JPH052605A - Machine translation system - Google Patents

Abstract

PURPOSE:To division/synthesis-process a sentence which is difficult to translate in terms of a long sentence and a morpheme. CONSTITUTION:A translation part 3 divides input data forming an original sentence into the unit of the sentence based on information on a translation dictionary part 5, and the sentence is divided into a translation unit. The divided parts of the translation unit are separately translated and the parts of the translation unit which are translated at the time of generating the translation sentence are synthesized. Plural sentence acknowledgement means for the text of the original sentence are prepared, and the sentence acknowledgement means adjusted to a purpose and a use is selected and used. Furthermore, a translation word is obtained by noticing sentence end candidate information of a list guide sentence and a list sentence end connection word as a list sentence recognition means among the sentence recognition means.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a machine translation system, and more particularly to a machine translation system for performing morphological analysis by improving translation accuracy and editing work efficiency. For example, it is applied to an English-Japanese machine translation device and a natural language processing technique.

[0002]

2. Description of the Related Art In a conventional machine translation system, the sentence length has a great influence on the processing speed and translation accuracy of the system. That is, the processing speed increases exponentially with respect to the length of the sentence, and the translation accuracy sharply decreases when the translation length exceeds a certain length. For this reason, it is often practiced to use a machine translation system to translate a sentence by dividing it. But this is because the user couldn't translate
We deal with the places where we think that we can't translate by designating the translation sentence unit by preediting. Moreover, the result obtained by these processes is not in the form of a sentence, so if you do not bother to edit it so that it becomes a sentence in the pre-editing or change it considerably in the post-editing, you will not get the translation you want There is a problem.

In the conventional machine translation system, it is considered only from the viewpoint of translation from text to text, and not from the viewpoint of translation from text to text. For that reason, the document format information such as the format, column positions, and the number of spaces between words in the document is not taken into consideration, and the document and the sentence are reformatted for the convenience of the system, and even the original image is broken. . End-of-sentence recognition is done by using end-of-sentence symbols such as periods as clues,
This is generally done by instructing the end of sentence by the user, and the information in the document format has not been used. For this reason, there were many failures in the recognition of title sentences and list sentences, which were usually written without a trailing mark such as periods, and the analysis rate was significantly reduced. Similarly, headwords that are often written without end-of-sentence symbols such as periods, and symbols are not separated, and thus become an element of syntactic analysis, resulting in a burden of syntactic analysis and failure of parsing. There were many things. Furthermore, in order to prevent or correct such a sentence recognition failure, the user has to rely on the user, and pre-editing and post-editing are time-consuming, which is one of the factors that reduce the efficiency of the whole translation process.

Further, the conventional sentence recognition is for so-called prose style and is not suitable for a format such as a manual in which title sentences and list sentences frequently appear. The sentence recognition was uniquely determined, and after translation, the user had to redo it. If the sentence recognition fails, it is up to that point, and another candidate that the user can fix by himself, unlike the translation result, was not prepared. Also, there was no way to deal with titles and lists by changing the sentence recognition. Also, although it had multiple spell check functions, its merits were not fully utilized, and only one sentence recognition function, which could be said to be the job of the same morphological analysis unit, was retained.

There are various documents that require machine translation. In the example of using English-Japanese machine translation, there is a typical genre of manual when it comes to the translation of industrial documents in the world. It is said that manuals are suitable for machine translation because of the large amount and relatively simple English. However, the description format of the manual is different from English such as magazines and papers. There are many manuals that are more like sentences than sentences, and that do not seem like sentences. It has titles, lists, headings, and none of them. Unlike ordinary sentences, these are not always a group of meanings that end with a sentence end recognition symbol such as a period. So-called sentences (one translation processing unit) from where to where
It is a difficult problem to catch.

Various document input methods are possible. Examples are keyboard input, disk media, communication, OCR (character recognition device), and the like. Above all, it mechanically reads characters printed on paper and converts them into electronic information O
CR is useful as a convenience. For example, OC
The alphanumeric information read by R is translated by an English-Japanese machine translation device. At this time, it is very convenient if spell checking and sentence recognition are performed just before being read by OCR and translated. Because OCR is not universal, some read errors always occur. Therefore, it is necessary for humans to read it to some extent. At this time, OCR
It will be very convenient if the spell check and sentence recognition are completed immediately after the start. At this time, concurrency is required,
Since speed is important, when recognizing a sentence, the word dictionary information is not referenced, and only the morphological features of the character string are used to make the determination. This method sometimes fails to process abbreviations that end with a period. In order to solve this, it is necessary to search the dictionary information. Therefore, a sentence recognition method with a search of dictionary information should also be adopted.

[0007]

[Purpose] The present invention has been made in view of the above-mentioned circumstances, and it is necessary for the system to automatically divide the sentences performed by these users and synthesize them to some extent. Providing a machine translation method for dealing with long sentences, sentences that are morphologically unsuccessful in translation processing (for example, list sentences, etc.), automatically or by simply specifying by the user. If multiple text sentence recognition means are prepared and the text of the original text does not have a sentence terminator, the sentence is certified based on the character attributes of the text of the original text and the layout information of the text of the original text. Components (words and symbols, etc.)
The vertical and horizontal layout information of the above, the empty sentence, the title sentence, and the list sentence are certified from the vertical layout information, and the symbols and entry words can be certified from the horizontal layout information. The purpose was to provide a machine translation method that allows a certain degree of qualification processing even in the case of texts, and that certifies texts according to various conditions such as texts, situations, and uses. Is.

[0008]

In order to achieve the above object, the present invention provides (1)
In the machine translation method that analyzes the input source language original sentence and outputs the translated sentence of the target language as a result, divides the input data that forms the original sentence into sentence units, and further divides the sentence into translation units Means and the translation unit portion divided by the dividing means are subjected to different translation processing, and the whole translation result is obtained by synthesizing the translation unit portion subjected to the translation processing at the time of translation generation. Processing means, and further, (2) providing a relationship between division units at the time of division processing, and having an analysis means for analyzing based on the division time information, and (3) the division time information. In order to obtain an appropriate translated word by using it, it has a control means for controlling the translation order, utilization, etc., and (4) the dividing means,
It is used to process list statements, and further (5)
A separate translation is held for each translation unit, and the user can select and use the translation unit by an editing operation. Further, (6) the user specifies whether or not to perform translation unit division, and (7) translation The user sometimes designates a unit for division processing, or (8) an input device for inputting the original text of the first language to the processing device, a display device for displaying the information input by the input device, and the information processing Second processing device
In a machine translation system including an output device for outputting a language, a plurality of sentence recognizing means for the text of the original sentence are prepared, and (9) in (8), the text of the original sentence has no sentence end symbol. In this case, the sentence recognizing means performs sentence recognition based on the character attribute of the original text and the arrangement information of the constituent elements (words, symbols, etc.) of the original text in the vertical direction and the horizontal direction. ) In (9), the sentence certifying means includes an empty sentence certifying means, a title sentence certifying means, a list sentence certifying means, a symbol part certifying means, and a headword certifying means.
1) In the above (10), the blank sentence recognizing means recognizes as a blank sentence when line breaks appear consecutively, and (12) In the above (10), the title sentence recognizing means: Sentences are recognized just before the target text to be certified, and when the target text starts at the beginning of the line and the next line starts in a case other than lowercase, each word except the preposition, article and conjunction When the beginning of is a capital letter, a symbol, or a number, it is recognized as a title sentence, and (13) In (10), the title sentence certifying means is sentence-certified immediately before the target text to be sentence-certified. When the target text starts at the beginning of the line and the next line starts with a letter other than lowercase, the number of words that make up the line is less than the specified value, and the indentation of the next line is deeper. (14) In (10), the list sentence recognizing means has a type (any one of a numeric word, a letter of the alphabet, a symbolic character such as an asterisk) and an appearance position. When texts of the original sentence having the same head portion appear consecutively, it is recognized as a list sentence, and further, (15)
In the above (10), the symbol part recognizing means, when it is recognized as the list sentence, certifies the head part thereof as a symbol part, and (16) in the above (10), the symbol part. The certifying means certifies that, when a numerical word or an alphabetic character at the beginning of the original text is enclosed in parentheses, the portion enclosed in the parentheses is a symbol portion,
Further, (17) In the above (10), the symbol part recognizing means is configured such that when any of a parenthesis, a period, a colon, or a semicolon is added to the first numeric word or one letter of the text of the original text, The added portion is recognized as a symbol portion, and further, (18) the above (10)
In the above, when the text of the original sentence starts with a symbol character (such as an asterisk), the symbol portion recognizing means recognizes the symbol character as a symbol portion, further, (19) above (10)
In the above, the headword recognizing means is sentence-recognized immediately before the target text to be sentence-recognized, or when the target-word has already been recognized just before the target text, the target text is equal to or more than a specified value in a line. When there are multiple spaces, recognizing from the beginning of the line to the space as an entry word, further, (20) In (10) above,
The headword recognizing means has a plurality of spaces, each of which is equal to or less than a specified value, in a line when the target text has been recognized immediately before the target text to be sentence-recognized or when the target text has already been recognized just before the target text. If the character immediately following the multiple spaces begins with a non-lowercase letter, and there is an end-of-sentence at the end of the line, or if the entire line is already end-of-sent, then there is a space from the beginning of the line to just below the last multiple spaces. In the case where the number of words up to a plurality of spaces is 1, in any one of the cases, the part from the beginning of the line to the space is recognized as a headword, and (21) In (10) above, The symbol part certified by the symbol recognizing means is not subject to the syntactic analysis, or (22) an input device for inputting an original sentence of the first language to a processing device; In a machine translation system including a display device that displays information input by a device, a processing device that processes information, and an output device that outputs a second language, a temporary sentence of text in the first language ( Displaying the certification result (1 unit of translation processing) or (2
3) an input device for inputting the original sentence of the first language to the processing device,
In a machine translation system including a display device that displays information input by the input device, a processing device that processes information, and an output device that outputs a second language, a text sentence of the first language ( (1 unit of translation processing) highlighting a part of the recognition result, or (24) an input device for inputting the original sentence of the first language to the processing device, and a display device for displaying the information input by the input device In a machine translation system including a processing device for processing information and an output device for outputting in a second language, a plurality of text sentence (one unit of translation processing) certification results of the first language are prepared. Select by referring to the plurality of sentence recognition results, or (2
5) An input device for inputting the original text of the first language to the processing device,
A display device for displaying the information input by the input device, a processing device for processing the information, an output device for outputting the second language, and a plurality of sentence certifying means for certifying the sentence of the original text. In the machine translation system consisting of, as the list sentence certifying means among the sentence certifying means, by the part of speech of the sentence end candidate information serial word in the list introduction sentence,
Changing the syntactic role given to the sentence end candidate information, and giving a translated word of the sentence end candidate information according to the part of speech;
(26) In (25), the list sentence end connective word that is likely to occur at the end of the list sentence approved by the list sentence identifying means is specified, and the list sentence end connective word is added to the beginning of the next sentence that is not a symbol part. Or (27) an input device for reading an original sentence in the first language with a character recognition device, a display device for displaying information input by the input device, a processing device for processing the information, and a second language In a machine translation system equipped with an output device that outputs a document, it is determined whether the document is a horizontal writing document or a vertical writing document, the line range specifying means for measuring the line spacing, and the word spacing Calculating means for determining the position and the length, indent calculating means for obtaining the indent distance from the length of the interword distance, and determining the type of indent from the distribution of the indent distance,
Document format estimating means for recognizing linguistic features such as title lines or list sentences based on the type of indentation determined by the indent calculating means, and characters for recognizing words or characters based on the document format estimating means Or (28) an input unit for inputting an original sentence in the first language, and a translation dictionary unit having a translated word in the second language corresponding to the word in the first language, semantic information, and the like. A morphological analysis unit that morphologically analyzes the original text input by the input unit; a syntax analysis unit that parses the original text input by the input unit; and a conversion that converts / generates the second language.
In a machine translation system that includes a generation unit and an output unit that outputs a translated sentence that is converted / generated by the conversion unit, in the case where a sentence immediately before a sentence that failed in analysis is a title sentence that also serves as the title of the sentence, The translation processing is performed by concatenating the parsing failure sentence and the immediately preceding title sentence as one sentence. Further, (29) in (28), the sentence immediately before the parsing that fails the parsing is the title of the sentence. In the case of a double-ended title sentence, the translation processing is performed by giving syntactic information such as a part of speech, a subject, or a predicate to the first word of the parsing failure sentence and the title portion immediately before it.
(30) If the translation process of (29) above is unsuccessful in the analysis, perform the translation process of (28) above;
(31) Delete the translated words of the part corresponding to the title from the translated text obtained by the translation process of (28) or (29) to generate a sentence, and (32) add the sentence of (28) or (29) above. It is characterized in that the translated word in the part corresponding to the title is replaced with a pronoun from the translated text obtained by the translation process to generate a sentence. Hereinafter, description will be given based on examples of the present invention.

FIG. 1 is a block diagram for explaining an embodiment of a machine translation system according to the present invention, in which 1 is an input unit and 2 is
Is an original sentence storage unit, 3 is a translation unit, 4 is an edit control unit, 5 is a translation dictionary unit, 6 is a translated sentence storage unit, 7 is a display control unit, 8 is a display unit, 9 is a printing unit, and 10 is a dictionary editing unit. is there.

In the embodiment of the present invention, an English-Japanese machine translation apparatus for inputting an English sentence and obtaining a translated sentence of a Japanese sentence will be described. First, the input original text is input by the input unit 1 including a keyboard and sent to the original text storage unit 2. The translation unit 3
Based on the information in the translation dictionary unit 5 under the control of the editing control unit 4,
The original sentence in the original sentence storage unit 2 is translated for each predetermined processing unit. The translated texts obtained by the translation processing are sequentially stored in the translated text storage unit 6. The edit control unit 4 drives the display control unit 7,
The original sentence stored in the original sentence storage unit 2 and the translated sentence stored in the translated sentence storage unit 6 are displayed on the display unit 8 in association with each other. The operator performs post-editing processing while looking at this display. The post-editing process is executed by the edit control unit 4 according to the control information input from the input unit 1. The dictionary editing unit 10 corrects / changes / deletes the contents of the translation dictionary unit 5 under the control of the editing control unit 4. These results are reflected during translation. After the post-editing process, the completed translated sentence and the original sentence are output by the printing unit 9.

FIG. 2 is a block diagram of the translation unit. In the figure, 11 is a morphological analysis unit, 12 is a syntax analysis unit, 13 is a syntax conversion unit,
Reference numeral 14 is a morpheme generation unit, 15 is a word dictionary, 16 is a grammar rule, and 17 is a generation rule. Here, four major processes of the translation unit are shown. First, the morpheme analysis unit 11 uses the word dictionary 15 to perform normal morpheme division processing and the like. The syntactic analysis unit 12 obtains information on each word and creates a tree structure according to the grammar rule 16. The syntax conversion unit 13 converts the tree structure of the input language into the tree structure of the output language. The morpheme generation unit 14 translates the obtained tree structure using the generation rule 17 for each node.

FIG. 3 is a flow chart for explaining the operation of the morphological analysis unit in FIG. Hereinafter, each step will be described in order. step1 ； First, read the character string of the input language. step2 ； Next, the lexical analysis and sentence recognition processing is performed by the English word dictionary 1
9 is used. step3 ; In the list part condition determination, it is determined whether or not the sentence includes a list part by the list part condition. When the list part is recognized as the list part, the position of the symbol part of the list part and the position of the sentence following the symbol part are given to the pattern matching part. step4 ; Then, the pattern matching unit identifies whether the sentence-recognized sentence unit can be divided into processing units by using the pattern recognition rule 18. step5 ; Next, the copy unit recognized in the above step 3 is used as an analysis unit using the English word dictionary 19 to perform the process after the syntactic analysis.

The list sentence is a general term for sentences led by symbols. For example, *, *, ▲. ▼ (▲. ▼ means a white circle). Symbols such as (1), (2), ... a), b), ... are used. Each list sentence may be a sentence or may form a sentence as a whole. Corresponding to each sentence is relatively easy. Simply recognize that there is a special symbol in front of the sentence and analyze the rest as a sentence. The problem is a sentence that has a list format (hereinafter referred to as the list part) that is embedded in the sentence.
It can be solved by synthesizing it later.

The list part is analyzed under the following conditions. (1) The line is broken and indented, and if the beginning is the following, it is assumed to be a list sentence Number word Alphabet 1 character A symbol such as asterisk (2) The line is broken and indented, and the beginning is In case of, it is confirmed as a list sentence. Numbers with parentheses ex) (1), (a), ... 1), a), ... Number symbol e with period (colon, semicolon)
x) 1. a. (3) In the above case, if it exists in a sentence without a line feed indent, it is assumed to be a list part.

The first symbol and the like in the list part is called a symbol part. The difference between hypothesis and definite is that in the case of definite, the list part is unconditional, and in the case of hypothesis, there is a similar list part that follows, it must have the same kind of symbol part, and the appearance column position must be the same. However, in the case of (3), it is determined only by the type of the symbol part, not the position of the appearance column. The conditions that the list section must satisfy are called list section conditions.

FIG. 4 is a diagram showing pattern matching rules. As a result, the target to be divided in the sentence is specified, the relationship between the divided processing units is determined, and information such as the order / hierarchical relationship, roles, and goals of the divided processing units is added. The description format of the rule is as follows. [] Indicates a unit to be divided, and a matching condition is written inside the unit. "~" In [] indicates matching with an arbitrary character string. “<Numerical value>:” in [] is the minimum number of matching characters. If the number of characters is less than this, even if the matching is successful, it is ignored. This is specified when it seems that it is better to process simultaneously as usual rather than dividing a too small range. $ <Written after pattern description
The description including the number> indicates the operation when the matching is successful. The number after the $ indicates the previous [] numbered in the order they appear. $ <Number> = category indicates a goal when analyzing the divided parts, and $ <number><−
$ <Number> indicates the subordinate relationship of the divided parts. [KP ~] *
The symbol "*" in (1) represents repetition and the symbol "KP" represents a symbol part of the list part of the list sentence.

For example, according to rule 1, when a plurality of units including a character string "regarding to" come after a character string that is a unit of 10 characters or more, each is regarded as a translation unit, and
It shows that the lump before to is analyzed as a sentence, and the analyzing to is analyzed with a prepositional phrase as a goal. Further, it indicates that the prepositional phrase should be processed to be included in the first sentence. Rule 2 divides a sentence embedded in quotation marks and processes it as a processing unit. Rules 3 and 4 are written to process patterns such as to infinitives that often appear in bullets, and if clauses below the if clause.

Rule 3 means the following: The first [~] is an arbitrary character string. [To:] indicates that a sentence pattern of ~ to: ~ comes. [KP-] * indicates that the symbol KP that represents the symbol part included in the sentence recognized as a list sentence by the list condition processing matches the symbol KP, followed by an arbitrary character string multiple times. . As the action of this rule, the first [~], that is, $ 1 is analyzed as SE (sentence),
The form of $ 2 + $ 3, that is, "to: KP ~" is analyzed as prepf, that is, a prepositional phrase. However, ':''KP' is treated as if it does not exist in the course of analysis. This $
If there are multiple processes that analyze 2 + $ 3 with prepf as the goal, they are performed for each. Furthermore, $ 1 <-
As a result, $ 2 + $ 3 is treated as a subordinate node in the morpheme generation unit. According to this rule, $ 1 is analyzed in the form of $ 1 + <prepf>, that is, a dummy morpheme unit <prepf> for the first matching.
Is added to indicate that the analysis is performed.

You must supply the RECFM values if: 1. The data set resides on an unlabeled tape volume or 2. the data set is nor included in a DSCB. Is divided into. You must supply the RECFM values <if-clause>. (Goal: sentence) 1. if the data set resides on an unlabeled tape volume or <KF>. (Goal: if clause) 2. if the data set is nor included in a DSCB. (Goal: if clause) <> corresponds to the processing unit of the sentence divided by the dummy morpheme unit. That is, in the analysis, these are analyzed as if there are symbols without translations. In addition, information such as goals written after the sentence is passed to the subsequent phase together with the processing unit. In other words, syntactic analysis can be processed as a completely different sentence.

In the generation, these analysis units are put together as a sentence unit. First, the conventional generation process is performed. However, the dummy morpheme unit remains the same. You are
The value of if-clause> RECFM must be given. 1. If the dataset is on an unlabeled tape volume or <KF> 2. If the dataset does not contain a DSCB, then these three statements result in the parsing unit generation. This is combined to obtain the entire translation result. You can 1. REC if the dataset is on an unlabeled tape volume, or 2. REC if the dataset does not contain a DSCB.
You must give the FM value.

FIG. 5 is a flow chart for explaining the pattern matching process. Hereinafter, each step will be described in order. step1 ; First, it is determined whether the processing unit is a divided processing unit. step2 : If the processing unit is divided, normal generation processing is performed. step3 ； Judge whether all the solutions of one sentence are complete. step4 ； If the solutions are complete, solution composition processing is performed. step5 ； If the solution is not complete, keep the solution. step6 ; If it is not the processing unit divided in step 1, the normal generation processing is performed and the processing ends.

FIG. 6 is a flow chart for explaining the operation of the edit control section. The steps will be sequentially described below. step1 ; First, it is judged whether or not the translation in the translation section is completed. If finished, the translated text is displayed. step2 ; Next, if the translation is not completed, it is judged whether or not there is a dictionary editing instruction. If there is a dictionary edit instruction, the dictionary edit process is performed. step3 ; Next, if there is no dictionary editing instruction, it is determined whether there is a translation instruction. If there is a translation instruction, the translation process is performed. step4 ; Next, if there is no translation instruction, it is judged whether or not there is an input instruction of the original sentence. If there is an input instruction of the original sentence, the original sentence is stored in the original sentence storage unit and the original sentence display is changed. step5 ; Next, if there is no input instruction, it is determined whether there is an edit instruction. If there is an edit instruction, edit processing is performed. step6 ; Next, if there is no editing instruction, it is determined whether there is another translation instruction. If there is a translation instruction, the translation is displayed and the original text display is changed. step7 ; Next, if there is no other translation instruction, other processing is executed and the process returns to step1.

As described above, in the separate translation display, not only the display / selection processing of a plurality of solution candidates for one sentence, which is conventionally performed, but also the analysis candidates are displayed for each analysis unit. As a result, the user can be presented with a solution candidate for each part that is easier to understand than simply displaying a large number of different translations. This is shown in FIG. Further, in the above examples, the analysis unit is divided by the morpheme processing, but it is possible to easily extend to allow the user to specify the processing or to prohibit the division processing. Also, although the examples so far have been described in English and Japanese, it is clear that it is also useful to carry out the method of the present invention for Japanese-English and other language pairs.

8 (a) and 8 (b) are block diagrams for explaining an embodiment of the machine translation type sentence end recognizing device according to the present invention, in which 41 is an original text and 42 is a sentence end recognizing device. , 42a is an end-of-sentence recognizing device when there is no end-of-sentence symbol, 42a
b is an end-of-sentence recognizing device when there is an end-of-sentence symbol, 43 is a post-sentence recognizing text, 44a is a blank sentence certifying means, 44b is a title sentence certifying means, 44c is a list sentence certifying means, 44d is a symbol part certifying means, and 44e is a headline. It is a word recognition means. Figure 8
In (a), there is a sentence end recognizing means for the case where the original text has a sentence end symbol and the case where there is no sentence end symbol. Further, when there is no end-of-sentence symbol, it has a plurality of end-of-sentence recognizing means as shown in FIG.

FIG. 9 is an overall configuration diagram for explaining the machine translation system. In the figure, 61 is a basic dictionary, 62 is a grammar rule, 63 is a field dictionary, 64 is a translation unit, 64a is a preprocessing unit, 64b is a morphological analysis unit, 64c is a syntax analysis unit, 64
d is a syntax generation part, 64e is a post-processing part, 65 is an original text, and 66 is a translated text. Morphological analysis unit 64b
Then, the dictionary of the original text 65 is looked up. When dictionaries are used, basic dictionaries and field-specific dictionaries are used. The syntax analysis unit 64c later analyzes the information of each word according to the grammar rule 62, and creates a tree structure from the analysis result. The syntax generation unit 64d translates the obtained tree structure for each node.

Next, an embodiment of the method of configuration (11) will be described. For example, when the following document is input from the input unit, a line feed code sequence is searched for. If the line feed code is continuous, it is recognized as an empty sentence, and the part up to immediately before the continuous line feed code is recognized as a sentence. In the example below,
Let'§ 'be the line feed code. In this case, the first line “The Machine Translation
"System" is certified as a sentence. Also, for example, even when the user specifies the end of a sentence, if it is known that the end of a sentence will always be recognized if there is an empty sentence, it is easy to specify the end of the sentence or modify it. Will be realized.

Next, an embodiment of the schemes (12) and (13) will be described. When sentence recognition is performed immediately before and starts at the beginning of a line, and the next line begins with a non-lowercase letter, the following checks are performed. FIG. 10 shows the flowchart. (1) Check whether the beginning of each word is capital letters, symbols, or numbers. However, even if the first letter of the word is lowercase, if it is a preposition, article, or conjunction, the processing is continued. In the flowchart of FIG. 10, INITIAL
It corresponds to the check of CAP. (2) Check whether the number of words forming a line is smaller than a specified value and the indentation of the next line is deeper.
If any of the above is true, the line is recognized as a title sentence and sentence recognition is performed. For example, If the document is input, the first line applies to (1) above, is recognized as the title sentence, and is separated from the second line.
Also, for example, the prescribed value of the number of words in the above condition (2) is set to 6
In that case, the first line of the following example is recognized as the title sentence and separated from the second line. In this case, the first line is sentence certified.

Next, an embodiment of the schemes (14) and (15) will be described. Here, as an example, the configuration (1
Although the explanation will be given according to the algorithm flow including the means of the configuration (15) in 4), it is also possible to create the two means separately. The flowcharts are shown in FIGS. 11 (a) and 11 (b). First, input in line units, and if you can no longer input, end. In listflag, it is turned on when the list statement is certified. When this flag is turned on at the end (where it starts with B), it indicates that the last line is a list sentence. Next, it is judged whether the sentence is a symbol sentence by a symbol sentence authorization routine. The symbolic sentence here is a convenient name as a general term for a sentence that begins with a symbolic character such as a numeric word, one letter of the alphabet, or an asterisk. In addition, the symbolic sentence recognition routine is to check whether the input is a symbolic sentence,
End-of-sentence certification is not performed. If it is not a symbol sentence, the process returns to input, and if it is a symbol sentence, the following process is performed to certify the list sentence.

(1) The input example is recognized as a symbol sentence. (2) If it is a symbol sentence by the symbol sentence authorization routine,
The column position and type of the symbol part are obtained. Each for example
It is stored in memory such as colml and kindl. (3) Next, enter a new line. Hereinafter, the line currently being read will be referred to as the “current line” and the line immediately before it will be referred to as the “previous line”. (4) If the current line is a blank line, sentence end certification (construction 11) is performed (is being performed) by a blank sentence. Only separation (configuration 15) is performed. (5) If the character at the colml position on the current line is the same as the kind indicated by kindl, it is judged as a list sentence because the symbol sentences are continuous. At that time, sentence recognition (construction 14) and symbol part separation are performed on the immediately preceding line as a list sentence. (6) Also, the listflag is turned ON to indicate that the list sentence certification has started. In this case, enter the next line again and certify by the same process.

If the characters in the collm are spaces, the following three cases are possible. (1) When the list statement is written over multiple lines. “* HP VISTA, HP SINE, and HP Model Data Manager installation… 1 and update procedure.… 2” When reading the 2nd line, there is a space from the beginning of the line to colml, and it is not a symbol sentence, so go to the input of the next line. . (2) When the list sentence has a list sentence below. "* Event History ... 1 * System Exception reports, including ... 2-System Error Summary ... 3-Subsystem Exception reports for processors, channels, ... 4 DASD and tape device ... 5 * Threshold Summary for some tape device ... 6" Line 3 When is read, there is a space from the beginning of the line to colml and it is a symbol sentence, so the list sentence recognition process of line 2 is performed, the symbol position and character type are changed, and then the next line is entered.

(3) When a space happens to come. “-System Error Summary… 1－ Detail Edits of selected records… 2 All of these reports types are discussed in more detail in v… 3” When the line 3 is read, there are no spaces from the beginning of the line to colml, but 1, 2 and so on. Since the list sentence is being processed, the list sentence certifying process of the line 2 is performed, and the process returns to the check whether the line 3 is the list sentence.

By the above flow, for example, (2) in the above example
When a document such as is input, the result is as follows. (1) "*" (2) "Event History" (3) "*" (4) "System Exception reports, including" (5) "-" (6) "System Error Summary" (7) "-" ( 8) “Subsystem Exception reports for processors, channels,” DASD and tape device ”(9)“ * ”(10)“ Threshold Summary for some tape device ”

Next, an embodiment of the method of configuration (16) will be described. The source text is processed line by line. When the beginning is a number word enclosed in parentheses or a single letter of the alphabet, it is recognized as a symbol part. Accreditation of the sign part means that end-of-sentence recognition is performed immediately before and after the sign part. "The customer is responsible for (1) Adequate site and system planning and preparation. (2) Receipt, unpacking, and placement of the 5363 System Unit." In the above example, the part of "(1)" in the second line is It is recognized as a sign part, and the sentence ending with "for" on the first line is recognized.
Furthermore, the sentence end is recognized immediately after "(1)". In addition, since the third line is performed in the same manner, as a result, the sentence is recognized as follows. (1) "The customer is responsible for" (2) "(1)" (3) "Adequate site and system planning and preparation." (4) "(2)" (5) "Receipt, unpacking, and placement of the 5363 System Unit. ''

Next, an embodiment of the method of configuration (17) will be described. The source text is processed line by line. The beginning is a numeric word or an alphabetic character,
When any of closing brackets, period, colon or semicolon is added to it, it is recognized as a symbol part. Accreditation of the sign part means that end-of-sentence recognition is performed immediately before and after the sign part. "The customer is responsible for 1) Adequate site and system planning and preparation. 2) Receipt, unpacking, and placement of the 5363 System Unit." In the above example, the part "1)" in the second line is the symbol part. It is certified that there is, and the sentence end "for" is certified.
Furthermore, the sentence end is recognized immediately after "1)". In addition, since the third line is performed in the same manner, as a result, the sentence is recognized as follows. (1) “The customer is responsible for” (2) “1” ”(3)“ Adequate site and system planning and preparation. ”(4)“ 2) ”(5)“ Receipt, unpacking, and placement of the 5363 System Unit. "

Next, an embodiment of the system of the structure (18) will be described. The source text is processed line by line. When the beginning starts with a symbol character (such as an asterisk), it is recognized as a symbol part. Accreditation of the sign part means that end-of-sentence recognition is performed immediately before and after the sign part. "The customer is responsible for * Adequate site and system planning and preparation. * Receipt, unpacking, and placement of the 5363 System Unit." In the above example, the part marked "*" in the second line was identified as the symbol part. The first sentence, “for”, is recognized as the end of the sentence. Furthermore, the sentence end is recognized immediately after "*". In addition, since the third line is performed in the same manner, as a result, the sentence is recognized as follows. (1) "The customer is responsible for" (2) "*" (3) "Adequate site and system planning and preparation." (4) "*" (5) "Receipt, unpacking, and placement of the 5363 System Unit . ''

Next, an embodiment of the method of configuration (19) will be described. In the line where the sentence recognition is performed immediately before, if there are multiple spaces more than the specified value in the line, the part from the beginning of the line to that part is separated as a headword. For example, the specified value is set to 5. If five or more spaces appear consecutively, it is assumed that there is no syntactic connection between the words up to that point and the spaces after it, and the sentences up to immediately before the spaces are recognized. In addition, the sentence recognition being performed immediately before includes the beginning of document input. "CAUTION When you select [ABORT], all data is lost, the ~ acquired is set to 0, and you must select [START] ~ collection. ~" In the above example, "CAUTION" and "When" in the first line Since there are 5 spaces above the specified value between, the "When" portion is recognized as the entry word, and sentence end recognition is performed immediately after that. The flowchart is shown in FIGS. 12 (a) and 12 (b). However, the flow of the configuration (20) is also included here as an example.

Next, an embodiment of the system of the structure (20) will be described. In the line where the sentence recognition is performed immediately before, even if there are multiple spaces in the line that are equal to or greater than the specified value, the sentence is recognized up to immediately before the multiple spaces if either of the following conditions is satisfied. (1) When there is a sentence end symbol at the end of the line. (2) If the entire line is already end-of-sentence certification. (3) When there is a space from the beginning of the line to just below the last multiple spaces. (4) When the number of words up to a plurality of spaces is 1. In addition, the sentence recognition being performed immediately before includes the beginning of document input. "CAUTION When you select [ABORT], all data is lost, the 〜 acquired is set to 0, and you must select [START] 〜 collection. 〜" In the above example, the first line "CAUTION" and "When" There are only two spaces below the specified value between, but since the space from the beginning of the line to just below the last multiple spaces is "Whe
The "n" part is recognized as a headword, and sentence end recognition is performed immediately after that. FIG. 12 shows the flowchart. However, the flow of the configuration (19) is also included here as an example.

Next, an embodiment of the method of the structure (21) will be described. If the sign part is separated, it is not subject to parsing. "* Adequate site and system planning and preparation. * Receipt, unpacking, and placement of the 5363 System Unit." It is useless because it is not. Therefore, in order to reduce the burden of syntactic analysis, syntactic analysis is not performed, but it is only reflected directly in the Japanese sentence.

FIG. 13 is a block diagram for explaining another embodiment of the machine translation system according to the present invention. In the figure, 21 is a cathode ray tube (CRT), 22 is a keyboard, and 23 is an optical character reader (OCR). ), 24 is an input document, 25 is a spell check unit, 26 is a pre-editing unit, 27 is a translation body unit, 28 is a post-editing unit, 29 is a dictionary, 30 is a grammar rule, 31 is an output document, and 32 is a printer. . The input sentence obtained by any of file input, keyboard input, and OCR input is stored as the input document 24, and then the spell check unit 25.
Checks the spelling with reference to the dictionary 29, and the pre-editing unit 2
6 can be used to perform pretreatment. The translation main unit 27 is the dictionary 2
9 and the grammar rule 30 are used for translation, and the output sentence obtained as a result can be edited by the post-editing unit 28 using the translation information. The edited output sentence is stored in the output document 31.
The input sentence and the output sentence can be printed using the printer 32.

FIG. 14 is a diagram showing four processes of the translation main body unit in FIG. 8, in which 33 is a dictionary, 34 is an input unit,
Reference numeral 35 is a morphological analysis unit, 36 is a syntax analysis unit, 37 is a conversion unit, 38 is a generation unit, 39 is an output unit, and 40 is a grammar rule. The four main processings of the translation main body are that the morphological analysis unit 35 performs dictionary lookup of the input text,
The syntactic analysis unit 36 obtains the information of each word to obtain the grammar rule 4
0, parsing is performed to create a tree structure from the analysis result, the converting unit 37 transforms the input language tree structure into an output language tree structure, and the generating unit 38 translates the obtained tree structure for each node. It is to be.

The present invention belongs to the preprocessing (preediting) section or the morphological analysis section. Here, the input text is English. The sentence that is recognized here means one unit that is subjected to translation processing, and is slightly different from one sentence that is commonly known. Therefore, a title, a list, a headline, or a so-called sentence may not be prepared, but if it is a unit that should be translated by itself, it is a unit of translation processing and can be called a sentence.
When input text is subjected to translation processing, sentence recognition processing and dictionary lookup processing work first. Sentence recognition processing includes processing that uses only the morphological characteristics of text, and processing that performs dictionary lookup to grasp the morphological characteristics of words. In any case, it has a sentence recognition algorithm. Sentences of titles, lists, and headlines that deviate from the so-called sentences described earlier in this algorithm
Credits may or may not be accredited.

Consider, for example, the following simple sentence recognition algorithm. The following patterns are recognized as sentences (one unit of translation processing). 1. It ends with the end-of-sentence symbol (.?!;). 2. There are two line breaks at the end. 3. Within a certain number of words (e.g., 8 words), a newline follows, and the next character is non-lowercase (ignoring indentation). 4. At the beginning of the next line is an ordinal number or symbol for the list. 5. There are two or more spaces or tabs behind. 6. Exclusion rules Do not end with a pattern that consists only of numbers and periods (this is one of the ordinal numbers).

These are summarized as follows. Ordinary sentence rules 1 Title rules 2,3 List rules 4,6 Heading rules 5 This is an example, and the actual rules are more complicated than this. Especially 1 is considerably simplified. For example, abbreviations (such as Mr.) end with a period, but usually do not indicate the end of a sentence. It is necessary to pull dictionary information to process this. Looking at these rules, in the conventional sentence recognition, the ordinary sentence rule (1) or at most the title rule (2, 3) was used. In other words, the list rules (4, 6) and the heading rule (5) are new concepts of sentence recognition. If there are no steps in this sentence recognition rule, several sentence recognition results will be obtained. Results obtained by using some rules can be displayed as provisional sentence certification candidates, and sentence recognition results obtained by using some rules can be highlighted. The user can correct the sentence recognition result obtained by using a certain rule, or can refer to the sentence recognition result by another rule as a so-called another candidate and replace it there. That is, sentence recognition is performed on the input text using a sentence recognizing means having a sentence recognizing algorithm (belonging to the morphological analysis section or the preprocessing section), and the result is displayed on the display means. This state is shown in FIG.

FIG. 15 is a flow chart for explaining sentence recognition. Hereinafter, each step will be described in order. First, input text (step1) and perform sentence recognition processing (step2). The sentence certification candidates obtained as a result of the processing are displayed (step 3). You can also display other candidates for sentence recognition (step 4). It is also possible to highlight some sentence recognition candidates (step 6). The user corrects / selects the display of steps 3, 4, and 6 (step 5). If the user approves (step 7), the sentence certification ends. FIG. 16 is a block diagram of the sentence recognizing unit, in which 51 is a preprocessing / morphological analysis unit, 52 is a sentence recognizing unit, 53 is a candidate selecting unit, 54 is a displaying / editing unit, and 55 is a target document.

FIG. 18 is another block diagram of the sentence recognizing section. In the figure, 52a is a first sentence certifying means, 52b is a second sentence certifying means, and other parts which have the same functions as those in FIG. The same reference numerals are attached. It is possible to prepare a plurality of sentence recognition means such as (1) having an algorithm using only morphological features, and (2) additionally using dictionary information for reference. Both have advantages in speed and accuracy. (1) is superior in speed but slightly inferior in accuracy.
(2) is superior in accuracy but slightly inferior in speed. In other words, prepare a plurality of sentence recognition means (belonging to the morphological analysis unit or preprocessing unit) that has a sentence recognition algorithm for the input text, and perform sentence recognition using one or some of them, and the result Is displayed on the display means.

Next, the relationship between the list sentence and the preceding sentence (list introduction sentence) for guiding the list sentence described before the list sentence will be described. A description will be given of the translation in the case where the sentence introduction candidate information of the list introduction sentence, for example, ":" (colon), ";" (semicolon), ")" (closing bracket) is added. Here, the following two examples will be described in the case where a colon is added. First, "When:"
Will be described separately in the case where the sentence is at the end of the sentence, it is difficult to independently translate the list introduction sentence, and the case where “following:” is at the end of the sentence and the list introduction sentence is independently translatable.

(1) Syntactically connect a plurality of list sentences with a list introduction sentence (parent sentence).

[0048]

[Table 1]

Direct links are beneficial only when it is not possible ~. Direct links are beneficial only when two computers transfer ~. Direct links are beneficial only when two conputers are ~. It was done. In the example of Table 1, there are other cases such as proverbs such as be: verb for: such as are: and preposition find: such as verb he :.

(2) Although not syntactically linked, the following list sentence is a description of the list introduction sentence (parent sentence),
It indicates (semantic) that it is a supplement or a list of contents.

[0051]

[Table 2]

In the case of (2), the colon as the sentence end candidate information has a special role only when the word "following:" comes. Prediction code (syntactic role) in colon as an algorithm for translation
And translate it using normal grammar. To the list introduction sentence (parent sentence), a period is added to the end of the sentence, or a plurality of prediction codes obtained by adding a period prediction code to the prediction code given to the colon are given (FIGS. 19 to 22 described later). The flow chart shown in is in this form). The prediction code given to the colon is changed according to the part of speech of the word immediately preceding the colon. When the word before the colon is an idiom, take the part of speech of the entire idiom, not the component.

[0053]

[Table 3]

* 1 This prediction code shall not be continuous nouns. The reason is * 2 "mak
See the e: example. * 2 When the word immediately before the colon is a verb or auxiliary verb, and when it is a multi-part-of-speech word, give the period information in addition to the prediction code and leave it to the parsing. In the case of "will:", will is a multi-part word of an auxiliary verb and a noun. In this case, the colon has v's prediction code,
Give information as a period. Example "make:" make is a multipart word of a verb and a noun. In this case, the colon has an N prediction code,
Give information as a period. However, when make is taken as a noun, it is inconvenient to take a colon as a noun, so the prediction code for N is
It should not be a series of nouns. Also, “as follo
In the case of ws ", the colon is N. * 3 In the case of nouns, only" following "is treated separately. This is a multi-part-of-speech word, but the colon is treated only as a period. * 4 In the case of articles , One colon doesn't correspond to one prediction, so give multiple predictions.

Further, as shown in Table 4, translation information such as a translated word (or translated sentence) is given to the colon.

[0056]

[Table 4]

FIG. 19 and FIG. 20 are flow charts showing the processing after the list sentence guided by the list introduction sentence accompanied by a colon is recognized. Hereinafter, each step will be described in order. step1 : Check the surface layer of colon serial words. step2 : Check whether it is "following". step3 : If "following" is selected in step2, a period prediction is given to the colon. step4 : If it is not "following" in step 2, then it is checked whether it is "as follows". step5 : In the above step4, if "as follows", give the colon a "noun + period" prediction. step6 : If it is not "as follows" in step 4, then it is checked whether it is "to". step7: In the step6, if "to", give a prediction and both the prediction of "verb + period" of "noun + period" in the colon. step8: In the step6, "to" Otherwise, check the part of speech of the colon series language. step9 : Check whether it is a conjunction. step10 : In step 9, if it is a conjunction, the colon is given a "sentence" prediction. step11 : In the above step9, if it is not a conjunction,
Next, check whether it is a preposition. step12 : In the above step11, if it is a preposition,
Give the colon a "noun + period" prediction.

Step 13 : If it is not a preposition in step 11, it is next examined whether or not it is a verb. step14 : In step 13, if the verb is a verb, the colon is given a noun + period prediction. However, the prediction of this noun is a noun that is not continuous. step15 : If it is not a verb in step 13,
Next, check whether it is an auxiliary verb. step16 : In the above step15, if it is an auxiliary verb,
Give the colon the "verb + period" prediction. step17 : If it is not an auxiliary verb in step15, then it is checked whether or not it is a noun. step18: In step 17, if it is a noun, give the colon a "period" prediction. step19 : If it is not a noun in step17,
Next, check whether it is an article. step20: in the step19, if the article, give a prediction of "noun + period" in the colon. If it is not an article, it ends.

21 and 22 are flowcharts showing the translation processing of the list introduction sentence. The prediction code given to the colon is also given as the syntactic role of the list statement. This improves the translation accuracy of list sentences. The process of storing the prediction code given to the colon is performed immediately before the end of the flowchart. Next, when the sentence is a list of children of the previous sentence, the syntactical role of the child list sentence is specified in the stored prediction code, and parsing is performed. The flowcharts shown in FIGS. 21 and 22 are
Steps 1 to st of the flowchart shown in FIGS. 19 and 20
Up to ep20, it is the same, but the following points are different. That is,
Step 23 is provided as a storage unit A for storing the prediction code excluding the period given to the colon. Immediately after starting the flowchart, check whether it is a list statement (step2
1) If it is a list sentence, the list introduction sentence is translated with the storage unit A as a goal (step 22).

Next, at the end of the list sentence, "and""then"
The translation process when a connective word (connective or sentence adverb, etc.) such as “or” comes will be described. That is, a method of transforming a list sentence so that the subsequent translation process will be successful, focusing on a word that tends to occur at the end of the list sentence, will be described by the following embodiment. FIG. 23 shows a flowchart. Here, it is assumed that a list sentence as shown in Table 5 is given as an input sentence.

[0061]

[Table 5]

The input sentence is recognized by the sentence end recognizing means of the morphological analysis unit as shown in Table 6.

[0063]

[Table 6]

In the figure, the leftmost number is the sentence number added by the sentence recognizing means. Sentence numbers (2) (4) (6)
The symbol part is separated and processed as one sentence. Corresponding to each sentence number, an information flag indicating whether or not the sentence is a symbol part is set and stored as shown in Table 7.

[0065]

[Table 7]

First, it is checked whether or not there is a sentence to be read on the stack of the original sentence storage unit (step 1). If it exists, the top sentence of the stack is read (step 2), and it is checked by the flag information table in Table 7 whether or not the sentence is a sentence in the symbol part (step 3). If there is no sentence to read, the process ends. If the read sentence is a sentence in the symbol part, this sentence is taken out from the stack (step
4) Go to step A. If the sentence is not in the symbol part, it is checked whether the sentence end is a period (step 5). If it is a period, the sentence is taken out from the stack (step 4) and the process proceeds to step A. If the sentence end is not a period, it is checked whether or not a specific word exists immediately before that (step 6).
Specific words are connectives and sentence adverbs such as “and”, “then”, and “or” that tend to occur at the end of sentences in list sentences, and these words can be prepared in advance by the system. It is also possible for the user to set it before the translation process.

If the word at the end of the sentence does not belong to the previously designated word group, the sentence is taken out from the stack (step 4) and the process proceeds to step A. If the word belongs to the designated word group, the word is stored in the temporary storage unit, the word is deleted from the original sentence (step 7), this sentence is taken out from the stack, and the next sentence is read (step 8). In the case of Table 6, neither the sentence number (1) nor the sentence number (2) has the specified word at the end of the sentence. Since the word "then" exists in the sentence number (3), "then" is deleted from the sentence of (3). Here, the next sentence is read, and if the sentence is a sentence in the symbol part, it is taken out from the stack and the next sentence is read. If it is not a sentence in the symbol part, only the first character at the beginning of the previously stored word "then" is changed to upper case and added to the beginning of the sentence, the first letter of the sentence is converted to lower case, and step B is performed. Proceed (step 10). As a result, the sentences of sentence number (3) and sentence number (5) are as shown in Table 8. The sentence with the sentence number (5) is sent to step B.

[0068]

[Table 8]

In this way, the same processing is repeated, and the original input list sentence is transformed as shown in Table 9 and provided for the subsequent translation processing.

[0070]

[Table 9]

Tables 10 and 11 show the conventional translation result and the translation result of the present invention.

[0072]

[Table 10]

[0073]

[Table 11]

Next, in the case of a document in which image data is formed when reading with a scanner, the language processing of the title sentence and list sentence will be described. That is, the indent determination using the image information will be described below. FIG. 24 is a flowchart. First, the specification of the row range will be described. The image of the document having the format shown in Table 12 is read (step 1).

[0075]

[Table 12]

In the image, if there are no pixels in columns and they are arranged in parallel at equal intervals, there are no characters (that is, line spacing), and the blank columns are It can be determined that it is a horizontal writing document if it is arranged horizontally, and a vertical writing document if it is arranged vertically (horizontal writing in this example). If the space between lines is specified, the part in which the characters are written is also known, so the position of each line is measured in pixel units. Further, the range of rows can be easily specified by searching for a portion having no pixels in the vertical direction for horizontal writing and a portion having no pixels in the horizontal direction for vertical writing (step 2). After the position of the line is specified, the distance between lines is calculated (step 3). First, all line spacing is measured throughout the document. Since the lines should be arranged at equal intervals, the length in which the distribution is most concentrated in the measurement result is determined as the line-to-line distance. It is assumed that blank lines are inserted between lines that are extremely far apart from each other. The number of blank lines can also be easily estimated by comparing with the line width (step 4). This state is shown in Table 13.

[0077]

[Table 13]

Next, calculation of the distance between words and the distance between characters (step 5)
Will be described. After being divided for each row, a range without pixels in the vertical direction is searched for within the range of one row. In the case of English, the distance between words is longer than the distance between characters, so the length of the blank range within a line should be roughly divided into two. Therefore, the one with the longer blank length is determined as the inter-word distance and the one with the shorter blank length is determined as the inter-character distance. It may not be clear that there are two blank length distributions due to image reading errors, image noise, or blurring. May be determined to be a space between words or a space between characters. This allowable range can be set arbitrarily. This tabulation process is performed on all the rows.
For example, assume that the image positions shown in Table 14 are obtained.

[0079]

[Table 14]

From the tabulation result shown in Table 15, it is determined that the character distance is 2 [pixels] and the word distance is 5 [pixels]. When the inter-character distance and inter-word distance are determined, the position and length of the word can be known.

[0081]

[Table 15]

Next, the indent calculation (step 6) will be described. Further, a line starting from a position extremely longer than the interword distance can be determined to be indented. Similar to the inter-character distance and the inter-word distance, different indent depths (indent types) can be distinguished by searching for the indent distance over the entire document and obtaining the distribution of lengths.

Next, the document format estimation (step 7) will be described. If you want indentation, blank line, line length,
The format of the document can be estimated. For example: -The document start point or blank line immediately before the line-The line length is 2/3 or less of the maximum width of the document-If there is no indent, there is a high possibility that it is the title line. The linguistic properties of the title line include: -the beginning is likely to be a number or one letter of the alphabet-is likely to be capitalized for each word-is likely to be a noun phrase. Lines that are the items of a list (for convenience, are called list sentences) are adjacent to lines that have the same indentation immediately before or immediately after that. is there. The linguistic properties of the list sentence include: ・ The beginning is likely to be a number or symbol ・ It is likely to be a noun phrase.

Next, the word recognition (step 8) will be described. After estimating the formal characteristics of each line, the character recognition device (O
A word or character is recognized by collating the image in the range of the line with the standard pattern for recognition by (CR) or the like.
At that time, the recognition accuracy can be improved by adding the linguistic feature of the line to the matching score. For example, it is assumed that the recognition range is obtained as shown in Table 16 from the above example.

[0085]

[Table 16]

This line satisfies the following conditions which are the formal characteristics of the title line. -The document start point or a blank line immediately before the line-The line length is 2/3 or less of the maximum width of the document-There is no indent, so this line is likely to be the title line.
It is highly possible that it has the following linguistic features.・ There is a high probability that it will start with a number or a single letter ・ It is likely that each word has been capitalized ・ There is a high possibility of a noun phrase. If the case is similar, such as iS, you can decide which one to use. Next, the application to machine translation will be described. Even in machine translation, title sentences, list sentences, and the like cannot be correctly translated unless language processing different from general sentences is performed. Therefore, OCR
As a result, the translation rate can be expected to improve if not only the characters but also the formal characteristics of each line are added.

The above description is for the indent determination using the image information. Next, document format estimation using the indent depth will be described. First, the distribution of the indent depth will be described. As shown in FIG. 25, for documents using indents with different depths,
Detect indentation.

First, the distance of the start position of each line from the beginning (left end) of the file is calculated over the entire document.
The unit of the distance may be the number of blank characters (the number of columns) or the actual length ([mm], [inch], etc.). It can be said that the distance at which the distribution is concentrated in the tabulation result is the indent depth, and there are as many types of indents as there are concentrated points. For example, it is assumed that the distances have the distribution shown in Table 17 as a result of tabulation. In this distribution, 0, 10, 16, 22, and 28 are concentrated. Therefore,
This document has no indentation: 0 [column] 〃 depth 1:10 [column] 〃 depth 2:16 [column] 〃 depth 3:22 [column] 〃 depth 4:28 [column] Can be presumed to be used. Since the sentence numbers are also aggregated, it is easy to see from Table 17 how much each line is indented.

[0089]

[Table 17]

The results are shown in FIG. In the figure, b: indent depth 1:10 [column] a: 〃 depth 2:16 [column] c: 〃 depth 3:22 [column] d: 〃 depth 4:28 [column] Next, the document format analysis based on the indent depth will be described. The indentation depth represents the document nesting structure. For example, the indent of FIG. 27 (a) represents the document structure of FIG. 27 (b). In other words, if the indent depth of the line is the same as the previous line, the line is at the same level as the previous line.・ If the indent depth of the line is shallower than that of the previous line, the principle that the line is not related to the line immediately before is established.

Next, the specification of the contents of each range will be described. After estimating the range of the document structure based on the indent, the content of each range is further estimated. The contents include a title part, a body part, a list part, a headline part, and the like. The morphological characteristics of the title part are as follows: ・ The number of words that are composed is likely to be small ・ The beginning is likely to be a number or alphabet ・ The indentation depth is different from the line immediately before or after it ・ Indentation depth But there are shallow. The morphological features of the list part are: -It is likely that the number of words that are composed is small-It is likely that the beginning is a number, alphabet, or symbol-The indent depth is the same as the line immediately before or after, or both of them・ There is a deep indent. Similarly, the text portion, the headline portion, and the like also have morphological characteristics. In this way, the content of the document format can be estimated by using the indentation depth and the morphological characteristics of the line.

Next, the use of document format information in machine translation will be described. Even in machine translation, the title sentence,
List sentences cannot be translated correctly unless they are processed differently from ordinary sentences. Therefore, if the document format information is added to the translation unit at the time of analyzing the document format, the appropriate translation processing can be performed.
We can expect an improvement in translation rate. The present invention has described the case where the entire document has already been digitized. However, when specifying the document format, the morphological content of the line and
Since the document format of the line is determined based on the relationship between the line and its adjacent line, it is easy to extend the method described in the present invention to the case where the document is input line by line. is there. For that purpose, the line is not determined even after the input of the line is completed, and the immediately preceding line may be determined when the next line is input.

The above description is about the document format estimation using the indent depth. Next, the machine translation focusing on the table of contents will be described. First, the table of contents will be described. When translating a document by machine, translate the table of contents first. The table of contents section here refers to a set of lines composed of: chapter number, chapter title, number of pages, etc. Therefore, even if the electronic document does not have a table of contents, the method described in the present invention can be applied by extracting a chapter part from the text and creating a new table of contents.

[0094]

[Table 18]

Next, the use of the table of contents information will be described.Select word   The chapter title is a simple expression of the contents of the chapter.
Contains many frequently used words (important words) and jargon
Is expected to exist. Therefore, it is a set of chapters
Translate the table of contents before the text and confirm the translation result
By doing so, it is possible to reduce the translation load of the text section.
It For example, the table of contents in Table 18 is machine-translated first and the translated text is confirmed.
If you decide “Parallel Distributed Processing” ＝ “Parallel Distributed Processing”
Reason " “Distributed Representations” = “Distributed Representations” “PDP” = “PDP (parallel
(Disposal) “Cognitive Science” = “Cognitive science” It is possible to memorize translated word pairs such as. These simple
Words are important words and are keywords. this
If these words occur in the text, the
Word pairs can be used to give correct translations. In translation
It is easy to understand that technical words and important words are translated correctly.
It is very effective in improving sex. Also, it is the eyes to translate first
The same effect can be obtained not only in the next part but also in the index part.

[0096]Separation of chapter parts   The translation unit (character string) in the table of contents section is the chapter title in the text section.
Should appear as. The title section is different from normal English
And the symbols that indicate the end of the sentence (such as ".", "?", "!")
There are often no cases. Therefore, in machine translation,
It is difficult to automatically separate the part from the text part,
Authorize the character string in which the subject part and the body part are connected as one translation unit
Therefore, the translation process often failed. Therefore,
Perform the following steps to separate the chapter and body parts
To do. Specify the table of contents in the document to be translated.
  From the table of contents and the specified character string ・ Chapter number ・ Chapter ・ Number of pages The character string corresponding to is extracted and stored. Extraction work used
May be done manually by the person in charge, or the table of contents may have a uniform pattern.
Since it is expected that there will often be
Set of characters] [Set of alphanumeric characters] [Set of numbers] [End of line]
Automatic extraction may be performed according to any extraction pattern. Eye
Translate the next part and confirm the translation result. Final result is original translation
It is stored in association with the unit. Translate the text section. So
At the time of, the chapter number and title (in the number of pages if you know the number of pages
If there is a translation part that matches () also, it is judged as the chapter part
And separate it from the text section. Translated parts separated by
Without processing, the translation stored in is used as the translation for that part.
In addition, the chapter part is extracted from the text part to create the table of contents part.
For example, a string pattern, for example, [start of line] [collection of numbers
Specify], [Set of alphanumeric characters], [End of line], etc.
A dynamic search function may be provided.

FIG. 28 is a flow chart for explaining the machine translation focusing on the table of contents. Hereinafter, each step will be described in order. step1 : First, determine whether to translate the table of contents. step2 : When translating, specify the table of contents. step3 : Extract and store table of contents information. step4 : Translate the table of contents. step5 : Memorize the translation pair in the table of contents. step6 : Read the text part for each translation unit. step7 : Check whether the chapter part is included. step8 : If the chapter part is included, the chapter part is separated. step9 : Give the translation of the table of contents to the separated part. step10 : Translate the remaining part. step11 : It is checked whether or not the main body part ends. If it is not completed, the process returns to step 6, and if it is not completed, the process ends.

According to the above description, the input original sentence is recognized as a sentence at the time of morphological analysis, using information such as the character attributes and constituent elements of the text of the original sentence. As a sentence recognition method for a title sentence, when the target text (title) to be sentence-recognized is immediately preceding, the target text (title) starts from the beginning of the line and the first word is other than a preposition, article, or conjunction. Words, symbols, starting with capital letters
It is a number, and if the next line starts with a letter other than lower case, it is recognized as a title sentence. As a result of sentence recognition, even if the sentence immediately following the title sentence is connected to the title sentence to form one sentence, if the above conditions are satisfied, the title sentence and the next sentence continue. Translation processing such as analysis, conversion, generation, etc. is performed for each sentence as a separate sentence.

In the method of advancing the translation process while cutting out the input original text, sentence by sentence, a means for recognizing a sentence and a break between sentences is required. Especially recently
There is also a method of recognizing a document captured as an image by OCR and subjecting the text data obtained as a result to a translation process as it is. In such a case, it is necessary to have a means for recognizing a sentence and a break between sentences. Is important. However, in the conventional sentence end recognition method, if a title and the sentence immediately following it form a sentence (which is often seen in manuals, etc.), the result of sentence recognition at the time of morphological analysis indicates that the title part and the sentence immediately follow. Sentence is analyzed as a separate sentence and the translation process is executed, so even if the title part also serves as the first word of the sentence immediately after it, the next sentence following the title is a non-sentence ( It becomes a sentence with no subject) and analysis fails. In order to solve such a problem, it is preferable to use the embodiments described below.

FIG. 29 is a diagram showing an example in which a title part (eg, italic part) of a manual or the like also serves as the beginning of the next sentence (starting with Represents).
Is a flowchart. When you execute the translation process, "it
The "alic" part and the sentence starting with "Represent" are recognized as separate sentences in the morphological analysis (step 1). In the next process, the syntactic analysis (step 2), the "italic" part is analyzed as a title sentence. However, the sentence starting with "Represent ~" becomes a sentence without a subject and fails in parsing.When the parsing fails, it is judged whether the sentence immediately before it is a title sentence (step 4), and the title If it is a sentence, concatenate the title sentence and the analysis failure sentence (step
5) Then, the translation process is executed again. The result of concatenating the statements is "it
alic represents characters youtype from the keyboa
It becomes the sentence "rd", and as a result of executing the translation process again, the translated sentence "Italics represent characters input from the keyboard." is obtained.

FIG. 31 is a flow chart for explaining another embodiment. As in the example shown in FIG. 39, the "italic" part and the sentence starting with "Represent ~" are recognized as different sentences in the morphological analysis (step 2). Similarly, in the parsing of the next process (step 3),
The sentence starting with "Represent ~" is a parsing failure. By combining the sentence that failed to parse with the title part (step 7), “italic represents characters you type from t
The sentence "he keyboard" is created, and for the word at the beginning of the sentence (the title part) and the next word (the first word of the parse failure sentence), for example, the subject noun in the title part, the parse failure sentence The syntactic information of a pre-determined pattern such as a predicate verb is given to the first word of the step (step 8), and then the translation process is performed. Is deleted (step 11) and the translation process is performed again.

When the example sentence shown in FIG. 29 is translated by the above-mentioned method, a sentence starting with "Represents ~" is converted into a title "it".
Since the analysis is performed by linking with "alic", FIG.
The translation result as shown in (a) is obtained. The italic translation of the title part (・ italic) and the translation of the sentence at the beginning of the sentence immediately following it are also the same italics, and the same words are consecutive, so as shown in Fig. 32 (b) Omit the translation of and the particle that accompanies that translation. When the translated word of the title sentence and the translated word at the beginning of the sentence immediately after the title sentence are consecutive with the same word, the translated word at the beginning of the sentence immediately after the title sentence is replaced with a demonstrative pronoun (FIG. 32 (c) -p).

[0103]

As is apparent from the above description, the present invention has the following effects. (1) In claim 1, the original sentence is divided into a plurality of parts,
By translating each of them, it is possible to realize faster and more accurate translation, and to construct a system with good visibility. (2) In claim 2, since these pieces of information can be used during the translation processing, highly accurate translation can be performed. (3) In claim 3, by using these pieces of information at the time of generation, a correct translation can be obtained without being affected by the division processing. (4) In claim 4, since the dividing means is applied to the processing of the list sentence, the translation processing of the list sentence is facilitated. (5) In the configuration 5, by holding the translation in each translation unit, more detailed translation information can be obtained than in the translation of a single sentence, and the translation accuracy when the translation is included is improved. To do. (6) In configuration 6, if the user does not want the division processing, it can be canceled. (7) In the configuration 7, the user can operate the division processing unit, and the division processing desired by the user can be performed. (8) Since a plurality of end-of-sentence recognizing means are provided in claims 5, 6 and 7, flexible end-of-sentence recognizing is performed, and it is possible to select and use the end-of-sentence recognizing means suitable for the purpose and application. (9) In the configuration 11, since the empty sentence is recognized as the sentence end, the sentence end recognition failure when there is no sentence end symbol can be eliminated. In addition, the failure of end-of-sentence certification can be easily corrected manually. (10) Since the titles are recognizable in the configurations 12 and 13, it is possible to recognize a title sentence that is prominently written without an end-of-sentence symbol. (11) Since the list sentence is recognizable in the configuration 14, it is possible to recognize the list sentence that is remarkable without being accompanied by the sentence terminator. (12) In the configurations 15, 16, 17 and 18, since the symbol recognizing means is provided, it is possible to recognize a prominent headword that is written without an end-of-sentence symbol. (13) Since the headwords are separated in the configurations 19, 20 and 21, the prominent headwords that can be written without the end-of-sentence symbol can be recognized. (14) In the configuration 21, since the symbol part is removed from the elements of the syntactic analysis, the syntactic analysis load can be reduced and erroneous analysis can be prevented. (15) Effect corresponding to the configuration 22; The sentence recognition result of the first language can be confirmed once. (16) Effect corresponding to the configuration 23: The sentence recognition result of the first language can be confirmed once, and moreover, it becomes easier to confirm by emphasizing a specific part. (17) Effects corresponding to the configuration 24: The sentence recognition result of the first language can be confirmed once, and the user can select and correct it. (18) In claim 8, the syntactic role of giving the sentence end candidate information is changed depending on the part of speech of the list introduction sentence, and the translated word of the sentence end candidate information can be given according to the part of speech of the list introduction sentence. (19) In claim 9, the list sentence end connective word that is likely to occur at the end of the list sentence is specified, and the list sentence end connective word is added to the beginning of the next sentence that is not a symbol part. Becomes (20) According to the tenth aspect, since the indent determination is performed using the image information, the language processing of the title sentence, the list sentence and the like becomes easy. (21) In claim 11, it is determined whether or not the sentence immediately before the sentence that has failed to be analyzed is a title, and if it is a title, it is concatenated with the sentence immediately after that, and the sentence is analyzed as one sentence to obtain a correct sentence. You can get the translation result as. (22) In claim 12, by giving syntax information of a fixed pattern such as a part of speech, a subject, or a predicate to the title and the first word of the sentence immediately after the title, more accurate analysis and translation processing are performed, and the translation is performed. The accuracy can be increased. (23) In the configuration 30, if the parsing fails as a result of the parsing using the given syntactic information, the given syntactic information may not be applicable, so ignore the information and request again. By performing the processing of item 1,
You can get the translation result as a correct sentence. (24) By deleting the translated word of the part that also serves as the title of the translation result obtained in the structure 31 from the translated sentence, it is possible to avoid the continuation of the translated word at the beginning of the sentence that is the same as the translated word of the title portion, and to make the translated text easy to read. can do. (25) In the configuration 32, by replacing the part of the translated word described in (31) to be deleted with a pronoun, it is possible to improve the understandability of the translated text.

[Brief description of drawings]

FIG. 1 is a configuration diagram for explaining an embodiment of a machine translation system according to the present invention.

FIG. 2 is a configuration diagram of a translation unit.

FIG. 3 is a flowchart for explaining the operation of the morphological analysis unit.

FIG. 4 is a diagram showing a pattern matching rule.

FIG. 5 is a flowchart for explaining a pattern matching process.

FIG. 6 is a flowchart for explaining the operation of the edit control unit.

FIG. 7 is a diagram showing a separate translation display screen.

FIG. 8 is a configuration diagram for explaining an example of sentence end recognition of a machine translation system according to the present invention.

FIG. 9 is an overall configuration diagram for explaining a machine translation system.

FIG. 10 is a flowchart for explaining configurations (5) and (6).

FIG. 11 is a flowchart for explaining configurations (7) and (8).

FIG. 12 is a flowchart for explaining headword identifying means.

FIG. 13 is a configuration diagram for explaining another embodiment of the machine translation system according to the present invention.

14 is a configuration diagram for explaining a process of the translation main body unit in FIG.

FIG. 15 is a flowchart for explaining sentence recognition.

FIG. 16 is a configuration diagram of a sentence authorization unit.

FIG. 17 is a diagram showing application results of “Rules 1, 2, 3” and “Rules 1 to 6” for text examples.

FIG. 18 is another configuration diagram of the sentence authorization unit.

FIG. 19 is a flowchart showing a process (1) after a list sentence written in a list introduction unit accompanied by a colon is recognized.

FIG. 20 is a flowchart showing a process (1) after a list sentence written in a list introduction unit accompanied by a colon is recognized.

FIG. 21 is a flowchart showing a translation process of a list introduction sentence.

FIG. 22 is a flowchart showing a translation process of a list introduction sentence.

FIG. 23 is a flowchart for performing a translation process focusing on a list sentence end connecting word.

FIG. 24 is a flowchart for performing indent determination using image information.

FIG. 25 is a diagram showing an example of a document using a plurality of types of indents.

FIG. 26 is a diagram showing a result of indent detection.

FIG. 27 is a diagram showing an indent example and a document structure expressed by the indent example.

FIG. 28 is a flowchart for explaining machine translation focusing on the table of contents.

FIG. 29 is a diagram showing an example in which the title portion of a manual or the like also serves as the beginning of the next sentence.

FIG. 30 is a diagram showing a flowchart for connecting a title sentence and a sentence immediately after it to form one sentence.

FIG. 31 is a diagram showing another flowchart for connecting a title sentence and a sentence immediately after it to form one sentence.

FIG. 32 is a diagram showing a result of translation.

[Explanation of symbols]

1 ... Input unit, 2 ... Original sentence storage unit, 3 ... Translation unit, 4 ... Edit control unit, 5 ... Translation dictionary unit, 6 ... Translation sentence storage unit, 7 ... Display control unit, 8 ... Display unit, 9 ... Printing unit, 10 ... Dictionary editor.

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Masumi Narita             1-3-3 Nakamagome, Ota-ku, Tokyo Stocks             Company Ricoh (72) Inventor Miwa Shiraishi             1-3-3 Nakamagome, Ota-ku, Tokyo Stocks             Company Ricoh (72) Inventor Yoshihisa Oguro             1-3-3 Nakamagome, Ota-ku, Tokyo Stocks             Company Ricoh (72) Inventor Norikazu Ito             1-3-3 Nakamagome, Ota-ku, Tokyo Stocks             Company Ricoh (72) Inventor Sakiko Honma             1-3-3 Nakamagome, Ota-ku, Tokyo Stocks             Company Ricoh

Claims (12)

[Claims] 1. An input source language original text is analyzed and processed.
As a result, in a machine translation system that outputs a translated sentence of a target language, input data that forms an original sentence is divided into sentence units, and a dividing unit that divides the sentence into translation units, and a translation unit divided by the dividing unit. Machine translation which performs a separate translation process for each part and synthesizes the translation unit parts that have been subjected to the translation process at the time of translation generation to obtain a whole translation result. method. 2. The machine translation system according to claim 1, further comprising an analysis unit which gives a relationship between the division units at the time of the division processing and analyzes based on the division time information. 3. The machine translation system according to claim 1, further comprising control means for controlling a translation order and utilization so as to obtain an appropriate translated word by using the division time information. 4. The machine translation system according to claim 1, wherein the dividing unit is used to process a list sentence. 5. An input device for inputting an original sentence in a first language to a processing device, a display device for displaying information input by the input device, a processing device for processing the information, and an output in a second language. A machine translation system comprising an output device, wherein a plurality of sentence recognizing means for the text of the original sentence are prepared. 6. If the original text does not have a sentence terminator,
6. The machine translation method according to claim 5, wherein the sentence recognizing means is configured to perform sentence recognition based on character attributes of the original text and vertical and horizontal arrangement information of constituent elements of the original text. 7. The machine according to claim 6, wherein the sentence recognizing means comprises an empty sentence recognizing means, a title sentence recognizing means, a list sentence recognizing means, a symbol part recognizing means, and a headword recognizing means. Translation method. 8. An input device for inputting an original sentence in a first language to a processing device, a display device for displaying information input by the input device, a processing device for processing the information, and an output in a second language. In a machine translation system comprising an output device and a plurality of sentence certifying means for certifying a sentence of the original text, a serial word of sentence end candidate information in a list introduction sentence as a list sentence certifying means of the sentence certifying means. A machine translation method, wherein the syntactic role given to the sentence end candidate information is changed according to the part of speech, and a translated word of the sentence end candidate information is given according to the part of speech. 9. A list sentence end connecting word which is likely to occur at the end of the list sentence certified by the list sentence recognizing means is specified, and the list sentence ending connective word is added to the beginning of the next sentence which is not a symbol part. The machine translation system according to claim 8. 10. An input device for inputting an original sentence in a first language, a display device for displaying information input by the input device, a processing device for processing the information, and an output device for outputting in a second language. In a machine translation system equipped with, determine a horizontal document or a vertical document and determine the position and length of the word by determining the line range specifying means for measuring the line-to-line distance and the word-to-word distance and the character-to-letter distance. Calculating means, and an indent calculating means for obtaining an indent distance from the length of the distance between words, and determining an indent type from a distribution of the indent distance,
Document format estimating means for recognizing linguistic features such as title lines or list sentences based on the type of indent determined by the indent calculating means, and characters for recognizing words or characters based on the document format estimating means A machine translation method comprising a recognition means. 11. An input unit for inputting an original sentence in a first language, a translation dictionary unit having a translated word in a second language corresponding to a word in the first language and semantic information, and the input unit. Morphological analysis unit for morphologically analyzing the original text, a syntax analysis unit for parsing the original text input by the input unit, a conversion / generation unit for converting / generating into the second language, and conversion by the conversion / generating unit -In a machine translation system including an output unit that outputs a generated translated sentence, if a sentence that has failed to be analyzed is a title sentence that also serves as the title of the sentence, the analysis failure sentence and the immediately preceding title sentence are concatenated. The machine translation method is characterized by performing translation processing as if it were one sentence. 12. When the sentence immediately before the sentence that failed in the analysis is a title sentence that also serves as the title of the sentence, the part of speech, the subject, the predicate, etc. for the first word of the sentence for which the analysis failed and the title portion immediately before the sentence 12. The machine translation system according to claim 11, wherein the translation processing is performed by giving the syntax information of.