JP4531325B2 - Machine translation apparatus and machine translation method - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a machine translation apparatus and a machine translation method, and more particularly, to divide a sentence in a source language into phrases and clauses and output a phrase or clause in a target language obtained by translating each phrase or clause. The present invention relates to a machine translation apparatus and a machine translation method capable of performing the above.
[0002]
[Prior art]
In many existing machine translation apparatuses, translation processing is executed in units of one sentence. Therefore, the longer the length of one sentence, the higher the possibility that the syntax will be erroneously analyzed. If a long sentence is translated in units of one sentence, the possibility of outputting a meaningless translated sentence increases. Even if the syntax can be parsed correctly, a translated sentence that is logically correct but very difficult to read is often output. Therefore, the current machine translation apparatus is mainly used in such a way that a user edits the output result for use.
[0003]
In addition, it is better to obtain a translation result in a unit smaller than one sentence in a state where there are few parse errors, rather than obtaining a translation result in one sentence unit when there is a high possibility that the syntax analysis is incorrect. There has been proposed a machine translation apparatus that outputs a translation result in a unit smaller than one sentence such as a word, a phrase or a clause without outputting the translation result in one sentence unit. Examples of such machine translation devices include the following.
[0004]
That is, the dictionary lookup, morpheme analysis, and translation selection similar to those of a translation unit in a normal sentence unit are performed, and the syntax structure of the original sentence and the translation word for each word of the original sentence are obtained. After that, when generating a translation, the source text is divided into phrases and sections based on the syntactic structure, and the words in the source language that make up the divided source phrases and sections (hereinafter referred to as source words) are supported. The translations to be translated are grouped into phrases and sections of the original text, and the translations within each phrase and section in the appropriate order (only the translations belonging to the phrase or section are extracted from the order of translations when translating one sentence at a time) The translated phrases are assembled in the order obtained).
[0005]
Here, the phrase “phrase” of the translated phrase does not mean “phrase” in English grammar, but is a phrase used to control the above “phrase” and “section”. The same applies to the phrase “phrase” used below.
[0006]
[Problems to be solved by the invention]
However, the machine translation apparatus that outputs the translation result in a unit smaller than the conventional one sentence has problems as shown in Examples 1, 2, and 3.
[Example 1]
Original
"It would push them to the end of the creditors' line
and make it unlikely. ''
Results of dividing the original text into phrases
(1) It would push them to the end of the creditors' line
(2) and
(3) make it unlikely.
Phrase translation results
(A) It puts them at the end of the creditor's line
(B) and
(C) will make it unlikely
[0007]
In this example, when the original text and the translation result are compared in phrase units, the auxiliary verb “would” exists in the original text in (1), but this is not reflected in the translated text in (a). . The effect of the auxiliary verb “would” appears in the form of “to be” only in the translated phrase of (c). The reason for this is that the result of normal translation processing per sentence is
“It will put them at the end of the creditor's line and make it unlikely.”
This is because the effect of the auxiliary verb appears only at the end of the sentence, and the phrase division position of the original sentence is faithfully reflected in the translated sentence.
[0008]
However, comparing the original phrase in (1) with the translated phrase in (a), it is not natural that the effect is not shown in the translated phrase even though the original phrase has the auxiliary verb “would”. It is. Semantically, the auxiliary phrase “would” of the original phrase is effective for both of the two clauses of the original sentence that share the subject and are in parallel. In some cases, it is only necessary to change the form of the translation at the end of the sentence, but in the case of phrase translations where each section is separated, it is not necessary to change the translation form explicitly for each section. It's not a good translation.
[0009]
[Example 2]
Usually, when a machine language translation device generates a target language translation from a source language source, several translations that do not correspond to any word in the source language are generated in addition to the source language translation. Although used, the generation position of a translation that does not correspond to such a source language word is determined by a very simple rule. For example, if the target language is Japanese, a translation that does not correspond to a word in the target language is generated in the same phrase as the previous translation, based on the translation generation order when translated in single sentences. It is supposed to.
[0010]
When the translation process from English to Japanese is performed according to such rules, the following results are obtained.
Original
`` Keating and other Lincoln exectives are the subject of
a class-action lawsuit alleging they misled investors about
the safety of junk bonds issued by American Continental. ''
Results of dividing the original text into phrases
(1) keating and other Lincoln exectives are the subject of
a class-action lawsuit
(2) alleging they misled investors about the safety of junk bonds
issued by American Continental
Phrase translation results
(A) Keating and other Lincoln executives are the subject of the class action that he claims
(B) they misled investors about the safety of Jack Bond issued by American Continental
[0011]
In this case, it is difficult to understand that “assert” in the translated phrase of (a) takes the object, and that the object is the entire translated phrase of (b). In order to improve this, the case particle “to” at the end of the translated phrase in (b) and receiving the entire translated phrase in (b) is replaced with the “assertion” of the translated phrase in (a). It becomes easier to understand if it is generated immediately before. However, as described above, the case particle “to” is a word that does not correspond to the word of the original phrase, and therefore, when translated in units of one sentence, the translated word “mistake” generated immediately before “to”. Are generated in the same phrase as (b).
[0012]
[Example 3]
Original
`` The plan was opposed by Keating and some of his former top
executives, who invested in the company's securities. ''
Results of dividing the original text into phrases
(1) The plan was opposed by Keating and some of his former top executives
(2) who invested in the company's securities
Phrase translation results
(A) The plan was opposed by Keating and some of his former top executives
(B) Invested in company securities
[0013]
In this example, the original phrase in (2) is a related clause, but if you look at only the translated phrase in (a), it is a clause that lacks the case (in this case, the action main case), so this is all Then it is difficult to take a meaning as a phrase.
[0014]
Therefore, the object of the present invention is to eliminate the above-mentioned problems that hinder the comparison between the original text and the translation result in units of phrases or to read only the translated phrase, and for the user to grasp the contents of the translated sentence. An object of the present invention is to provide a machine translation apparatus and a machine translation method that can be made easier.
[0015]
[Means for Solving the Problems]
In order to achieve the above object, a machine translation apparatus according to the first aspect of the present invention provides:
Language generation means for generating a translated word string of the input source text composed of source language words;
Division position determining means for determining a position at which the input original is divided into phrases or clauses;
A translated word storage means for acquiring the translation of each word constituting the clause divided at the position determined by the divided position determining means from the generation result by the language generating means, and storing the translated word for each clause and in units of words;
A clause that is divided at the position determined by the division position determination means and that has a case element missing. According to the division rules A relative clause detection means to detect;
An antecedent acquisition unit for acquiring an antecedent of a relative in a relative clause detected by the relative clause detection unit;
The translation of the antecedent obtained by the antecedent acquisition means is acquired from the noun translation storage means, and the missing words in the translations of a plurality of words constituting the relevant clause stored in the nodal translation storage means A relative translation generation means for generating a translation of the relative at the position of the case element;
It is characterized by having.
[0016]
In the above configuration, when the relative clause that is separated at the position determined by the division position determination unit and lacks the case element is detected by the relative clause detection unit, the antecedent of the relative clause is obtained by the antecedent acquisition unit, A translation of the antecedent is acquired by a relative translation generation means, and the relative word of the relative is placed at the position of the missing case element in the translation of a plurality of words constituting the relative clause stored in the clause storage. Generated as a translation.
In this way, the translation of the antecedent is inserted as the translation of the relative clause into the translation of the relative clause separated by the division position determining means and lacking the case element, and translation is performed for each clause.
[0017]
The invention according to claim 2 is the machine translation device according to claim 1,
A translation of the relative verb generated in the translation of a plurality of words constituting the relative clause in the relative translation storage unit by the relative translation generation unit, and the relative verb stored in the clause translation storage unit It is characterized in that the translation of the antecedent in the clause containing the antecedent is the same.
[0018]
The invention according to claim 3 is the machine translation device according to claim 1 or 2,
Immediately after the translation of the relative term generated by the relative term translation generation unit, postfix generation means for generating a postfix representing the case of the translation of the relative term is provided.
[0019]
The invention according to claim 4 is the machine translation device according to any one of claims 1 to 3,
The apparatus further comprises auxiliary modifier generation means for generating a modifier related to the translation of the relative verb generated by the relative translation generation means and improving the readability of the clause in which the translation of the relative is generated. It is said.
[0020]
The machine translation method of the invention according to claim 5 is:
A language generation step of generating a translated word string of an input original composed of source language words by a language generation means;
A division position determination step for determining a position at which the input original text is divided into phrases or clauses by a division position determination means;
The translation of each word constituting the clause divided at the position determined by the division position determination means is obtained from the generation result by the language generation means, and is stored in the phrase translation storage means for each clause and in units of words. A translation storage step;
The related clause detection means is a clause that is divided at the position determined by the division position determination means and lacks a case element. According to the division rules A relative clause detection step to detect;
An antecedent acquisition means for acquiring an antecedent of a relative in the relative clause detected by the relative clause detection means by an antecedent acquisition means;
A plurality of words constituting the related clause stored in the noun translation word storage means obtained by the noun translation generation means, obtained from the noun translation word storage means, obtained by the antecedent acquisition means. Step of generating a relative translation at the position of the missing case element in the translation of
It is characterized by having.
[0021]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, the present invention will be described in detail with reference to the illustrated embodiments.
FIG. 1 is a block diagram showing an embodiment of the machine translation apparatus of the present embodiment. In this embodiment, the machine translation apparatus will be described by taking as an example the case where the source language is English and the target language is Japanese.
[0022]
In FIG. 1, the machine translation apparatus is configured by connecting an input unit 1, a translation control unit 2, a main memory 3, a storage unit 4, an output unit 5 and a control unit 6 via a bus line 7.
The input unit 1 is composed of an input device such as a keyboard, mouse, electronic pen, tablet, scanner, character recognition device, or a communication device connected to a communication line. Entered.
The translation control unit 2 includes a translation module 2A and an output format shaping unit 2B. The translation control unit 2 performs translation processing in units of phrases on the original text in the source language input from the input unit 1 to generate a translation in the target language. .
In the main memory 3, various buffers for holding data generated in the process in which the translation control unit 2 generates a translation sentence are set.
The storage unit 4 stores data such as a dictionary for translation, grammar rules, and other rules necessary for translation, which are used when the translation control unit 2 generates a translation.
The output unit 5 is composed of a display device such as a CRT (cathode ray tube) or LCD (liquid crystal display device), a printing device such as a printer, or a communication device connected to a communication line, and generates a generated translation or the like. Output.
The control unit 6 is mainly configured by a CPU (Central Processing Unit), and controls the input unit 1, the translation control unit 2 and the output unit 5 to execute various processes such as a translation process.
[0023]
In general, machine translation has an analysis level as shown in FIG. When an original text in the source language is input at the upper left in the figure, level L1 dictionary lookup, level L2 morphological analysis, level L3 syntax analysis, and so on are advanced. Machine translation is roughly divided into the following two types depending on the level of analysis to be performed.
The first analyzes the intermediate language, which is a concept that does not depend on either the source language or the target language of level L6, and generates context from the target language of level L7, semantic generation of level L8, syntax generation of level L9 This is a pivot method in which the generation and generation of level L10 morpheme is advanced to generate a translation in the target language.
The other is to analyze the level L2 morphological analysis, level L3 syntax analysis, level L4 semantic analysis and level L5 context analysis to obtain the internal structure of the source language, This is a transfer method for generating a translated sentence of the target language after converting into the internal structure of the target language at the same level as the obtained internal structure of the source language.
[0024]
Hereinafter, each analysis will be described in order.
(A) Dictionary lookup and morphological analysis
The dictionary stored in the dictionary memory 4a of the storage unit 4 is subtracted to divide the input sentence into morpheme strings (word strings). Then, grammatical information such as parts of speech of each divided word and translations are obtained, and the tense, person, number, etc. are analyzed.
(B) Parsing
The structure (structure analysis tree) of the input sentence such as dependency between words obtained as a result of the morphological analysis is determined.
(C) Semantic analysis
From the plurality of syntax candidates obtained as a result of the above parsing, the semantically correct one is discriminated from the one that is not.
(D) Context analysis
Understand the topic and remove omissions and ambiguities.
It is assumed that the translation module 2A in this embodiment executes the translation process by a transfer method for performing analysis up to at least level L3 syntax analysis in FIG.
[0025]
As shown in FIG. 1, the translation module 2A has a dictionary lookup / morpheme analysis unit 2a, a syntax analysis unit 2b, a syntax conversion unit 2c, and a target language generation unit 2d in order to execute the above-described translation processing. is doing. Furthermore, each translation phrase is constructed from a division position determination unit 2e that determines a division position for dividing a sentence into phrases by referring to data being translated, and a target language word generated by the target language generation unit 2d. It has a translated phrase construction unit 2f.
[0026]
The main memory 3 includes a source buffer 3a, a dictionary buffer 3b, a pre-conversion syntax buffer 3c, a post-conversion syntax buffer 3d, a translation buffer 3e, a source / translation correspondence information buffer 3f, a division position buffer 3g, and a phrase. For each word, a translated word position buffer 3h, a translated phrase buffer 3i, an output format buffer 3j, an auxiliary verb buffer 3k, a phrase translated flag buffer 3l, and a tree structure pointer buffer 3m are set. In addition to the dictionary memory 4a, the storage unit 4 stores a grammar rule memory 4b in which grammar rules used by the syntax analysis unit 2b during parsing, a postfix table memory 4c, an auxiliary translation table memory 4d, Other translation rule memory 4e, phrase division rule memory 4f, and auxiliary verb code table memory 4g are set.
[0027]
Furthermore, as shown in FIG. 2, the translation module 2A in the present embodiment generates a tree structure trace in order to generate a translated phrase whose contents are accurate and easy to understand when translating in phrase units. Means 11, node information acquisition means 12, auxiliary verb sentence translation method determination means 13, auxiliary verb detection means 14, parallel clause detection means 15, dictionary description translation generation means 16, auxiliary word generation means 17, and auxiliary verb translation generation means 18. Yes.
Further, as shown in FIG. 3, the translated phrase constructing unit 2f includes a missing phrase detecting means 21, a related clause detecting means 22, a generating position correcting means 23, an auxiliary symbol generating means 24, an auxiliary modifier generating means 25, An antecedent acquisition unit 26 and a postfix generation unit 27 are provided.
[0028]
FIG. 4 is a flowchart of the phrase unit translation processing operation executed by the input unit 1, translation control unit 2, and output unit 3 under the control of the control unit 6.
The phrase unit translation processing operation in the present embodiment will be described below with reference to FIG.
[0029]
In step S1, the original text is input from the input unit 1. If it does so, the character string which comprises the said original text will be recognized by the input part 1, and an input character string will be divided | segmented into a partial character string based on a blank character (space). Then, each divided partial character string (morpheme candidate) is stored in the original sentence buffer 3a with position information added as shown in FIG.
Further, a translation process (hereinafter referred to as phrase translation) in units of phrases is commanded from the input unit 1 and recognized by the input unit 1. Then, the phrase translation flag is set in the phrase translation flag buffer 3l of the main memory 3 in accordance with the recognition result “phrase translation”.
In step S2, the dictionary lookup / morpheme analysis unit 2a of the translation module 2A executes the dictionary lookup / morpheme analysis process on the original sentence stored in the original sentence buffer 3a as described above, and obtains the obtained analysis. The result is stored in the dictionary lookup buffer 3b of the main memory 3.
In step S3, the syntax analysis unit 2b of the translation module 2A executes the syntax analysis process to determine the structure analysis tree of the original sentence. Then, the obtained structural analysis tree is stored in the syntax buffer 3 c before conversion in the main memory 3.
At step S4, the syntax conversion unit 2c of the translation module 2A converts the structure parse tree of the original sentence stored in the syntax buffer 3c before conversion into a Japanese structure parse tree, Stored in the syntax buffer 3d.
[0030]
In step S5, the target language generation unit 2d in the translation module 2A executes a target language generation subroutine, which will be described in detail later, to determine the translation of each word of the original sentence. Each determined translation is stored in the translation buffer 3e of the main memory 3. In addition, source / translation correspondence information is stored in the source / translation correspondence information buffer 3 f of the main memory 3.
In step S6, the division position determination unit 2e of the translation module 2A determines the division position of the original sentence into phrases and stores it in the division position buffer 3g of the main memory 3.
In this determination of the division position into phrases, a phrase division rule as shown in FIG. 8 stored in the storage unit 4 is applied to the structural analysis tree stored in the syntax buffer 3c before conversion. Is done by. In this case, the actual operation of determining the division position is to search the nodes of the structural analysis tree in a predetermined order, check whether each phrase division rule is applicable at each node, and if applicable, that node. Is determined as a phrase division position.
[0031]
In step S7, the translated phrase construction unit 2f of the translation module 2A converts the source / translation word correspondence information created in step S5 (see FIG. 15) and the division position information created in step S6 (see FIG. 16). Based on this, the translated words constituting each phrase (hereinafter referred to as translated words) are obtained. Then, referring to the translated word buffer 3e, a word position (translated word position) on each translated word is obtained and set in the translated word position buffer 3h for each phrase in the main memory 3. The format of the translated word position buffer 3h for each phrase is shown in FIG.
In step S8, the translated word position of the translated words determined as a result of the execution of the target language generation subroutine in step S5 by the translated phrase constructing unit 2f has not yet been set in the translated word position buffer 3h for each phrase ( In other words, with respect to the translated word (for which the phrase to which it belongs is not determined), as described in detail later, the phrase to which the translated word belongs is determined, and the translated word of the translated word is assigned to the corresponding phrase in the translated word position buffer 3h for each phrase. The word position is set.
[0032]
In step S9, the translated phrase construction unit 2f determines whether or not each translated word position set in the translated word position buffer 3h for each phrase needs to be corrected, and the translated word position that needs to be corrected. If exists, the corrective action is performed.
The correction operation is performed as described in detail later according to the correction rules stored in the other translation rule memory 4e of the storage unit 4.
[0033]
In step S10, the translated phrase constructing unit 2f refers to the translated word buffer 3e based on the translated word position for each phrase stored in the translated word position buffer 3h for each phrase created in steps S7 to S9. Then, the translated word string for each phrase is obtained, and the obtained translated word string for each phrase is stored in the translated phrase buffer 3i.
In step S11, the output format shaping section 5B prepares the format of the translated word string for each phrase stored in the translated phrase buffer 3i and stores it in the output format buffer 3j.
For example, as shown in FIG. 19, the output format is such that the starting positions of the original sentence and the translated sentence of each phrase are aligned, and the correspondence between the original phrase and the translated phrase that is the translated sentence is clear. Format.
In step S12, the output unit 5 outputs the original phrase and the translated phrase stored in the output format buffer 3j as shown in FIG.
[0034]
FIG. 5 is a flowchart of a target language generation subroutine executed by the target language generation unit 2d in step S5 in the flowchart of the phrase unit translation processing operation shown in FIG. Hereinafter, the target language generation subroutine will be described with reference to FIG.
[0035]
When the original structural analysis tree is converted into the Japanese structural analysis tree in step S4 in the flowchart of the phrase unit translation processing operation shown in FIG. 4, the target language generation subroutine starts.
In step S21, the parallel clause detection means 15 of the target language generator 2d determines whether or not there are parallel phrases or clauses in the Japanese structural parse tree obtained by the syntax converter 2c. The As a result, if it exists, a node corresponding to the parallel phrase or parallel node in the structural analysis tree is pointed and the process proceeds to step S23. On the other hand, if it does not exist, the process proceeds to step S22.
The above-mentioned “pointing to a node” means storing the node number in the structural analysis tree in the tree pointer buffer 3m of the main memory 3. In this way, information on nodes that are currently focused on is retained.
In step S22, the dictionary description translation generation means 16 and the attached word generation means 17 determine the translation of each word of the original sentence based on the translation stored in the dictionary lookup buffer 3b and store it in the translation buffer 3e. Also, the original / translated word correspondence information buffer 3f stores the original / translated word correspondence information.
After that, the process returns to step S6 in the flowchart of the phrase unit translation processing operation shown in FIG.
[0036]
In step S23, the contents of the auxiliary verb buffer 3k in the main memory 3 are cleared.
In step S24, the tree structure tracing means 11 determines the positional relationship between the lower nodes by referring to the parts of speech and grammatical elements of the lower nodes acquired by the node information acquisition means 12. Will point to the node corresponding to the first child of the currently pointed node.
[0037]
In step S25, the auxiliary verb detection means 14 determines whether there is an auxiliary verb at the currently pointed node. As a result, if there is an auxiliary verb, the process proceeds to step S26, and if not, the process proceeds to step S27.
The auxiliary verb is detected as follows. That is, the auxiliary verb here is an auxiliary verb on the English structural analysis tree, and therefore does not exist on the structural analysis tree after conversion (Japanese). Therefore, the tree structure tracing means 11 and the node information obtaining means 12 are controlled, and on the structure analysis tree before conversion (English) corresponding to the pointed node on the structure analysis tree after conversion (Japanese). This is done by tracing the nodes.
In step S26, the auxiliary verb code of the auxiliary verb detected as described above by the auxiliary verb detecting means 14 is referred to the auxiliary verb code table stored in the auxiliary verb code table memory 4g of the storage unit 4 in the format shown in FIG. Is obtained. The obtained auxiliary verb code is stored in the auxiliary verb buffer 3k.
[0038]
In step S27, the auxiliary verb sentence translation method determining means 13 refers to the contents of the phrase translation flag buffer 3l to determine whether or not a phrase translation is commanded. As a result, if a phrase translation is commanded, the process proceeds to step S28, and if not, the process proceeds to step S34.
In step S28, the dictionary description translated word generating means 16 and the attached word generating means 17 generate a translated word string suitable for the original word string belonging to the currently pointed node based on the translated word stored in the dictionary lookup buffer 3b. It is determined and stored in the translated word buffer 3e. At that time, the auxiliary verb translation generating means 18 transforms the end into an end form corresponding to the auxiliary verb code stored in the auxiliary verb buffer 3k. However, if the auxiliary verb code is not stored in the auxiliary verb buffer 3k, it is simply transformed into a final form.
[0039]
Here, the attached word generation means 17 generates a translated word that does not correspond to the original word. As an example of the generation method, there is a method of using information on the case structure. In this method, when generating a translated word corresponding to a source word, the case information of the target language (Japanese) is referred to by referring to the case information of the source word stored in the dictionary memory 4a. This is to determine the ancillary words.
[0040]
In step S29, the tree structure tracing unit 11 refers to the part of speech or grammatical element acquired by the node information acquiring unit 12 to determine whether or not the currently pointed node has a brother node. . As a result, if there is a younger brother node, the process proceeds to step S31, and if not, the process proceeds to step S30.
In step S30, the contents of the auxiliary verb buffer 3k in the main memory 3 are cleared.
After that, the process returns to step S6 in the flowchart of the phrase unit translation processing operation shown in FIG.
[0041]
In step S31, the tree structure tracing means 11 points to the brother node determined to be present in step S29.
In step S32, the dictionary description translation generation means 16 generates a translation of the conjunction at the brother node and stores it in the translation buffer 3e.
In step S33, the tree structure tracing means 11 points to the next brother node (that is, the node connected in parallel to the first child node by the connective).
After that, the process returns to step S25 to shift to the process for the brother node.
[0042]
If it is determined in step S34 that the phrase translation is not instructed in step S27, the tree structure tracing means 11 and the node information obtaining means 12 determine whether there is a brother node at the currently pointed node. Is determined. As a result, if there is a younger brother node, the process proceeds to step S35, and if not, the process returns to step S28 to generate a translation of the current node.
In step S35, the dictionary description translated word generating means 16 and the attached word generating means 17 determine a translated word string suitable for the original word string belonging to the currently pointed node and store it in the translated word buffer 3e. However, in this case, the end of the clause is transformed into a continuous stop type that does not reach the effect of the auxiliary verb, regardless of whether or not the auxiliary verb code exists in the auxiliary verb buffer 3k.
After that, the process returns to step S31, a point is set at the brother node determined to exist at step S34, and the process proceeds to the process for the brother node. If it is determined in step S29 that there is no longer a brother node, the process returns to step S6 in the flowchart of the phrase unit translation processing operation shown in FIG. 4 via step S30.
[0043]
In the following, the input section 1 includes a parallel clause including an auxiliary verb.
"It would push them to the end of the creditors' line
and make it unlikely. ''
The phrase unit translation processing operation shown in FIG. 4 and FIG.
[0044]
First, the original text read from the input unit 1 is stored in the original text buffer 3a as shown in FIG. Then, dictionary lookup / morpheme analysis processing and syntax analysis processing are performed to obtain a structure analysis tree of the original text as shown in FIG. 9 and stored in the syntax buffer 3c before conversion. Further, a syntax conversion process is performed to obtain a Japanese structural analysis tree as shown in FIG. 10, and the result is stored in the converted syntax buffer 3d.
... Steps S1 to S4
[0045]
Furthermore, since there is a parallel node in the Japanese structural analysis tree, the node number “1” (see FIG. 10) of the “parallel node” is stored in the tree structure pointer buffer 3m. The node “main node (node number = 2)”, the node “connective (node number = 18)” and the node “main subject missing main node (node number = 19)” belonging to the node “parallel node” Of the child nodes, the node “main node” is determined to be the first child, and the node number “2” of the node “main node” is stored in the tree structure pointer buffer 3m.
... Steps S21, S23, S24
[0046]
Here, the node in the structural analysis tree (FIG. 9) before conversion corresponding to the node “main node (node number = 2)”, which is the first child, is the node “main node (node number = 2)”, and its lower level Has a node “auxiliary verb (node number = 6)”. Therefore, the auxiliary verb code “9” of the auxiliary verb “would” corresponding to the node “auxiliary verb (6)” is obtained by referring to the auxiliary verb code table shown in FIG. 11, and set in the auxiliary verb buffer 3k as shown in FIG. Is done.
... Steps S25 and S26
[0047]
Since “phrase translation” is commanded, the source word string belonging to the currently pointed node “main clause (2)” in the converted structural analysis tree shown in FIG.
"It them creditors' line of end to push"
Thus, the translated word string of the node “main clause (2)” is determined based on the translated word stored in the dictionary lookup buffer 3b. In this case, the end of the clause is transformed into a final form “~” according to the auxiliary verb code “9”. As a result, the translation for each source word belonging to the node “main clause (2)” in the source sentence is
"That"
"They"
"a creditor"
"of"
"line"
"of"
"End"
"I will put it"
Is determined. In addition, translations that do not correspond to the original words
","
"Ha"
"O"
Is generated and set in the translated word buffer 3e as shown in FIG.
Here, as described above, the translated word that does not correspond to the original word is generated by referring to the case information stored in the dictionary memory 4a of the storage unit 4 by the attached word generating means 17, specifically, 34 pays attention to the contents of the “case name” stored in the dictionary memory 4a as shown in FIG. 34, and the attached words corresponding to the contents of the “case name” are placed after the storage unit 4 as shown in FIG. It is generated from the word table 4c.
... Steps S27, S28
[0048]
If the above phrase translation is not specified, the node “main clause (2)” currently pointed to includes the node “connective (18)” and the node “subject missing main clause (19)”. Since there is a younger brother node, the translation of the source word string belonging to the node "main clause (2)" whose end has been transformed into a continuous discontinuation form that does not reach the effect of the auxiliary verb "would" is determined and the translation buffer 3e Will be stored. Therefore, as shown in FIG. 14, the translation of the original word “push” in the original sentence is “insert”.
... Steps S27, S34, S35
[0049]
After that, the brother node “connective (18)” exists at the node “main clause (2)” currently pointed to. Therefore, the younger brother node “connective (18)” is pointed, and the translated word “and” of the conjunction “and” is determined and set in the translated word buffer 3e as shown in FIG. Further, the node “main subject lacking main clause (19)” which is the next brother node relating to the node “main clause (2)” is pointed.
... Steps S29, S31 to S33
[0050]
Hereinafter, the node in the structural analysis tree before the conversion corresponding to the node “subject missing main clause (19)” is the node “subject missing main clause (23)”, and the node “auxiliary verb” does not exist below the node. Therefore, the source word string belonging to the node “subject missing main clause (19)”
"It unlikely make"
From this, the translated word string of the node “subject missing main clause (19)” is determined. However, in this case, since the node “auxiliary verb” does not exist in the lower node, the content of the auxiliary verb buffer 3k is not updated. Therefore, the end of the section is transformed into the final form “~” according to the auxiliary verb code “9” stored in the auxiliary verb buffer 3k, similarly to the end of the “main section (2)”. As a result, the translation for each source word belonging to the node “subject missing main clause (19)” in the source sentence is
"That"
"I don't think so"
"I will do"
"."
Is determined. In addition, translations that do not correspond to the original words
"O"
"To state"
Is generated and set in the translated word buffer 3e as shown in FIG.
Here, since there is no younger brother node in the node “subject missing main clause (19)”, the target language generation subroutine is terminated.
... Steps S25, S27 to S30
[0051]
As described above, the translated word string obtained for each section is set in the translated word buffer 3e. At this time, word position information on the translated sentence is added to each translated word. . In this case, the word position information is word position information not in units of phrases but in units of one whole original. The reason for this is to make it possible to share many of the configurations with a conventional machine translation apparatus that outputs a translation sentence by sentence.
At the same time, source / translation correspondence information indicating the correspondence between the position of each source word on the source sentence and the position of the corresponding translation word on the target sentence is set in the source / translation word correspondence information buffer 3f as shown in FIG. The The original word position information necessary at this time is stored in the original sentence buffer 3a in the format shown in FIG.
[0052]
Next, the phrase division rule shown in FIG. 8 is applied to the English structural analysis tree as shown in FIG. 9 stored in the syntax buffer 3c before conversion to determine the phrase division position. Is done.
First, to the node “main node” of node number “2”,
・ Phrase division rule 2
Separate the parts that are grouped at the nodes of the "main section"
Applies. Here, the part grouped by node “main node” with node number “2”
"It would push them to the end of the creditors'line"
Therefore, the original sentence is divided between the word “line” and the word “and”.
Next, to the node “subject missing main clause” of node number “23”,
・ Phrase division rule 2
Separate the parts that are grouped together at the nodes of the “main section lacking case elements”
Applies. Here, the part grouped by the node “subject missing main node” of node number “23” is
"Make it nulikely"
Therefore, it is divided between the word “and” and the word “make”.
[0053]
As a result of the application of the phrase division rule as described above, the original text is finally divided between the word “line” and the word “and” and between the word “and” and the word “the”. In place
"It would push them to the end of the creators'line"
"And"
"Make it unlikely"
It is divided into three phrases.
The word positions “1”, “12”, and “13” of the first words “It”, “and”, and “make” of each phrase divided in this way are referred to the original buffer 3a in FIG. can get. Each obtained word position is stored in the divided position buffer 3g of the main memory 3 as shown in FIG.
... Step S6
[0054]
If it does so, the word position of the translation word which comprises each obtained phrase will be calculated | required as follows.
For example, the first phrase
"It would push them to the end of the creditors'line"
In the case of FIG. 16, the start word position “1” of the first phrase in the original sentence is obtained by referring to the “word position” column of “serial number 1” in the division position buffer 3g as shown in FIG. . Further, by referring to the “word position” column of “serial number 2”, the start word position “12” of the second phrase and the end word position “11” of the first phrase are obtained. From the above, the first phrase of the original text is composed of each original word whose position is “1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11”. You can see that
Next, referring to the original word / translation word correspondence information buffer 3f as shown in FIG. 15, the translated word position "1, 12, 4, 11, 10" of the translated word corresponding to the original word constituting the first phrase. , 7, 6, 8 ". Furthermore, when the items are rearranged in ascending order, except for overlapping ones, the translated word position “1, 4, 6, 7, 8, 10, 11, 12” of the translated word constituting the translation of the first phrase in the original sentence. Is obtained. The translation word position of the translation constituting the translation of the first phrase obtained in this way is stored in the “translation word position” column of phrase number 1 in the translation word position buffer 3h for each phrase.
[0055]
Similarly, the translation word position “14” of the translation constituting the translation of the second phrase of the original text and the translation word position “16, 18, 20” of the translation constituting the translation of the third phrase are obtained. Are stored in the “translated word position” column of the corresponding phrase number.
... Step S7
[0056]
Next, among the above-mentioned translated words, translated words whose translated word positions are not stored in the translated word position buffer 3h for each phrase, that is, translated words “ha”, “ha”, “ha” The phrase to which “in state” belongs is determined and the respective translated word positions “2”, “5”, “17” and “19” are obtained, and the “translated word” of the corresponding phrase in the translated word position buffer 3h for each phrase. It is stored in the “word position” column.
... Step S8
As a result, the translated word position for each phrase is as shown in FIG.
1st phrase
1,2,4,5,6,7,8,9,10,11,12
Second phrase
14
3rd phrase
16, 17, 18, 19, 20
It becomes.
[0057]
Since there is no translated word presenting the translated word position that needs to be corrected in the translated word for each phrase thus obtained, a translation for each phrase is constructed as follows.
First, the translation word position of the translation constituting the first phrase is determined by the column “translation word position” of phrase number 1 in the translation word position buffer 3h for each phrase shown in FIG.
1,2,4,5,6,7,8,9,10,11,12
And obtained. Next, from the translated word buffer 3e shown in FIG. 13, the translated word “it” at the translated word position 1 is obtained, and the translated word “ha” at the translated word position 2 is obtained. In the following, the translated words “them”, “o”, “creditor”, “no”, “line”, “no”, “end”, “ni” and “will be put” are obtained. And the character string obtained by connecting each translated word thus obtained in ascending order of translated word position
"It will put them at the end of the creditor's line"
Is stored in the "translation phrase" column of phrase number 1 in the translation phrase buffer 3i, as shown in FIG.
Similarly, the translated word “and” that makes up the second phrase and the translated words “it”, “to”, “probably”, “to state” and “do” that make up the third phrase The character string obtained by connecting the obtained translated words in ascending order of the translated word positions is stored in the “translated phrase” column of the corresponding phrase number in the translated phrase buffer 3i. As a result, as shown in FIG. 18, a translation for each phrase is constructed.
... Steps S9, S10
[0058]
When the translated phrases for each phrase of the original sentence are obtained as described above, the output format shaping unit 2B aligns the positions of the original phrase and the translated phrase head as shown in FIG. Is displayed.
... Steps S11 and S12
[0059]
Thus, in the present embodiment, the target language generation unit 2d of the translation module 2A is provided with the auxiliary verb sentence translation method determining means 13, the auxiliary verb detection means 14, the parallel clause detection means 15, and the auxiliary verb translation generation means 18. Further, the main memory 3 is provided with an auxiliary verb buffer 3k for storing auxiliary verb codes of auxiliary verbs detected for each clause by the auxiliary verb detecting means 14. Then, when the original text in which the parallel clause including the auxiliary verb exists is input to the input unit 1, the target language generation unit 2d sets the end of the translation of each clause constituting the parallel clause to the final form corresponding to the auxiliary verb. A modified target language is generated.
Therefore, the original text with parallel clauses containing auxiliary verbs
"It would push them to the end of the creditors' line
and make it unlikely. ''
The phrase translation result for
It would push them to the end of the creditors' line
It will put them at the end of the creditor line
and
And
make it unlikely.
Will make it unlikely
The effect of the auxiliary verb “would” will appear as “put in” for the translation of the first phrase, and the effect of the auxiliary verb “would” does not enter as the translation of the first phrase. Nature is eliminated.
[0060]
Next, the source text that has a structure that separates one section from another
"Keating is the subject of a class-action lawsuit alleging
they misled investors about the safety of junk bonds. ''
However, the phrase unit translation processing operation shown in FIG. 4 and FIG.
[0061]
Similar to the above-described original sentence, dictionary lookup / morpheme analysis processing, syntax analysis processing, and syntax conversion processing are performed. As a result, a structural analysis tree in English is obtained as shown in FIG.
... Steps S1-S4
Next, since there are no parallel clauses in the original sentence, the translated words are arranged in the order of the original words in the converted Japanese structure analysis tree, and the translated words of the original sentences are determined. The determined translated word is stored in the translated word buffer 3e as shown in FIG. Further, the original / translated word correspondence information buffer 3f stores the original / translated word correspondence information as shown in FIG.
... Step S5
[0062]
Next, for the node “main clause (16)” in the structure analysis tree before conversion as shown in FIG.
・ Phrase division rule 2
Separate the parts that are grouped at the nodes of the "main section"
Applies. Here, the part grouped by node “main node” with node number “16”
"They misled investors about the safety of junk bonds"
Therefore, the original text is divided between the word “alleging” and the word “they”.
Then, the word positions “1” and “10” in the original sentence of the first word of each phrase divided in this way are shown in FIG. 23 together with the number of the phrase division rule applied at the time of phrase division determination. It is stored in the division position buffer 3g.
... Step S6
[0063]
Next, referring to the column of “word position” in the division position buffer 3g as shown in FIG. 23 and the original word / translation word correspondence information buffer 3f as shown in FIG. 22, the translated word constituting each phrase in the original sentence is referred to. The translated word position is obtained. Further, the translated word position of the translated word whose translated word position is not stored in the translated word position buffer 3h for each phrase is determined. Thus, the translated word position for each phrase is determined as shown in FIG. 24 and stored in the translated word position buffer 3h for each phrase.
... Steps S7, S8
[0064]
Next, according to the correction rule memorize | stored in the other translation rule memory 4e of the memory | storage part 4 by the missing phrase detection means 21, the production | generation position correction means 23, and the auxiliary symbol production | generation means 24 in the said translation phrase construction part 2f, A determination as to whether correction is required for the translated word position and a correction operation are performed as follows.
The above amendment rule at that time is
(I) When all of the following determination rules (1), (2), and (3) are satisfied, the symbol “˜” and the translated word at the end of phrase b are generated at a discontinuous portion of phrase a.
(1) There are discontinuous parts in the translated word positions of the translated words that form a phrase a, and all the translated word positions of the translated words that form the other phrases b are one in the phrase a. Enter a discontinuous spot.
(2) The phrase division rule applied when dividing the phrase b is rule number 2.
(3) There is no original word corresponding to the translated word at the end of phrase b.
It is.
[0065]
First, in the first phrase stored in the translated word correspondence information buffer 3h for each phrase shown in FIG. 24 by the above-described missing phrase detecting means 21, each translated word position is a translated word position after sorting. Since the translated word position “15” is located next to the translated word position “2”, it is detected that this location is a discontinuous location. In addition, it is detected that all the translation words (translation word positions 4 to 14) belonging to the second phrase enter this discontinuous portion. Therefore, judgment rule (1)
“There is a discontinuous part in the translated word position of the translated word that forms a phrase a, and all the translated word positions of the translated word that forms another phrase b are one discrete in the phrase a. Enter in place "
Is determined to be satisfied.
Further, the content of the division position buffer 3g shown in FIG. 23 is referred to, and the application rule number at the time of phrase division relating to the second phrase is “2”. Therefore, decision rule (2)
“The phrase division rule applied when dividing phrase b is rule number 2.”
Is determined to be satisfied.
Also, the translated word position of the translated word at the end of the second phrase is “14”, but referring to the original / translated word correspondence information buffer 3 f shown in FIG. 22, the translated word position corresponds to the translated word at the translated word position 14. It can be seen that there are no source language words. Therefore, decision rule (3)
“There is no source word corresponding to the translated word at the end of phrase b”
Is determined to be satisfied.
That is, since the original text satisfies all the determination rules (1), (2), and (3), it is determined that the content of the translated word position buffer 3h for each phrase needs to be corrected.
[0066]
Then, according to the correction rule (I), the translated word “to” at the translated word position 14 located at the end of the second phrase is a translated word that is a discontinuous part in the first phrase. It is moved before the translated word “assert” at position 15. That is, in the present embodiment, the attached word generating means in claim 4 is realized by the attached word generating means 17 of the target language generating section 2d and the generation position correcting means 23 of the translated phrase constructing section 2f. Furthermore, the symbol “˜” is generated before the translated word “to” at the discontinuous portion of the first phrase by the auxiliary symbol generating means 24 of the translated phrase constructing unit 2 f.
As a result, the contents of the translated word position buffer 3h for each phrase change as shown in FIG. Here, the translated word position “−100” means the translated word position of the translated word “˜”. Further, when the value of the translated word position is a negative number, the translated word stored in the auxiliary translated word table memory 4d of the storage unit 4 is used instead of using the translated word stored in the translated word buffer 3e. It is meant to be used. In the case of this example, referring to the auxiliary translation word table in the auxiliary translation word table memory 4d as shown in FIG. 26, the translation (symbol) “˜” whose code is “−100” is obtained.
... Step S9
[0067]
Then, based on the contents of the translated word position buffer 3h for each phrase, the translated word buffer 3e and the auxiliary translated word table are referred to, and a translation for each phrase is constructed. As shown in FIG. Stored. Then, the head positions of the original sentence phrase and the translated sentence phrase are aligned and displayed as shown in FIG.
... Steps S10-S12
[0068]
Thus, in the present embodiment, the translation word position correction rule (I) is stored in the other translation rule memory 4e of the storage unit 4. Then, the blank phrase detection means 21 in the translated phrase construction unit 2f satisfies all of the determination rules (1), (2) and (3), so that the contents of the translated word position buffer 3h for each phrase are corrected. If it is determined that it is necessary, the translated word that is the translated word located at the end of the second phrase according to the modified rule (I) by the generated position correcting means 23 and the auxiliary symbol generating means 24 of the translated phrase building unit 2f. The symbol “˜” is added in front of the translated word “to” at the position 14, and the translated word “insist” at the translated word position 15, which is a discontinuous portion in the first phrase, is moved.
[0069]
Therefore, according to the present embodiment, the original text having a structure in which one section is separated from another section so that it is completely removed.
"Keating is the subject of a class-action lawsuit alleging
they misled investors about the safety of junk bonds. ''
The phrase translation result of
Keating is the subject of a class-action lawsuit alleging
Keating is the subject of a class action claiming that
they misled investors about the safety of junk bonds.
They misled investors about the safety of junk bonds
It is easy to understand that it has a sentence structure that is separated from the place of “~” in the first phrase so that the second phrase that is the case element of the sentence is completely removed, and Each translated phrase is easy to read.
[0070]
Next, the original with a relative clause that would cause the case element to be missing when separated
"The plan was opposed by some of top executives
who invested in the company's securities. ''
However, the phrase unit translation processing operation shown in FIG. 4 and FIG.
[0071]
In the same manner as described above, dictionary lookup / morpheme analysis processing, syntax analysis processing, and syntax conversion processing are performed. As a result, a structural analysis tree in English is obtained as shown in FIG.
... Steps S1-S4
Next, since there are no parallel clauses in the original sentence, the translated words are arranged in the order of the original words in the converted Japanese structure analysis tree, and the translated words of the original sentences are determined. Then, the determined translation is stored in the translation buffer 3e as shown in FIG. Further, the original / translated word correspondence information buffer 3f stores the original / translated word correspondence information as shown in FIG.
... Step S5
[0072]
Next, for the node “relative pronoun (15)” and the node “subject missing main clause (16)” in the structure analysis tree before conversion shown in FIG.
・ Phrase division rule 5
If there is a contact point of “relative” right next to the node of “main clause lacking case element”, put together the contact point of “relative” and the contact point of “main clause lacking case element” Separate as a single phrase
Applies. Here, the part where the node “relative pronoun” of the node number “15” and the node “subject missing main node” of the node number “16” are combined is
"Who invested in the company'securities"
Therefore, the original text is divided between the word “executives” and the word “who”.
Then, the word positions “1” and “10” in the original sentence of the first word of each phrase divided in this way are shown in FIG. 32 together with the number of the phrase division rule applied at the time of phrase division determination. As shown, it is stored in the division position buffer 3g.
... Step S6
[0073]
Next, referring to the “word position” column in the division position buffer 3g as shown in FIG. 32 and the original word / translation word correspondence information buffer 3f as shown in FIG. 31, the translated words constituting each phrase in the original sentence are referred to. The translated word position is obtained. Further, the translated word position of the translated word whose translated word position is not stored in the translated word position buffer 3h for each phrase is determined. Thus, the translated word position for each phrase is determined as shown in FIG. 33 and stored in the translated word position buffer 3h for each phrase.
... Steps S7, S8
[0074]
Next, other translations in the storage unit 4 are performed by the relative clause detection unit 22, the generation position correction unit 23, the auxiliary modifier generation unit 25, the antecedent acquisition unit 26, and the postfix generation unit 27 in the translated phrase construction unit 2f. In accordance with the correction rules stored in the rule memory 4e, whether or not each translated word position needs to be corrected is determined and corrected as follows.
The above amendment rule at that time is
(II) If all the following judgment rules (4) and (5) are satisfied, the word corresponding to the antecedent of the relative in phrase a is obtained from the immediately preceding phrase, and an appropriate modifier for that word is obtained. At the same time, it is generated at the missing position of phrase a.
(4) The rule division rule applied when dividing a certain phrase a is rule number 5.
(5) There is a “main clause” in which the case element is missing in phrase a.
It is.
[0075]
First, the content of the division position buffer 3g shown in FIG. 32 is referred to by the relative clause detection means 22, and the application rule number at the time of phrase division relating to the second phrase is “5”. )
“The rule division rule applied when dividing a phrase a is rule number 5.”
Is determined to be satisfied.
In addition, referring to the structural analysis tree before conversion shown in FIG. 29, it can be seen that the second phrase has a node “subject missing main clause (16)” which is a main clause with a missing case element. Therefore, judgment rule (5)
“There is a“ main clause ”with missing case elements in phrase a”
Is determined to be satisfied.
That is, since the original text satisfies all the determination rules (4) and (5), it is determined that the content of the translated word position buffer 3h for each phrase needs to be corrected.
[0076]
Then, the antecedent acquisition means 26 searches for the antecedent of the relative pronoun “who” in the second phrase in the first phrase according to the correction rule (II), and the noun “executives” is the target. Is determined to be an antecedent of Further, since the noun “executives” is the word whose original word position is “9” in the original text, the antecedent is referred to by referring to the original word / translation word correspondence information buffer 3f shown in FIG. 31 and the translated word buffer 3e shown in FIG. As a translated word corresponding to the word “executives”, “executive” whose translated word position is “11” is obtained. Then, by the generation position correcting means 23 of the translated phrase building unit 2f, the translated word “stem” of the translated word number 11 among the translated words belonging to the phrase number 2 in the translated word position buffer 3h for each phrase shown in FIG. Is added.
[0077]
At that time, it is necessary to determine the generation position of the translated word “executive” in the second phrase. This is based on the case grammatical viewpoint, as in the case of normal translated word generation. Determined by 2d.
Here, in the present embodiment, it is assumed that the generation position of the case element of the sentence is described in the verb dictionary. And in the case of the original text, the central verb of the clause forming the second phrase is “invested”, and the syntax of the clause is
Subject + verb + preposition verb
Is a deformed form (the subject is missing). This has already been determined by the syntax analysis unit 2b as shown in FIG. The translation of the word “invest” is also determined based on the result of this parsing process.
[0078]
FIG. 34 is a conceptual diagram of data relating to the translated word “invest” selected in the original sentence in the original word “invest” stored in the dictionary data memory 4 a in the storage unit 4. “First case” in FIG. 34 is the “subject” of the case in the original sentence, and “second case” in FIG. 34 is the “preposition phrase” in the case in the original sentence. It is assumed that the standard generation position of the first case at the time of Japanese generation is designated as “first”.
Then, since the translated word “stem” acquired by the antecedent acquiring unit 26 is a “subject” with respect to the original word “invested”, the generation position correcting unit 23 converts the translated word “stem” into the second phrase. It is generated as a translation of the relative pronoun “who” at the first position.
That is, in the present embodiment, the relative term translation generation unit in claim 1 is realized by the generation position correction means 23.
[0079]
Japanese case elements include postfix words such as case particles. Information for generating the postfix is also stored in the dictionary data memory 4a. Therefore, the postfix word generating means 27 obtains the first case name “acting subject” of the verb “invest” from the dictionary data memory 4a shown in FIG. 34, and the postfix word of the storage unit 4 as shown in FIG. Referring to the postfix word generation information stored in the table memory 4c, the case name is “acting subject” and the clause including the case (the clause forming the second phrase) is the main clause. “Ha” is generated as a postscript of the translated word “executive”.
Further, the modifier of the translated word “senior” is determined by the auxiliary modifier generation means 25 of the translated phrase construction unit 2 f as follows.
That is, the translated word “executive” is a noun because it is a noun. Therefore, by referring to the information stored in the auxiliary translation table memory 4d of the storage unit 4 as shown in FIG. 26, “that” defined as the modifier of the body is selected, and the generation position is the translation word Set immediately before “Executive”.
[0080]
As a result, the contents of the translated word position buffer 3h for each phrase are as shown in FIG. If the value of the translated word position is a negative number, the translated word stored in the translated word buffer 3e is not used, but the postfix table memory 4c (FIG. 35) and the auxiliary translated word table stored in the storage unit 4. Referring to the memory 4d (FIG. 26), the code “−101” generates the translated word “that”, and the code “−1” means that the translated word “ha” should be generated.
... Step S9
[0081]
Based on the contents of the translated word position buffer 3h for each phrase, the translated word buffer 3e (FIG. 30), the postfix word table memory 4c (FIG. 35), and the auxiliary translated word table memory 4d (FIG. 26) are referred to. A translation for each phrase is constructed and stored in the translated phrase buffer 3i as shown in FIG. Then, the head positions of the original sentence phrase and the translated sentence phrase are aligned and displayed as shown in FIG.
... Steps S10-S12
[0082]
Thus, in the present embodiment, the translation word position correction rule (II) is stored in the other translation rule memory 4e of the storage unit 4. Then, the related clause detection means 22 in the translated phrase construction unit 2f satisfies all of the determination rules (4) and (5), so that it is determined that the content of the translated word position buffer 3h for each phrase needs to be corrected. Then, the antecedent acquisition means 26 of the translated phrase construction unit 2f follows the correction rule (II) and the word “executives” that is the antecedent of the relative pronoun “who” in the second phrase and the translated word “ "Executive" get. Then, a translation word “stem” is generated among the translation words belonging to phrase number 2 by the generation position correction means 23 of the translated phrase construction section 2 f, and further, the auxiliary modifier generation means 25 and the postfix word generation means 27 The modifier “that” and the postfix “ha” are generated and added to the translated word “executive” which is a relative translation.
[0083]
Therefore, according to the present embodiment, the original text having a relative clause that causes the case element to be lost when separated.
"The plan was opposed by some of top executives
who invested in the company's securities. ''
The phrase translation result of
The plan was opposed by some of top executives
The plan was opposed by some top executives
who invested in the company's securities.
The executive invested in company securities
The case element “is its executive” missing from the second phrase is supplemented so that the second phrase, which is a related clause, can be read easily.
[0084]
Note that the algorithm of the phrase unit translation processing operation in the present embodiment is not limited to the flowcharts shown in FIGS.
[0085]
【The invention's effect】
As is clear from the above, the machine translation device of the invention according to claim 1 and the machine translation method of the invention according to claim 5 detect the related clauses that are separated and lack each element by the related clause detection means, A predecessor of a relative clause in a relative clause is acquired by an antecedent acquisition unit, a translation of the antecedent is obtained by a relative term translation generation unit, and a plurality of words constituting the relative clause stored in the nodal translation storage unit Is generated at the position of the missing case element in the translated word, as a translated word of the relative word, so that the relative word is separated in the translated word string of the relative clause that is separated at the position determined by the division position determining means and missing the case element. It is possible to obtain a translation result in which the translation of the antecedent is inserted as a translation.
[0086]
Therefore, the original text with a relative clause that would lose the case element when separated
"The plan was opposed by some of top executives
who invested in the company's securities. ''
The phrase translation result for
"The plan was opposed by some of top executives
The plan was opposed by some top executives
who invested in the company's securities.
The executive invested in company securities. ''
It is obtained as follows.
That is, according to the present invention, the missing case element in the relative clause is compensated, and a translation result that is very easy to understand even when read alone is obtained.
[0087]
Moreover, since the machine translation apparatus of the invention according to claim 3 has a postfix generating unit that generates a postfix immediately after the translation of the relative term generated by the relative term translation generating unit, You can get translation results of relative clauses that are very easy to understand by reading alone.
[0088]
In addition, the invention according to claim 4 Mysterious Since the machine translation device has auxiliary modifier generation means for generating a modifier relating to the translation of the relative phrase generated by the relative translation generation means, the relative clause in which the translation of the relative phrase is generated Can be further improved.
[Brief description of the drawings]
FIG. 1 is a block diagram of a machine translation apparatus according to the present invention.
FIG. 2 is a detailed configuration diagram of a target language generation unit in FIG. 1;
FIG. 3 is a detailed configuration diagram of a translated phrase construction unit in FIG. 1;
4 is a flowchart of phrase unit translation processing operation executed under the control of the control unit in FIG. 1;
FIG. 5 is a flowchart of a target language generation subroutine executed in the phrase unit translation processing operation shown in FIG. 4;
FIG. 6 is an explanatory diagram of an analysis level by machine translation.
FIG. 7 is a conceptual diagram showing an example of the contents of an original text buffer in FIG.
FIG. 8 is a diagram illustrating an example of a phrase division rule.
FIG. 9 is a conceptual diagram illustrating an example of a parse tree before conversion.
FIG. 10 is a conceptual diagram illustrating an example of a parse tree after conversion.
11 is a conceptual diagram showing an example of the contents of an auxiliary verb code table memory in FIG. 1. FIG.
12 is a conceptual diagram showing an example of the contents of an auxiliary verb buffer in FIG. 1. FIG.
13 is a conceptual diagram showing an example of the contents of a translated word buffer in FIG. 1. FIG.
FIG. 14 is a conceptual diagram showing the contents of a translated word buffer corresponding to FIG. 13 when a phrase translation is not commanded.
15 is a conceptual diagram showing an example of the contents of an original word / translation word correspondence information buffer in FIG. 1; FIG.
16 is a conceptual diagram illustrating an example of contents of a division position buffer in FIG. 1. FIG.
FIG. 17 is a conceptual diagram illustrating an example of the contents of a translated word position buffer for each phrase in FIG. 1;
18 is a conceptual diagram showing an example of the contents of a translated phrase buffer in FIG. 1. FIG.
FIG. 19 is a diagram illustrating a display example of a phrase translation result.
FIG. 20 is a conceptual diagram of a parse tree before conversion different from FIG. 9;
FIG. 21 is a conceptual diagram showing the contents of a translated word buffer different from FIG. 13;
FIG. 22 is a conceptual diagram showing the contents of an original / translation word correspondence information buffer different from FIG. 15;
FIG. 23 is a conceptual diagram showing the contents of a division position buffer different from FIG.
FIG. 24 is a conceptual diagram showing the contents of a translated word position buffer for each phrase different from FIG. 17;
25 is a conceptual diagram showing the contents of a translated word position buffer for each phrase after correction is made to FIG. 24. FIG.
FIG. 26 is a conceptual diagram showing an example of the contents of an auxiliary translation word table memory in FIG. 1;
FIG. 27 is a conceptual diagram showing the contents of a translation phrase buffer different from FIG.
FIG. 28 is a diagram showing a display example of phrase translation results different from FIG.
29 is a conceptual diagram of a parse tree before conversion different from FIGS. 9 and 20. FIG.
30 is a conceptual diagram showing the contents of a translated word buffer different from those in FIGS. 13 and 21. FIG.
31 is a conceptual diagram showing the contents of an original / translation word correspondence information buffer different from those in FIGS. 15 and 22. FIG.
32 is a conceptual diagram showing the contents of a division position buffer different from those shown in FIGS. 16 and 23. FIG.
33 is a conceptual diagram showing the contents of a translated word position buffer for each phrase different from those in FIGS. 17 and 24. FIG.
34 is a conceptual diagram showing a part of the dictionary memory in FIG. 1;
FIG. 35 is a conceptual diagram showing an example of the contents of a postfix word table memory in FIG. 1;
36 is a conceptual diagram showing the contents of a translated word position buffer for each phrase different from those in FIGS. 17 and 24. FIG.
FIG. 37 is a conceptual diagram showing the contents of a translation phrase buffer different from those shown in FIGS. 18 and 27.
38 is a diagram showing a display example of phrase translation results different from those in FIGS. 19 and 28. FIG.
[Explanation of symbols]
1 ... input unit, 2 ... translation control unit,
2A ... Translation module, 2a ... Dictionary lookup / morpheme analysis unit,
2b ... syntax analysis part, 2c ... syntax conversion part,
2d: target language generation unit, 2e: division position determination unit,
2f ... Translated phrase construction part, 3 ... Main memory,
3c: syntax buffer before conversion, 3d: syntax buffer after conversion,
3e ... translated word buffer, 3f ... original / translated word correspondence information buffer,
3g: Split position buffer,
3h ... translated word position buffer for each phrase,
3i ... translation phrase buffer, 3k ... auxiliary verb buffer,
4 ... storage unit, 4a ... dictionary memory,
4c: Postfix table memory, 4d: Auxiliary translation table memory,
4g ... auxiliary verb code table memory, 5 ... output section,
6 ... control unit, 11 ... tree structure tracing means,
13 ... Auxiliary verb translation method determining means, 14 ... Auxiliary verb detecting means,
15 ... Parallel clause detection means, 18 ... Auxiliary verb translation generation means,
21 ... A missing phrase detecting means, 22 ... A related clause detecting means,
23 ... Generation position correcting means, 24 ... Auxiliary symbol generating means,
25 ... auxiliary modifier generation means, 26 ... antecedent acquisition means,
27: Postfix word generation means.

Claims (5)

Language generation means for generating a translated word string of the input source text composed of source language words;
Division position determining means for determining a position at which the input original is divided into phrases or clauses;
A translated word storage means for acquiring the translation of each word constituting the clause divided at the position determined by the divided position determining means from the generation result by the language generating means, and storing the translated word for each clause and in units of words;
Related clause detection means for detecting a clause that has been divided at the position determined by the division position determination means and is missing a case element according to a division rule ;
An antecedent acquisition unit for acquiring an antecedent of a relative in a relative clause detected by the relative clause detection unit;
The translation of the antecedent obtained by the antecedent acquisition means is acquired from the noun translation storage means, and the missing words in the translations of a plurality of words constituting the relevant clause stored in the nodal translation storage means A machine translation device comprising: a relative translation generation means for generating a translation of the relative verb at the position of the case element. The machine translation device according to claim 1,
A translation of the relative verb generated in the translation of a plurality of words constituting the relative clause in the relative translation storage unit by the relative translation generation unit, and the relative verb stored in the clause translation storage unit A machine translation device characterized in that the translated word of the antecedent in a clause including the antecedent is the same. In the machine translation device according to claim 1 or 2,
A machine translation apparatus comprising postfix word generating means for generating a postfix representing the case of the relative word translation immediately after the relative word translation generated by the relative translation generation means . The machine translation device according to any one of claims 1 to 3,
The apparatus further comprises auxiliary modifier generation means for generating a modifier related to the translation of the relative verb generated by the relative translation generation means and improving the readability of the clause in which the translation of the relative is generated. A machine translation device. A language generation step of generating a translated word string of an input original composed of source language words by a language generation means;
A division position determination step for determining a position at which the input original text is divided into phrases or clauses by a division position determination means;
The translation of each word constituting the clause divided at the position determined by the division position determination means is obtained from the generation result by the language generation means, and is stored in the phrase translation storage means for each clause and in units of words. A translation storage step;
A relative clause detection step of detecting a relative clause, which is a clause divided by the division position determination means and lacking a case element , according to a division rule, by a relative clause detection means;
An antecedent acquisition means for acquiring an antecedent of a relative in the relative clause detected by the relative clause detection means by an antecedent acquisition means;
A plurality of words constituting the related clause stored in the noun translation word storage means obtained by the noun translation generation means, obtained from the noun translation word storage means, obtained by the antecedent acquisition means. A machine translation method, comprising: a relative translation generation step for generating a translated word of the relative verb at the position of the missing case element in the translated word.

Images