning section 7114 discriminates the dictionary reference units assumed to be "p", which has been completed with the dictionary queries in the section 7106 from the dictionary reference units, which represent character strings after the dictionary reference units of the dictionary query by the counter n have been completed.

The query section 7116 is a function section having a function similar to the dictionary retrieval section 6106 which retrieves the dictionary 7018 to take out the dictionary information based on the character string discriminated in the section 7114 and the same at the morpheme Processing information creating section 7118 transmits. The section 7118 has the same function as the section 7108 , with further specific information provided on each information which has been recognized as an ordinal number or a time factor in the polling section 7116 .

The query section 7120 and the processing sections 7122 and 7124 collectively synthesize a sequence of dictionary reference units up to " p + n " obtained from the adaptation retrieving section 7112 by the processing in the section 7118 into a dictionary reference unit. Then, the result is stored in the table 7126 , which is a buffer for storing the dictionary information completed by the fetching. The result of the morphological analysis is transferred from the table 7126 to the section I 7020 for syntactic analysis.

Now, the collective composition by means of the morpheme processing information will be explained with reference to the flowcharts shown in Figs. 55A and 55B. For example, the following character string should be input to the input processing section 7100 ( 7300 ).
Entered string: "26 Jan, '80 he."
The section 7102 divides the input character string by the dictionary reference unit for retrieving the dictionary 7018. ( 7302 ) "26" in the input character string is separated by separating the dictionary reference unit as a unit. It is now judged whether the separated part of the reference unit for the input character string has been finished or not; if it is finished, the operation is ended ( 7304 ), while if it is not, the flowchart advances to the following step 7306 .

The dictionary 7018 is fetched for "26" in the input character string to take out the dictionary information indicating that "26" is a "number" ( 7306 ). Then, morpheme processing information indicates that "number, number" is a morpheme feature, ie it is a sequence of numbers and is treated as a grouped cardinal number ( 7308 ). It is now judged whether or not the group receiving the dictionary information has been provided with the morpheme processing information at step 7308 ( 7130 ). If provided, the flowchart proceeds to step 7314 to apply further processing based on the local analysis rule, and no designated group is recorded in table 7126 ( 7312 ); in the flow chart, the process returns to step 7302 . Consequently, at step 7314, it advances to "26" since this number is provided with the morpheme processing information. The processing in step 7314 is performed according to the operation of the flowchart shown in Fig. 55B.

First, an initial value "0" is set in a counter n which counts the number of matching reference units when the reference units in the adjustment retrieving section 7112 are retrieved ( 7410 ). Further, since the reference unit completed by the dictionary retrieval in the section 7010 is set as "p", the matching table 7128 for the p + n _th (n = 0) reference unit, ie, "26", becomes the section 7112 retrieved ( 7412 ). Since "26" has been provided by means of the morpheme processing information indicating at step 7308 that it is an ordinal number, and since these compilations have "each with an ordinal number" at its upper end and after the second term the compilation in the Matching table 7128 (see FIG. 57), the reference unit "26" is the same as the information of the fitting table 7128 and thus adapted to it. In this case, an adjustment is made for Ms-Me while the second matched set is set as "Ms", with the last data of the combination having the "ordinal" at its top than "me" in the lookup table 7128 .

Based on the fit result in the table 7128 at a (p + n _te (n = 0) dictionary reference unit, the match state is judged ( 7414 ) and if it is determined to be matched, the flowchart advances to step 7416 while if it is determined that it is not matched, the flowchart advances to step 7424 .

If it is judged that to be adjusted, "1" is set in the counter n to perform the separation for the dictionary separation unit at p + 1 _te (n = 1) in the input character string. The separation is performed in the same manner as in step 7302 . The dictionary 7018 is retrieved for "Jan.," in the input string, since this has been deposited as the dictionary reference unit closest to "26" for processing to provide the morpheme processing information ( 7420, 7422) ). This processing is performed in the same manner as steps 7306 and 7308 .

By repeating the operations from step 7412 to step 7422 , the flowchart is executed until "Jan 26, '80 he" closed. However, since "he" does not match the adjustment table 7128 at step 4712 , the flowchart advances to step 7424 at step 7415 . This means that although the data are matched by 26th Jan, '80 'with' cardinal number, month, year 'in the match table 7128 , they are not matched for '26, Jan., '80'.

Further, if the sentence is terminated on the input character string, for example, "Jan. 26, '80 '", ie, there is no next discard for the next dictionary reference unit, the flow advances to step 7424 in step 7418 . In the event that it is determined that is not matched at step 7414 , it is now judged whether or not the content of the counter n is more than 1 ( 7424 ), and if it is not more than 1, this is considered a single one Reference unit recorded in table 7126 ( 7434 ).

If it is not less than 1, an adjustment is made by taking the p + n (n = 3), ie, "he" in "6 Jan," 80 he "is" EOS ", thus ending the compilation is displayed ( 7426, 7428 ). If it is not matched, the flowchart advances to step 7434 . If it is matched, "26 Jan., '80", which is p- (p + n-1) for the dictionary reference unit, is collectively assembled according to the compilation result, which corresponds to the compilation Ms in the fitting table 7128 , and Result is recorded in table 7126 ( 7430 ). It is then noted that the reference units have been terminated at the (p + n-1) th unit and (p + n -1) is reset to "p" ( 7432 ).

Now, the eighth embodiment will be explained. Fig. 63 shows the entire structure of the eighth embodiment of a speech analyzer having features of the invention applied to an English / Japanese automatic translator.

In Fig. 63, there are shown an input section 8801 , an English text 8002 , a pre-editing section 8003 , a morphological analysis section 8004 , a syntactic analysis section I 8005 , a syntactic sentence analysis section Ii 8006 , an operation display section 8007 , a word dictionary 8008 , a rule file 8009 , a control section 8010 , a texture conversion section 8011 , a translated sentence generation section 8012 , a post-editing section 8013 , an output section 8014, and a Japanese sentence 8015 . This translator, as shown in Fig. 63, has the input section 8001 through which the English text 8002 to be translated into Japanese is input. The input section 8001 may include , for example, a keypad with character keys such as alphanumeric and function keys, an optical character reader (OCR) for reading the English text and / or a file memory means for reading the English text recorded in a storage medium such as a magnetic disk ,

The English text entered in the section 8001 is read into the pre-editing section 8003 , with pretreatment for the translation being performed. In this case, mainly sentence recognition and unknown word processing are performed. This acts as part of the morphological analysis.

The pre-processed English data will be provided to section 8004 together with the information received during pre-editing. The section 8004 divides the data by referring to the word dictionary 8008 , decomposes the English morphemes, performs various types of merges such as processing unknown words, expressing a proper name, time and numbers, and performs processing for the entire sentence through, such as a supplemental question and an apposition recognition. The rules for a morphological analysis are included in rule file 8009 .

The English data, after being subjected to the morphological analysis, is delivered to the section I 8005 together with the dictionary information obtained by the morphological analysis. Section 1 8005 is a functional section that decomposes the surface-layer structure for the sentences by applying a grammatical rule to the English data, and this finds all the structural possibilities.

The English data, after being subjected to the syntactic analysis in the section I 8005 , is given to the section II 8006 together with the corresponding information, where a solution from the syntactic analysis result with respect to the surface layer by the syntactic analysis I is selected by applying a structural description. Consequently, a plausible so-called parsing or analysis tree of English description is prepared to form its structure. These so-called parsing or parsing rules are also included in rule file 8009 .

The English data which has been subjected to the syntactic analysis is transmitted as the data for the corresponding analysis tree to a texture conversion section 8011 . The section 8011 prepares a corresponding Japanese structure tree from the structure tree, which is an intermediate structure of the English sentence, to convert it into a Japanese underlying structure, from which the Japanese sentence can then be easily translated.

The data for the structure tree indicating the Japanese underlying structure that has been thus reformed is delivered to a translation section 8012 in which the translated sentence is generated. This is a function section for generating a Japanese sentence from the structure of the structure tree corresponding to the Japanese sentence.

The Japanese sentence data thus translated, that is, the data for the Japanese sentence, are delivered to the pre-editing section 8013 , which modifies the translated data for fetching the memory 8008 using the information used in the translation to obtain a more natural Japanese sentence perfect. The data for the Japanese sentence is transferred to the output section 8014 and output as the translated Japanese sentence 8015 from the output section 8014 . The output section 8014 includes a printer, a display, and / or a file storage device, such as a magnetic disk.

The flow of a series of translation operations is controlled by the control section 8010 , which controls the control for the entire device. The word dictionary 8008 includes dictionary data for English and Japanese words in the case of the illustrated embodiment, and also assigns a language part, etc., to the vocabulary, such as a connecting relationship, ie, a co-existing relationship, meanings, singular and plural forms. The rule file 8009 also contains rule data for the morphological analysis and for an English sentence decomposition.

The control section 8010 is connected to the operation display section 8007 which has operation keys for giving various commands from an operator to the device, such as a translation command key, a cursor key, etc. a display or display visually representing the input English sentence text Japanese sentence as a result of translation, intermediate data such as dictionary information and various commands to the operator. It can be designed so that most of these operation display functions are present in the keyboard if it is provided in the input section 8001 or present on the display if it is provided in the output section 8014 .

The illustrated embodiment relates to the automatic translation device, and it is used so that when a derived word is included in the English text 8002 , a grammatical feature, a semantic feature, the (Japanese) equivalent, etc., depending on the conditions for evaluating the derived word recognized by, for example, an addition, thereby increasing the reliability of the obtained decomposition result or the result of the translation. An Affix dictionary is used for the evaluation of dictionary information on unknown words in morphological analysis. The processing mode requires three kinds of processing, that is, prefix processing, prefix for the suffix, and the evaluated processing by the suffix. However, the data types essentially contain two types, ie, prefix and suffix rating data.

First, the prefix and suffix rating data become explained in detail.

(1) evaluation data prefix

If the prefix or part of a prefix is not in the dictionary registered word with the one described below (Part) matches and the remainder in the dictionary is present, the word becomes according to the dictionary data treated for its root. In the dictionary data is It is possible to add an internal feature to the series of the original one to add internal feature and a Japanese one Add suffix to the original translation word. For the word "electrochemical", for example the input "electrochemical" in the dictionary information Conservation table according to the dictionary data by the prefix entry "elektro" and the dictionary entry "chemically" formed as shown below is:

Präfixeintrag Dictionary entry electro chemical entered word Dictionary information table entry electrochemically electrochemically

The entry in the table then takes over all the dictionary Information to which the words correspond.

(2) suffix evaluation data

If the suffix or suffix part of a not in the dictionary registered word with the one described below (Part) matches and the rest of the part in the dictionary is present, this word is registered by new ones Dictionary data are created according to the information which is described in the suffix dictionary. In this Case becomes the first (Japanese) equivalent in the dictionary data, which correspond to the root part of the word taken out and for the (Japanese) equivalent in the new dictionary data used.

For example, for the entered word "controler"  an entry in the "controler" table based on the Suffix entry "- (e) r" and the dictionary entry "control" registered.

suffix entry Dictionary entry (Verb) - (e) R control noun verb input word Dictionary information table entry controler controler noun

Now the outline for the processing of a derived and the processing of an unknown word explained.

• (1) When for a word not registered in the dictionary a prefix or suffix at the beginning or end of a Word is included and the remaining part of the word in registered in the dictionary becomes an English part a language information, an internal feature and the Japanese Equivalent based on the dictionary and the Affix information composed.
• (2) Prefix and suffix are listed as a movie system and can be edited independently of a program.
• (3) First the possibility for the prefix is tried and if that is not possible, the possibility for the suffix tries. If both are included, will no attempt was made.
• (4) For the word that failed in the trial evaluation is processing for the end part as a unknown word performed.

In particular, an eighth embodiment will be explained with reference to a block diagram in FIG. 58. In Fig. 58, an input processing section 8020 , an input partitioning section 8021 , a delimitation table 8022 , a dictionary retrieving section 8026 , a derivation processing section 8023 , a reference dictionary 8024, and a dictionary information preserving table 8025 are illustrated . First, an English sentence is written in the input processing section from an input device having an original input file or a keyboard, an ORC and so on. Then, the dictionary reference unit is divided into the dedicated section, referring to the delimitation table, and if not to be finished, a dictionary attempt is performed using the reference dictionary. If there is an entry as a result of the experiment, the trial result is recorded in the dictionary information preservation table while a derivative word processing is performed if there is no entry.

Fig. 59 is a block diagram for explaining an embodiment of the derivative processing by means of a prefix. In Fig. 49, there are shown an adaptation section 8030 between the upper part and the prefix dictionary, a prefix dictionary 8031 , a dictionary retrieval section 8030 except for the prefix part, a speech part adaptation section 8033 in the prefix dictionary and the dictionary part of the entry, a dictionary information preparation section 8034 by the prefix evaluation, and a derivation processing section 8035 by the prefix.

Fig. 60 is a block diagram for explaining an embodiment of the derivation processing by means of a suffix. In Fig. 60, there are shown an adaptation section 8040 between the end part and the suffix dictionary, a suffix dictionary 8041 , a processing section 8042 for completing a (completely unknown) unregistered word, a dictionary prefetch section 8043 around the dictionary with respect to the part except the suffix Retrieve an adaptation section 8044 between the speech part of the root and the entry speech part in the suffix, a dictionary information preparing processing section 8045 with a suffix evaluation, a dictionary retrieval section 8046 for performing dictionary retrieval by adding the root change in the suffix to the dictionary Part except for the suffix part and unregistered word processing part 8074 by means of a suffix appraisal. First, an adjustment is made between the end part and the suffix dictionary. If it is not matched, the word is processed as a completely unregistered word, while if it is matched, the dictionary is retrieved for the part except for the suffix part. If there is no entry as a result of the dictionary retrieval, a dictionary retrieval is performed while adding the root change in the suffix dictionary to a part except for the suffix part. If, as a result, there is no entry, processing for an unregistered word is performed by means of the suffix evaluation. On the other hand, when there is an entry, matching between the speech part of the entry and the speech part of the root in the suffix dictionary is performed in the same way as in the case that the entry is the result of the dictionary retrieval for the part except for the suffix part is present. If it is matched, dictionary information preparation processing is performed by means of a suffix evaluation, whereas if it is not matched, unmatched word processing is performed by means of a suffix evaluation.

Fig. 61 is a block diagram showing details of the entire structure obtained by assembling the parts of Figs. 58 to 60; Fig. 62 illustrates the details of the completely unregistered word processing section 8052 shown in Fig. 61. The completely unregistered word processing section 8042 shown in FIG. 61 has a tried-and-true information preparation processing section 8050 for a noun and a corresponding section 8051 for a verb. However, since each of the parts in Figs. 61 and 62 has already been described in detail, they need not be explained again.

A ninth embodiment with features according to the invention will now be described. Here, Fig. 64 shows the entire structure of the ninth embodiment applied to an English-Japanese automatic translation device. This embodiment has an input section 9010 through which an English text 9012 to be translated into Japanese is input. The input section 9010 may comprise a keypad with character keys such as alphanumeric or function keys, an optical character reader (ORC) for reading the English text recorded on paper, and / or a file memory means for reading the English text onto a storage medium such as a magnetic disk , is recorded.

The English text input from the section 9010 is read into a pre-editing section 9014 in which the translation pre-treatment is performed. In this case, mainly sentence recognition and unknown word processing are performed. This then acts as part of a morphological analysis.

The pre-processed English data, together with the information obtained in the pre-editing, is transferred to a section 9016 for morphological analysis. The section 9016 divides the phrase for retrieving a word dictionary 9018 by dissecting the English morpheme, performs various arrangements such as unknown word processing, proper name processing, expression for time, number, etc., and also performs a Processing for the entire sentence such as a supplemental question and apposition recognition by. The rules for a morphological analysis are contained in a rule file 9036 .

The English data which has been subjected to the morphological analysis, together with the dictionary information obtained from the morphological analysis, is transmitted to a section I 9020 for syntactic analysis. The section I 9020 in this embodiment is a functional section which dissects the surface layer structure from bottom to top and right to left for the sentence by applying the CFG rule to the English data, and this finds all structural possibilities.

The English data, after having been parsed in section I 9020 , is submitted, together with the analyzed information, to a section II 9022 for syntactic analysis. The section selects a solution to the result of the syntactic analysis with respect to the surface layer structure in the section I by applying a structural description. In this way, a plausible parsing tree for an English sentence is prepared to form its structure. These so-called parsing rules are also stored in the rule file 9036 . The English data that has been subjected to the syntactic analysis is transmitted as the data for the so-called parsing tree to the texture conversion section 9024 . In section 9024 , the so-called parsing tree for the corresponding Japanese sentence is prepared for a structural tree, which is an intermediate structure of an English sentence and transformed into a Japanese underlying structure from which a Japanese sentence can then be easily translated. The data for the structural tree indicating the Japanese underlying structure that has been converted in this manner are delivered to a translation section 9026 in which the translated sentence is generated. This is a function to create a Japanese sentence from a tree of the Japanese structure tree. First, a sentence structure is created by changing the structure tree by exchanging the order so as to conform to the structure of Japanese, and then morphemic generation is performed to produce a translated sentence in the top-down and left-to-right form to generate the sentence structure tree.

The data for the Japanese structure tree thus generated is translation data which is output to a post-editing section 9030 in which the translation data is modified by retrieving the dictionary 9018 using the information used in the translation processing to be more natural complete Japanese sentence. The data for the Japanese sentence are transferred to an output section 9032 and then output as the translated Japanese sentence 9034 from the output section 9032 made. The output section 9032 includes, for example, a printer, a display and / or a file storage device such as a magnetic disk. The flow of a series of translation operations is controlled by a control section 9038 which controls the control of the entire device. The dictionary file 9018 stores dictionary data for English and Japanese words, and the rule file 9036 stores data for morphological and syntactic analysis in this embodiment.

The control section 9038 is connected to an operation display section 9040 which has operation buttons for issuing various commands from an operator to the device, for example, a translation command key or a cursor key, and a display or display which visually displays an input English text. display a Japanese sentence as the translation result, intermediate data such as dictionary information, and various commands for the operator. It may be configured so that most of the operation display functions are contained in a keyboard when it is placed on the input section 9010 or in a display when it is placed on the output section 9032 . In Syntactic Analysis Section I, the cfg rule is applied to the English sentence from top to bottom and from right to left for the English data after morphological analysis to derive all possible solutions to the sentence structure. The solutions are generally to be understood in the form of a tree structure. This shows a relationship for words or groups contained in each of the sentences which are related to each other in a subordinate or co-existing relationship, such as a modifying relationship or case relationship, a subordinate relationship, such as between parent, child and grandchild, and so on. in relationship. Each of the words or group is located at the node of the structure tree.

In this embodiment, the syntactic analysis Features that honor the form and vocabularies a sentence distinguished to the collective compilation with regard to the sentence structure. The Compilation in terms of the sentence structure will be here referred to as "unit" and "block".

The "unity" is a word or group of words, which is the smallest unit for the translation process form, which is identical to a word in the syntactic Analysis is treated, and the dictionary information  for each of the lawful elements contained therein is not used.

A "block" is a structural combination in which the syntactic analysis preferably for its interior is performed with respect to the outer part and which in an equivalent way as a unit with respect to them outer part is treated. For example, it can be a Phrase, a phrase, etc. as well as what the intermediate symbols, which is used in the cfg grammar have been. It can also have a nest structure, i. H. on Block may further include another block. Furthermore, can the notion of the block also a sentence, a paragraph, and whole Include sentences, each considered as a block can be. The processing, which is the partial syntactic analysis is given preference here as one partial dissection or analysis "designated". hereby can be uneconomical structural solutions that be reduced to the analysis efficiency to improve, and a more believable To obtain analysis result.

For the block in this embodiment, two features Are defined. One of them is a symbol in the cfg rule, referred to in the present specification as "target" which is composed as the result of the analysis is that in each of the constitutional elements inside of the block has been performed, d. H. a symbol that describes the structure or property of the block. The other is a symbol in the cfg rule that acts as a "Roll", which is carried by the block, if the syntactic analysis at the exterior of the block for the sentence, the phrase or the phrase, in which the block is contained. Ie. it is a symbol that the relationship or the role of the block to the others describes.  

For example, in the case of an English sentence I "White House is not white" the goal a sentence, and the Role is a noun (a phrase). Although the goal and the role is generally identical in most cases are, as in this case, sometimes different from each other.

When the structural arrangement of the input English sentence is recognized as a block for the embodiment shown in Fig. 62, the functional sections for evaluating their destination and their role are grouped in the structure as shown in Fig. 65. As can be seen from Fig. 65, the structural arrangement of the English sentence types pre-edited in the section 9014 is discriminated in the morphological analysis section 9016 using the word dictionary 9018 and the rule file 9036 .

The word dictionary 9018 stores dictionary information for English words and phrases. For example, as shown in Fig. 68, entries each having variations for each word are formed in this embodiment, and all the information is developed. For the part-of-speech information, for example, a number of part-of-speech information may be provided, as shown in FIG . The way to form dictionary 9018 is not limited to this example only.

The rule file 9036 stores these upper state data indicating the upper end of the block, the End state indicating the end, and block preparation information for creating the block with the destination and the role in the form of a table. An example of this is shown in FIG. 69. For example, a block is started by "conjunction" and ends with the end of the sentence. Thus, a block is formed starting from the top with "," which immediately precedes their conjunction, and the destination is a clause, while the role is a sentence. Further, another block is formed, starting from the conjunction to the end of the sentence, in which both the target and the role are a clause.

Further, a block is started by "relative pronoun" and is terminated by "," or at the end of the sentence. As in this case, the possibility for a number of end conditions for an upper condition is allowed. In the event that the block is terminated by "," a cluster ("preceded by the relative pronoun until the next occurrence") forms a block in which the target is a clause Clause and the role is an adverb or an adjective. This means that the entity acts as an adverb or adjective clause. In the case of an ending at the end of the sentence, an entity "," which precedes the relative pronoun at the end of the sentence, represents a block in which the target is a clause, and the role is an adverb or adjective. This is according to the conditions for forming a group, clause, or phrase that appear in standard modern English sentences. In the figure , the symbol "" represents a space. In the morphological analysis section 9016 , the English text input from the pre-editing section 9014 is first divided into sets as the translation units. In this case, spurious spelling or unregistered words are detected. The 9018 dictionary is retrieved at the respective sentence units and the dictionary information for each of the components is retrieved. Various arrangement modes are performed according to this dictionary information.

Fig. 66 shows a flow chart for the collective compilation of a block performed in the section 9016 for morphological analysis. First, a position pointer indicating the reading position for English sentence is set at the upper end ( 9100 ). The upper end does not mean the word at the top, but the upper (imaginary) end of the preceding sentence. The word extraction 9101 is executed at this point. As shown in Fig. 67, in the process 9101, a word is taken out by advancing the position by one ( 9111 ) if it is not at the end of the sentence ( 9110 ), and the dictionary 9018 is fetched for the word ( 9112 ). to write out the word information ( 9113 ).

In this way, when the word information is taken out in the processing operation 9101 , the start and end conditions table 9036 is referenced to judge whether or not there is anything matching the upper condition ( 9102 ). , In this way, steps 9101 and 9102 are repeated until the word matching the upper condition is detected.

If the above condition is satisfied , the next word and the subsequent words are taken out by the required number and reference is made to the match with the upper state of the block ( 9104 ). In this case, the dictionary is retrieved, if necessary, for each of the words. The position indicator is not advanced.

When it is compared with the upper state of the block at step 9104 , a word matching the block end state with respect to the upper state is retrieved ( 9105 ). Steps 9104 through 9106 are again run through until the word that matches the end condition is found. If a word matches the end condition ( 9106 ), a so-called cluster that includes the word is recognized as a block, and the block is written ( 9107 ). Specifically, a block is prepared to judge whether the block preparation condition at the location where the end condition is satisfied first is satisfied. With regard to the block preparation information table 9036 , the position for the word indicated by the pointer at the position where its advance in processing 9103 was stopped is set as the upper position for the block, and the position of the block Word that satisfies the end condition that appears first after that is set as the end position for the block. At the same time the goal and the role of the block are written.

If, as a result of such block recognition, "..., conjunction ...." For example, in the English sentence, as shown in Fig. 70, the formation from the top of the sentence is recognized as a block at the part preceding "," while the formation is "from, conjunction" at the end of the sentence Sentence is recognized as another block. In the figure , the inside of [] indicates a block. In this block, both the goal and the role are sentences. Further, an entity from the word after the conjunction at the end of the sentence forms another block in which both the goal and the role are sentences. On the other hand, the construct from the conjunction at the end of the sentence can be defined as a block. In this case, the target is a clause and a clause, and the role is an adverb.

The block can also be set as it is from the Position off that does not contain "," begins. Furthermore, the Punctuation u. ä. from the subject for the grammatical Analysis as the information is excluded which is present in the block. If in the same way ". . . Relative pronoun . . ." can exist, a "relative pronoun. be recognized as a block. In this block the goal is a clause or a sentence and the role is an adverb or an adjective.  

Of course, the block can also be executed in a nest structure. For example, if the English sentence has such a construction as "(beginning of sentence" ..., conjunctions, ..., relative pronoun ... ") end of the sentence, as shown in, for example, Fig. 71, then the figure of Fig ", Conjunction" until the end of the sentence a block BL1-BL1, in which, "relative pronoun. , ., "as another block BL2-BL2 is included.

In this way, the morphological analysis section 9016 discriminates the feature of the sentence in terms of the shape and the vocabulary to distinguish the structural composition as a block. Further, in addition to such block recognition, section 9016 performs various processing such as a proper name phrase, a derived unknown word, an abbreviated word, a numeric, a time, a dashed word, an apostrophe ('), and so on an apposition evaluation and additional question processing to prepare morphological analysis data.

The English sentence which has been subjected to morphological analysis in this way is transmitted to the syntactic analysis section I 9020 together with the analyzed information. Fig. 72 shows an example of the output data. Fig. 72 shows the result that an English sentence I "White House is not white" is input from the input section 9010 and is morphologically analyzed in the section 9016 . The block 1 is started at word position # 4 and terminated at position # 10, in which case both the target and the reel are free. In this way, block 2 is started at position # 5 and terminated at position # 6, where the target is a main-word group, while the role is a proper name. That is, the block "White House is'nt white." contains White House as a nest in another block. In a block, ie a smaller block "White House", each of the constituents internally functions as a proper name while having a position as a word clause relative to the exterior, ie, "is'nt white." , "White House" can be treated as a single entity.

Along with such block information, word information retrieved from the word dictionary 9018 is added and output from the morphological analysis section 9016 to the syntactic analysis section I 9020 . Section I 9020 analyzes the surface-layer structure of the English sentence by applying a context-free grammar rule stored in rules file 9036 to find all possible structural trees. In this case, if a block is included, the partial grammatical analysis described above is performed while referring to the local analysis. This can improve the efficiency and accuracy of the respective analysis.

In particular, the block inclusion relation becomes from the position information prepared for the block. Then the innermost block analyzed. The block with this grammatical Analysis has ended, is considered a unit and there will be no further processing in it Performed inside. In this way the area of the grammatical analysis gradually to the outside Blocks enlarged. Finally, the whole sentence is analyzed. The respective analysis is based on the cfg- Usually in the way from top to bottom and from right to done on the left in the English sentence. The grammatical Analyze is performed in a manner in which all possibilities are preserved, which by the grammatical rules allowed are.

Fig. 73 shows an example of such an analysis processing flow. First, based on the English data supplied to the section I 9020 , all the structural compositions for a sentence are recognized as blocks, and the objective and role are evaluated ( 9120 ). The way to a compilation is as shown in FIG . Then, if there is no block in such an array ( 9121 ), the sentence is parsed ( 9125 ) and only one sentence collectively assembled as a symbol is selected and the analysis for the sentence is ended ( 9126 ). Processing steps 9125 and 9126 are included in processing steps 9121 to 9124 .

If there is a block, the innermost block is parsed first ( 9122 ). In the example shown in Fig. 71, the inside of the block BL2-BL2 is analyzed. Although various solutions are generally obtained by the grammatical analysis, a solution in which the block is collectively assembled as a cfg symbol and which coincides with the destination of the block is selected ( 9123 ). In this case, with respect to the blanking target for the block, all that is arranged in a symbol is selected. Anything selected in this way is then treated as a single array having the role for the block ( 9124 ). In the block with the optional roll, a roll of the symbol selected in processing 9123 is defined as the roll. The processing operations 9121 to 9124 are then repeated one after another.

In this way, in the example of Fig. 71, the inside of the block BL2-BL2 is parsed first and then the inside of the block BL1-BL1 is parsed. In this case, the block BL2-BL2 is treated the same as a single word, and each of the constituents therein is not parsed.

When held in this manner, the data defining the structural composition and the subordinate relationship are delivered to the syntactic analysis section II 9022 . The data can then be easily recognized in the form of a tree structure, as described above. The data is then further transformed into the structure of the Japanese sentence in the texture conversion section 9024 and in the translation section 9026 ; a translated sentence is then generated for each of the nodes contained therein. The node processing in the structure is performed in the manner of from top to bottom and from left to right.

The set so generated, translated is then subjected to post-processing in the Nachredigierabschnitt 9030, is visually displayed on the display section 9040-service printer, for example, when a Japanese sentence 9034 in the output section 9032nd

In this way, according to this embodiment, the feature of the English sentence with regard to the form and the Vocabulary distinction, so the structural composition to distinguish as a block. For the block become a target, which can be the analysis result, and the structural role with which the block faces outward acts, evaluated. Then the surface layer structure becomes of the English sentence analyzed by a context-free Grammar rule is applied to all possible Figure out structure trees. This allows the number of Inefficient solutions reduced, the efficiency at the grammatical analysis improves as well created a more reliable analysis result.

Thus, there are different patterns for a to be added Expression, and it is difficult to do this in the grammatical Analysis, in particular for a context-free analysis to recognize, since it is in view of the above embodiment  in general, it is difficult to apposition recognition after to carry out the analysis is an ambiguous translation inevitable. Even if a rule has been prepared with which they are recognizable, it would be a risk that no appositional expression to be added as an identical one Case is detected, or the number of possible combinations is called. That means it will be a costly one local analysis between the parts, which in the appositional expression are included, and the other parts carried out.

In view of this, in the present embodiment alleviated the load during processing in the analysis step be by the appositional expression through the Feature of the sentence in terms of form or semantic Characteristics of words is detected. An evaluation of Apposition is performed by the next pattern as a block is detected.

For the English sentence structure "~, relative pronoun, ~.~" the relative pronoun is recognized by the fact that the speech sub-code for the word with a special code, for example "R" is provided. In this case, the interior that is with "," is surrounded, considered as a block on condition that that he is not with a block or a unit overlaps what has been displayed in the pre-editing is, and he does not "and" or "or" in the part after the second "," contains. For the English sentence structure "~, Relative pronoun ~. "Becomes the interior of", "and", " is considered as a block. The point can be another symbol used for the end of the sentence becomes.

To perform such apposition evaluation, the dictionary 9018 is configured to store the meaning information for words. The meaning information represents the difference for the article, location, person, etc., as shown in FIG. 74. Also for the block preparation condition, the table 9036 is formed as shown in Fig. 75 so that the upper end of the block is recognized by "proper name (person), noun (person)" as the upper condition, and the upper one End of the block is identified by "proper name (person), article noun (person)". It is thus possible to evaluate the appositional expression in terms of the morphological and semantic features without performing grammatical analysis and to perform a grammatical analysis according to the apposition score in the example of other processing operations shown in Fig. 64.

By the way, these groups are in the English sentence carry a very specific information and limited be used. When analyzed in the same way will be, as is done for the usual groups they analyzed for a completely different typesetting, and it is difficult, the original texture of the sentence by analyzing. Furthermore, this leads to large Losses.

For example, "let's" or "let us" immediately after punctuation, etc. as a command set with the causative Verb "let" analyzed, and it should be considered a group be analyzed, which the inviting character "let" ("let us," etc.). "Let" can also be different be used as a transitive verb and even as a noun in the meaning of "renting" and is not on the use of friendly invitations or Prompts and not limited to use as an auxiliary verb. Consequently, the analysis must be for the respective possibilities be carried out, reducing the efficiency is lowered. Furthermore, it is difficult to use as To induce invitation from the result of the analysis, as it except in terms of the sentence structure no difference  between use as a causation and as an invitation, and it is difficult this only in terms of the sentence structure to distinguish.

The uneconomical loss in the course of an analysis can be mitigated by letting let's or let us us ") immediately after the punctuation of the object the analysis is excluded. By separating from the basic use of the word, d. H. from the causative Use can be a semantic analysis readily be performed.

If "please", "let's" or "let us" appears at the beginning of the block, a flag for the Block information is set, and this will be for each of the Cases in a piece of information not granted to the unit. For example, the English sentence "let's go to school." ("let's go to school") processed as "go to school" ("go to school") on which "let's" ("let us") attached is.

To perform such "let" processing, a let information processing section 9200 is arranged between the morphological analysis section 9016 and the syntactic analysis section I 9020 in a modification of this embodiment. Fig. 77 shows the respective sections. In this figure , the same elements as those shown in Fig. 64 are denoted by the same reference numerals.

Further, dictionary 9018 is executed to store the let information for the word. As shown in Fig. 78, the let information provides "0" for ordinary words, "1" for "let's" and "let us" and "2" for "please".

The let information processing section 9200 has the task of taking the result of the morphological analysis together with the input English sentence from the section 9016 and adding the let information as the additional information to the information of a word during the analysis, as shown in FIG is shown. 79th In this case, a block is placed for the sentence. In the example shown in the figure, the block is 0 (start: 1, end: 10, destination: sentence, role: sentence). That is, the block in this example includes a clause, a group, etc. in addition to a sentence. In this case, the term for the block also includes a paragraph and the entire sentence, each of which can be considered as a block. Further, a "transitive verb root (appended with 's)" is written as a speech part to "let's" for the information of the word, and the let information is "1".

As shown in Fig. 81, processing for preparing a block of the set 9300 before starting the collective compilation of the block is performed to the input English sentence. The subsequent processing operations are the same as the flowcharts shown in FIG. 66. For example, in the English sentence, I said, "Let's go to school." (I said, "Let's go to school"). Block 0 is formed (Start: Start of sentence, End: End of sentence, Role : Sentence, goal: sentence).

As shown in Fig. 81, in the syntactic analysis section I 9020, each of the structural compositions is recognized as a block based on the English data supplied thereto, and its objective and its role are evaluated accordingly ( 9120 ). If the block does not exist in the array, ( 9121 ) the analysis is ended. If there are blocks in the input sentence , the innermost block is parsed first ( 9122 ). Although various solutions are generally obtained by analyzing, only the solution collectively arranged as a cfg symbol is selected among them ( 9123 ). The subsequent processing operations are the same as those in FIG. 23.

Such processing of "let" information is performed in the section 9200 corresponding to the processing flows shown in Figs. 83A and 83B. First, a pointer to the upper block is set ( 9330 ) to check the word located at the top or beginning of the block ( 9331 ). If the let information is "0", the pointer is advanced ( 9339 ) to pass the next word.

If the let information is not "0", the previous dictionary reference unit is checked ( 9322 ). If it is not an punctuation or if the pointer is not the top or the beginning, the pointer is advanced ( 9339 ) to move to the next word. If the previous dictionary reference unit is an punctuation in the check, or if the pointer indicates the top or the beginning, the innermost layer block containing the word is marked ( 9333 ).

Then, if the let information is "1" ( 9334 ), since immediately after the punctuation, this is "let's" or "let us", the role of the marked block is recognized as (Invitation Set ) ( 9336 ). If the information is "2" because it is "please", the role of the marked block is recognized as "prompt set" ( 9335 ). Then, the destination of the marked block is recognized as an instruction set ( 9337 ), and the word information which has been indicated by the pointer is deleted ( 9338 ). Then the pointer is advanced ( 9339 ) to move to the next word. The process is performed up to the word at the end position ( 9340 ).

Fig. 80 shows examples of the analysis result with which such let information processing has been performed in the example of the above-mentioned input sentence: I said, "let's go to school." (I said, "let's go to school.") When the section 9200 adds the let information to the information for the word, it removes the information concerning the let information from the table, and the block information is called "Imperative Theorem" is described as the target and as an "Invitation Set" as the role, as shown in FIG .

When the dictionary 9018 is retrieved in the treatment of hyphenated words in the English sentence as the entirety of a number of words which are hyphenated, if their entries are present in the dictionary 9018 , the method is successfully performed. If for the hyphenated word that is not registered in the dictionary 9018 , the whole part is treated as an unknown word, e.g. An adjective, since the dictionary information for each of the hyphenated words can not be exploited, it can not be translated. Furthermore, if the entry for the information for each of the constituents of the hyphenated words is present in the dictionary 9018 , they can not be neglected.

Moreover, if the analysis is performed by are decomposed into the respective components, the Art connect to hyphenated words extremely versatile.

To solve this difficulty, the whole hyphenated words are analyzed as an adjective in the sentence, and a corresponding analysis is made only for the inner part of the hyphenated words using the constituents of the hyphenated words, and the result of this analysis is then combined. This also allows parsing for hyphenated words by using the information for each of the constituents. That is, for the hyphenated words that are not registered in dictionary 9018 , the entire part is treated as an adjective in the same way. Here, the nouns associated by the hyphen are referenced in the dictionary, and the corresponding analysis is performed in a closed form only for the interior of the hyphenated words.

That is, when the hyphenated word is a word unregistered in the dictionary 9018 , block information is given that the whole part is regarded as one block, and the dictionary reference is performed for each of the constituents for the inside of the block respective unit information in which the hyphen is not included. For the word unregistered in the reference dictionary, final evaluation processing is performed along with unknown word processing as a kind of unknown word processing.

Such hyphenated word processing may be performed in the example of the structure shown in FIG. 64. In this case, the position of the word in the sentence is expressed not by the number attached to the word but by the number of characters from the beginning of the sentence, that is, by the number of characters.

In Fig. 84, an example of the hyphenated word processing performed in the morphological analysis section 9016 is shown. For the English sentence entered, for example, "The anti-war attitude is her open-door policy"("The anti-war attitude is her open-door policy."). If the position pointer is advanced to take out a word ( 9135 ), then it performs dictionary retrieval ( 9353 ). In this case, the hyphen is not used as a delimitation for the word. If the entry exists ( 9353 ), the word information is written out ( 9359 ).

This is then repeated until the end of the sentence.

In the event that the entry does not exist as a result of the dictionary reference 9352 , if there is no word containing a hyphen ( 9354 ), word information is spelled out ( 9359 ), whereas if there is a hyphenated word , a block is written with the hyphen ( 9355 ). In the hyphenated block, the start position is the start position for the hyphenated word, and the end position is the end position for the hyphenated word. The goal is optional, and the role is adjective / noun. The hyphen is then extracted to take out each of the constituent words ( 9356 ) and the respective constituent words are retrieved from memory ( 9357 ). The word information obtained as the result of the dictionary retrieval ( 9358 ) is written. When the word information is written out at steps 9359 and 9358 , in the case of a word not registered in the dictionary, it is written out as a speech part = a word not registered in the dictionary.

Fig. 85 shows examples of the block information and the word information of the English block which is collectively assembled into a block as a result of the processing of the inputted sentence example. In this example hyphenated words "anti-war" are registered in the dictionary 9018 , and the words "open-door" are not registered in the dictionary. Thus, the entry for the hyphenated word "anti-war" is written out as the information for the word. For the hyphenated words "open-door" they are broken down as "open" and "door" and written out as the information for the words, and block 1 (start: 30, end: 38, destination: optional, role: adjective / Noun) is written out as the information for the block. Although the form of the supplementary English question is extremely limited, its processing in the usual analytical procedure is very complicated. Furthermore, it is not easy to determine the verb to which a so-called supplementary question relates. After recognizing that it is a so-called supplemental question based on the feature of the sentence in terms of its shape, it is treated as a supplemental question relative to the associated structural composition, whereby the verb concerned by the supplementary question can be specified , That is, the part of the supplemental question in the English sentence is found out as a structural pattern, and the analysis is performed, the part of the supplementary question being considered as pure information with a certain kind of associated structural composition.

In the present embodiment, a unit or a block in the form of a symbol (a starting point display, indicating that this is a unit or Block endpoint is). In morphological analysis, the formed sentence text, in which also the recognition for the block is performed. In the present embodiment is "a quote" denoted as "Q", and a bracket is referred to as "P". For example this shall be determined as indicated below:

'. , , 'By (Q' ...) ',
"...." by (Q "...)",
(...) by ((P. ..)),
<. , , <by <(P. ..) <,
{. , , .} by {(P....}} and
[. , , .] by [(P....

The block detection is performed in the same way.

The start and the end symbol of the block are only under applied to the context where the block is opened by these icons or closed. The part immediately before  Start signal and immediately after the end signal should others be as alphanumeric symbols. The preceding symbols, which do not correspond to, are treated as pure symbols. The blocks can be interlocked several times, provided that they do not cross or overlap.

If, when processing the supplementary question, the follow the following phrases at the moment when the Pointer "," indicates, the entity (cluster) becomes "," until "?" deleted as a unit and a flag is considered a Block set. That is, the form of the supplementary question set includes:

", (Auxiliary) + (personal pronoun)?"
", (Auxiliary) not + (personal pronoun)?"
", (Auxiliary) + (personal pronoun) + not?"

Furthermore, the various auxiliary embodiments include the following a: am, is, are, what, were, do, does, did, have, has, had, wants, should, could, should, can, can not, could, may, might, must, ought, will not, shall not, need, dare, used. The species of personal pronouns close I, you, he, she, it, we, they are one.

These are called the information in the innermost block layer, to which they belong. For example in an English sentence: you said so, did not you ?, the whole part as a block in terms of structural composition in [You said so,] <with supplementary question <treated. Similarly, in the English sentence: I said, "You said so did not you? ", the sentence quoted" You said so did not you? " as a block 1 in terms of structural composition recognized, and further, the whole part as a Block 2 recognized in terms of structural arrangement. That is, [I said, [you said so,] <provided with a supplementary question <.  

The abbreviation word like "did not" is treated after it according to a predetermined table in a complete tendered form has been developed. For the Word that has a number of developed forms become these all spent.

In order to perform such processing for a so-called supplementary question, a corresponding section 9210 is arranged between the morphological analysis section 9016 and the syntactic analysis section I 9020 in another modification of the invention. Fig. 87 shows the respective portions. In these figures, the elements which are identical to those in Fig. 64 are denoted by the same reference numerals.

The section 9210 obtains the result of the morphological analysis along with the input English sentence from the dedicated section 9016 , and as shown in Fig. 88, a block for the sentence is placed. In the examples shown in the figure, the block is 0 (start: 1, end: 12, destination: sentence, role: role). In this case, in this modified embodiment, the word is represented by the number for the word. For example, in this modified embodiment, the block includes a sentence in addition to a clause and a group. In this case, the notion of the block also includes a paragraph and the whole sentence, each of which can be considered as a block.

The collective composition of the block for the input English sentence including the supplementary question may be the same as that in the flowchart described above and shown in FIG . That is, the processing 9300 for preparing the block of the sentence is performed before the start of the processing. For example, in an English sentence, I say, "It is good, is not it?" Block 0 formed (start: beginning of the sentence, end: end of the sentence, role: sentence, goal: sentence). In the syntactic analysis section I 9020 , analysis is performed by the same flowchart as shown in FIG. 82. Processing in the supplementary question processing section 9210 will be explained with reference to Figs. 90A and 90B. First, a pointer to the word at the top of the word information is set ( 9370 ). If it is not a comma, the pointer advances step by step ( 9384 ). This is repeated until the end of the sentence ( 9371 ). Then, it is checked if the word closest to the comma is a word belonging to the α group or a word belonging to the β group while the pointer is left in place he is ( 9373, 9379 ). In this case, it is then determined that what belongs to an auxiliary verb or verb is in the speech part and is not in negative form are words belonging to the α group, while what the negative form of the Auxiliary verb or the negative form of a verb in the speech part are words belonging to the β-group. If the word does not belong to any of the groups, the pointer is advanced ( 9484 ) and the operations are repeated until the end of the sentence ( 9371 ).

In the case where the word belongs to the α group, the step 9384 at which the pointer advances is performed when the word closest to the words of the α group is not a pronoun. If it is a pronoun, it is checked whether the next word is "not" or not ( 9375 ) and if it is not "not", it is checked if the word closest to the pronoun is a question mark or not ( 9377 ). If it is not a question mark, step 9384 is performed. If it is the question mark, the target is again written in a "negative sentence" and the role in the "Supplementary Question Set" for the innermost shift block ( 9378 ) and the ",. is deleted from the information table for the word ( 9383 ). The innermost layer block designates such blocks which satisfy the conditions: start position ≦ (position for ",") and which also satisfy the condition:
End position ≧ (position for "?") For the block position and with the minimum (end position start position).

If the word closest to the pronoun is "not" at step 9375 , it is checked whether the word closest to "not" is a question mark or not ( 9376 ). If it is not a question mark, step 9384 is executed, ie the pointer advances. If it is the question mark, the destination for the innermost layer block is written in an "operation set" while the role is written in a " supplementary question sentence" ( 9382 ), and ",. is deleted from the word information table ( 9383 ).

If, at step 9379, the word closest to the comma is a word belonging to the β group, step 9384 is performed if the word closest to the β group is not a pronoun. If it is a pronoun, it is checked if the word closest to it is a question mark ( 9381 ) or not, and step 9384 is entered if it is not a question mark. If it is the question mark, the destination of the innermost layer block is written in a "confirmation sentence " while the role is written in a " supplementary question mark" ( 9382 ), and ",. is deleted from the word information table ( 9383 ). Then the pointer is incremented ( 9384 ) and the operations are repeated until the end of the sentence.

For example, Fig. 88 shows the information for the blocks and the words which are sent from the morphological analysis section 9016 to the supplementary question processing section 9210 for the English sentence: "It is good, is not it?" has been obtained. The block information for block 1 is (start: 4, end: 12, destination: optional, role: optional). When subjected to the supplemental question processing in the section 9210 , the block information for the block 1 is rewritten in (start: 4, end: 12, destination: confirmation sentence, role: additional question sentence) and at the same time, the word information which is the supplementary question # 8- # 11 concerns deleted.

Claims (11)

A speech analyzer, comprising dictionary means containing morphological data of a particular language and additional data enabling multiple word combinations, comprising: first analyzing means for performing a morphological and syntactic analysis of an input string of said language using said morphological data, and with a second analysis device for performing a syyntaktischen analysis of the input string (word sequence) on the basis of the analysis result of the first analysis device, wherein

• a) the first analysis means ( 1016 ) is adapted to define a syntactic block using said additional data and the morphological data and to size syntactically related morphemes using the additional data and the morphological data, and
• b) the second analysis means ( 1020 ) is arranged to parse the string (word sequence) within the syntactic block and to treat the result of the structuring process or the associated morphemes as a single morpheme.

The speech analyzer according to claim 1, characterized in that said first analysis means ( 1016 ) is adapted for detecting a high degree of coupling for a number of combinations between a word retrieved in the dictionary means and other words stored in the dictionary means during the analysis To prepare preference data to give preference to a number of combinations for subsequent processing. The speech analyzer according to claim 1, characterized in that said dictionary means comprises a basic unit storage section ( 2026 ) for storing therein dimension units, and that said first analyzing means addresses said memory section ( 2026 ) with said basic units for said inputted character string, thereby morphologically to analyze whether the string expresses a dimension unit or not. The speech analyzer according to any one of claims 1 to 3, characterized in that the first analysis means ( 1016 ) has a pointer settable to a character at the front of the input string, and when a portion of the string is retrieved at fetching the dictionary means for retrieving the string starting from the character at which the pointer has been set, the pointer being settable on a string following the part of the retrieved string, the dictionary means being then addressable for the subsequent string to which the pointer is set. The speech analyzer according to any one of claims 1 to 4, characterized in that said dictionary means stores dictionary data stored therein at respective dictionary reference units, and said first analyzing means ( 1016 ) divides said input sentence into dictionary reference units and morphological analysis for the dictionary reference units, referring to the dictionary means, wherein the dictionary means stores as the dictionary data a discrimination notice indicating that the dictionary reference units represent numbers, and when the discrimination notice is included in the retrieved dictionary data and a dictionary reference another dictionary reference unit is retrieved with a different discrimination notice located near the former dictionary reference unit, a numerical value of the two dictionary reference units by the second analysis facility ( 1020 ) is formed. The speech analyzer according to any one of claims 1 to 5, characterized by feature information creating means (in 1020 ) for generating replacement feature information at a string not recorded in the dictionary means. The speech analyzer according to claim 6, characterized in that said feature information generating means (in 1020 ) is adapted to, with two consecutive strings (word sequences), one string (word string) having no feature information, compare them with the feature information of the other string (Word sequence) having feature information. A speech analyzer according to any one of claims 1 to 7, characterized in that when the string starts from a capitalized character and the preceding string (phrase) is at the end of the sentence, the first analyzer ( 1016 ) generates the capital letters at the beginning of the string are converted to a lower case letter and then addressed to the dictionary device by means of a fetch device for that string, and if not registered in the dictionary device, the string is parsed as a non-registered proper name. A speech analyzer according to claim 1, characterized in that the dictionary means comprises data for distinguishing dictionary reference units from a particular semantic element, the first analyzing means ( 1016 ) having a corresponding discrimination table for determining whether a sequence of dictionary reference units forming, with the specific semantic element, a composite unit expressing a specific meaning composed according to a certain rule, and the first analyzing means ( 1016 ) retrieving the dictionary means with respect to the respective dictionary reference units contained in the input sentence; and form the sequence of dictionary reference units with the specific semantic element into a single analysis unit, corresponding to the corresponding discrimination table. The speech analyzer of claim 1, characterized by control means ( 1038 ) for retrieving the dictionary means ( 1018 ) and controlling the first and second analysis means ( 1016, 1020 ) to perform respective analyzes. The speech analyzer according to claim 1 or 10, characterized in that said first analysis means ( 1016 ), when no dictionary data is included, when said dictionary means ( 1018 ) is retrieved for a number of hyphenated words, said dictionary means ( 1018 ) for each of said Retrieving the number of words, evaluating the whole part of the number of words as a compilation, and then evaluating the arrangement of the compilation as an adjective group.