JP6952967B2 - Automatic translator - Google Patents

Description

The present invention relates to an automatic translation technique, and in particular, a technique for handling a long sentence that is difficult to translate while following certain rules, such as claims in the claims of a patent application (hereinafter referred to as "claims"). Regarding.

It may be necessary to know the meaning of a claim written in a foreign language in order to confirm the technical content of a patent application written in a foreign language or to gain knowledge about the scope of rights. If you have a small number of subjects and you have knowledge of the language of the document in question, you may not have much trouble reading the original text. However, when the target document is large or the knowledge about the language of the document is poor, it is impractical to read all the claims of the target document. In such cases, it is common to try to understand the content by translating each claim into one's own language by automatic translation.

However, it is well known that the quality of complaints automatically translated into other languages is low. It is customary in the claims to state many limitations of the invention in one sentence. Therefore, the complaint text tends to be long. Moreover, in a technical field where many inventors are competing with development, it may not be possible to clarify the difference from the prior art unless a relatively large number of restrictions (constituent requirements) are added to the claim. As a result, claims are often sentences of length that are unlikely to occur in ordinary documents. Although automatic translation technology has been greatly developed and improved in accuracy in recent years, highly accurate translation cannot be expected yet, especially when the original text of the translation is long, such as a complaint.

A proposal for solving such a problem is made in Patent Document 1 described later. The technique described in Patent Document 1 is basically based on the idea that a long claim is divided into its components (hereinafter referred to as "structural parts") and translated for each structural part. .. This takes advantage of the fact that claims are generally divided into multiple structural parts, especially these days, and that the structural parts are separated by a specific delimiter pattern. ing. As the structural component of the claim, a transition clause inserted between the subject of the invention and one or more constituents and the subject of the invention and the constituents can be considered. The order of these descriptions varies depending on the language. For example, in the case of Japanese, one or more constituent requirements are located at the beginning, followed by a transitional phrase, followed by the transitional phrase, and the subject matter of the invention is described at the end. In the case of English, the opposite is true, with the subject of the invention being listed first, followed by a transitional phrase, followed by one or more constituents.

In the case of Japanese, as transitional phrases, "include (characterized by)", "provide (characterized by)", "consisting of (characterized by)", etc. are mainly used. In the case of English, "comprising:", "including:", "consisting of:", etc. are used as transition phrases.

Between one or more components, "," + line feed is mainly used in Japanese, and ";" is used in English.

By detecting such a delimiter pattern in the claim, the structure of the claim can be determined.

FIG. 1 shows a rough processing flow of the technique disclosed in Patent Document 1. With reference to FIG. 1, the claim translation procedure 30 in Patent Document 1 uses a computer, and has a step 40 of morphological analysis and parsing of the claim text to generate syntactic information, and a syntax. Step 42, which detects the delimiter pattern of structural parts in the analysis result and divides the text of the claim into each component, and for each structural part, the character string indicating the structural part name is specified in the text, and the character is specified. The column includes step 44 with a tag indicating that it is a structural part name. As a result of parsing, a tree-like syntactic structure is formed between structural parts. In Patent Document 1, this syntactic structure is expressed in the form of tags attached to structural parts and words inside them.

In the first embodiment of Patent Document 1, it is assumed that the subject claim is in English. In the case of English, the word corresponding to the name of the structural part is often written at the beginning of the text about each structural part, and most of the forms include the explanation about the structural part after that. Therefore, the process of step 44 can be easily performed. In addition, the explanation of structural parts is often patterned. In Patent Document 1, the explanatory text in which the description is patterned in this way is referred to as an explanatory pattern sentence.

The translation procedure 30 of Patent Document 1 further specifies an explanatory pattern sentence included in the structural component to be processed, divides the structural component to be processed into each pattern sentence, and the divided one is an explanatory pattern sentence. Step 46 to add a tag indicating that there is, step 48 to translate the structural part name of the structural part to be processed into the target language by automatic translation, and change the sentence structure to one suitable for translation for each explanation pattern sentence. Step 50, the step of translating each explanatory text whose sentence structure has been changed in step 50 into the target language 52 by automatic translation, and the structural part name and each pattern sentence translated into the target language in this way are used as tags. Therefore, it includes step 54 for displaying the tree.

The change in the sentence structure in step 50 is to change the description so that the description includes the subject and the predicate. For example, the English relative pronoun is omitted and a structural part name is inserted instead, or a noun or noun phrase modified by the explanation pattern sentence is inserted. In Patent Document 1, it is stated that a conversion pattern is set in advance for sentence conversion including this. This sentence structure conversion seems to be for matching the explanation pattern sentences in the word order of general sentences.

With reference to FIG. 2, in the automatic translation method according to Patent Document 1, more specifically, the following processing is performed. English claim 70 shall be the subject of translation. In Patent Document 1, an English claim is decomposed into a plurality of structural parts by finding an English delimiter pattern, and a tree structure is formed between them (process 72). As described above, by paying attention to the delimiter pattern in the English claim, it is possible to determine the structure of the English claim. According to the result, the English claim can be decomposed into multiple structural parts.

In the procedure according to Patent Document 1, as a result of the process 72, an English pattern 74 having a tree structure is formed. In the example shown in FIG. 2, the English pattern 74 includes subject 90, transitional phrase 92, and description 94. By automatically translating the subject 90, the transition phrase 92, and the explanation 94 separately (processing 76), a Japanese pattern 78 having a tree structure is generated. The Japanese pattern 78 has the same structure as the English pattern 74, and the text is converted into Japanese. That is, the Japanese pattern 78 includes a Japanese translation 100 of the subject, a Japanese translation 102 of the transition phrase, and a Japanese translation 104 of the explanation. In Patent Document 1, instead of generating one sentence of Japanese claim from Japanese pattern 78, Japanese pattern 78 is displayed as it is in a tree format.

According to Patent Document 1, since it is sufficient to automatically translate individual structural parts by performing such processing, it is said that the translation accuracy of the constituent explanatory pattern sentences is improved as compared with the conventional one. Furthermore, it is said that the content of the complaint will be easier to understand by displaying the translated sentence in a tree structure.

JP-A-2014-199476 (particularly FIGS. 5, 7, 8 and paragraphs 31, 41, 51-77).

Certainly, it seems that the method of Patent Document 1 makes the translation more accurate than translating the entire claim as a whole. However, when automatic translation is used, the technique of Patent Document 1 still has a problem to be solved.

The biggest problem is the change in the sentence structure that makes up the explanation. The change of the sentence structure is a process required because the explanation is not a normal sentence but a form that modifies each structural part. If the conversion pattern of the sentence structure is prepared in advance and the explanation pattern sentence is completely changed, there may be no problem even if each explanation sentence is translated using the existing automatic translation technique. However, the variations of sentences that can occur as claims are virtually infinite, and it is impossible to prepare sentence structure conversion rules for them in advance. Moreover, there is no clear disclosure in Patent Document 1 as to what kind of conversion pattern should be prepared. Therefore, it is unlikely that the description will be converted into a sentence structure suitable for automatic translation. Even if such conversion is not performed correctly and the explanatory text is not in the form of a sentence, the automatic translation device considers each explanatory text to be a normal sentence and performs translation. As a result, the resulting translation results are unlikely to be comprehensible. In particular, when a structural part is a noun phrase, if the automatic translation device treats it as a sentence and translates it, the adverse effect on the translated sentence becomes large.

For example, with reference to FIG. 3, consider an English claim 120 having two noun phrases 1 and 2 corresponding to explanatory text. According to the Japanese claim 122 obtained by automatic translation by the inventor of the present application in accordance with Patent Document 1, the translation for noun phrase 1 and noun phrase 2 is almost incomprehensible. As shown in FIG. 3, when processing is performed according to Patent Document 1, even though the sentence structure is not converted well, when each noun phrase or the like is translated as a sentence, the obtained translated sentence is obtained. Is difficult to understand and cannot serve the purpose of automatic translation.

Moreover, such a problem arises not only in the translation of the claims of the patent application, but also in the translation of a sentence consisting of a plurality of structural parts, which is also a long sentence and is described according to a specific rule. Is. For example, laws, contracts, and treaties, as well as instructions for using various machines, electronic devices, and software can cause such problems.

Therefore, an object of the present invention is to provide an automatic translation apparatus capable of improving the accuracy of automatic translation for a long sentence that is described according to a specific format and can be divided into a plurality of structural parts.

The automatic translation device according to the first aspect of the present invention translates a sentence in the first language into a sentence in a second language different from the first language. This automatic translation device identifies a sentence pattern of the first language and divides the sentence of the first language into structural parts, and a sentence pattern of the first language specified by the dividing means in advance. A pattern specifying means for specifying the pattern of the associated second language sentence, a pattern of the first language sentence divided by the dividing means, and a pattern of the second language specified by the pattern specifying means. With respect to the correspondence identification means for specifying the correspondence relationship of the structural parts and the grammatical characteristics of each structural part, and the word string constituting the structural part for each of the structural parts of the first language, the first A translation means that performs automatic translation using a model for translation from a language to a second language to generate a translation of the second language, and a translation means for each of the structural components of the first language. Substituting the translation of the second language into one of the structural parts of the pattern of the sentence of the second language specified by the pattern specifying means according to the correspondence of the structural parts specified by the corresponding specifying means. Includes an assignment means for generating a second language sentence, which is a translation of the first language sentence.

The text in the first language may be, for example, any of the claims of a patent application, laws, regulations, treaties, and instructions for use of various machines, electronic devices, and software.

Preferably, the automatic translation device receives the structural parts of the first language in which the correspondence relationship and the grammatical characteristics are specified by the correspondence identification means, which is provided between the correspondence identification means and the translation means, and receives the structural parts of the first language. For each of the structural parts, the word rearranging means for giving the translation means by rearranging the order of the word strings constituting the structural parts according to the order in which the translated words of the word strings appear in the second language. include.

More preferably, the word rearrangement means is a parsing means that performs parsing for each of the structural parts of the first language to generate a parsing tree for the first language, and a parsing means generated by the parsing means. In the conversion means for converting the parse tree of one language into the parse tree of the second language according to the conversion rules prepared in advance, and in the parse tree of the second language obtained by the conversion by the conversion means. It includes a rearrangement means for rearranging and outputting the words of the structural parts of the first language according to the order of appearance of the words.

The translation means is pre-optimized for the translation of the word string having the grammatical characteristics of the structural part with respect to the word string constituting the structural part for each of the structural parts of the first language. It may include a grammatical characteristic-specific translation means that performs automatic translation using a model for translation from language to second language to generate a translation of the second language.

More preferably, the pattern specifying means includes a dividing means for dividing the sentence in the first language into a plurality of structural parts by a predetermined dividing pattern included in the sentence in the first language, and a dividing means for dividing the sentence in the first language. The first language is based on the grammatical characteristic determining means for determining the grammatical characteristics of each structural component, the appearance order of the structural parts divided by the dividing means, and the grammatical characteristics determined by the grammatical characteristic determining means for each structural component. Including means for identifying the sentence pattern of.

The sentence in the first language may be a claim of a patent application written in the first language.

The translation model learning device according to the second aspect of the present invention translates a sentence of a specific grammatical characteristic of the first language into a sentence of a second language different from the first language by statistical translation. Train the translation model used. This device is a collecting means for collecting a plurality of translations consisting of a sentence having a specific grammatical characteristic of the first language and a translation of the sentence in a second language different from the first language, and a collecting means. To train the statistical model necessary for statistical translation from a sentence with a specific grammatical characteristic in the first language to a sentence in the second language, using the multiple translations collected by Including statistical learning means.

It is a flowchart which shows the procedure of the automatic translation processing of the claim of the conventional patent application. It is a figure which shows typically the translation process of the text in the automatic translation processing of the claim of the conventional patent application. It is a schematic diagram for demonstrating the problem in the automatic translation processing of the claim of the conventional patent application. It is a schematic diagram for demonstrating the learning process necessary for the process of automatic translation in 1st Embodiment of this invention. It is a schematic diagram for demonstrating the text translation process in the automatic translation process of a claim which concerns on 1st Embodiment of this invention. It is a block diagram which shows the whole structure of the automatic translation system which concerns on 1st Embodiment of this invention. It is a block diagram which shows the schematic structure of the model learning part for structural component translation which learns various models for automatic translation in 1st Embodiment of this invention. It is a block diagram which shows the schematic structure of the model learning part for parsing which learns the model for parsing each noun phrase and the verb phrase in the 1st Embodiment of this invention. It is a block diagram which shows the schematic structure of the claim translation part which automatically translates the claim of the patent application of English sentence into Japanese in 1st Embodiment of this invention. It is a block diagram which shows the whole structure of the automatic translation system which concerns on 2nd Embodiment of this invention. It is a block diagram which shows the schematic structure of the model learning part for structural component translation which learns various models for automatic translation in the 2nd Embodiment of this invention. It is a block diagram which shows the schematic structure of the claim translation part which automatically translates the claim of the patent application of English sentence into Japanese in the 2nd Embodiment of this invention. It is a figure which shows the appearance of the computer system which is the hardware flat form which realizes the automatic translation system which concerns on each embodiment of this invention. It is a block diagram which shows the hardware structure inside the computer system shown in FIG.

In the following description and drawings, the same parts are given the same reference numbers. Therefore, detailed explanations about them will not be repeated. The following explanation relates to the case of translating an English claim of a patent application into Japanese. However, the translation target of the following embodiments is not limited to such cases. For example, it can be applied when translating a Japanese claim into an English sentence. Furthermore, it can be applied to laws, regulations, treaties, and instructions for use of various machines, electronic devices, and software that are difficult to translate as well as claims of patent applications.

[First Embodiment]
<Basic idea>
In the claim automatic translation system according to the first embodiment described below, statistical machine translation is adopted. With reference to FIG. 4, in both English and Japanese, the claims of a patent application are described in which structural parts are arranged and described according to any of a certain pattern, although there are various variations. In the present embodiment, the arrangement pattern of the structural parts of such a claim is referred to as a claim pattern. This is similar to the conventional English pattern 74 shown in FIG. 2, and can be represented by a tree structure. An English claim pattern is called an English pattern, and a Japanese claim pattern is called a Japanese pattern. Further, by comparing the actual English claim with the corresponding Japanese claim, the English pattern and the Japanese pattern can be associated with each other. That is, a claim described according to a certain English pattern can be translated into a Japanese claim that matches the Japanese pattern corresponding to the English pattern.

These claim patterns can be obtained by classifying a plurality of claims. In each claim pattern, each structural part is given a grammatical characteristic of that structural part. For example, if the character string or word string forming the structural part is a noun phrase, the structural part is marked as a noun phrase, and if it is a verb phrase, it is marked as a verb phrase. Even if two claim patterns have the same appearance of the tree structure, if the grammatical characteristics of a certain structural part are different, they are different claim patterns. It should be noted that such a claim pattern is different from the "separation pattern" used in Patent Document 1.

By comparing the claim pattern of an English claim with the claim pattern of a Japanese claim corresponding to the English claim, the English pattern and the Japanese pattern can be associated with each other. Further, the corresponding claim patterns can be associated with each other. Such pairs of claim patterns are stored in advance as pattern classification data, and are used when learning a model for translation as described below. In the following embodiment, it is assumed that the English pattern and the Japanese pattern are associated with each other on a one-to-one basis.

In the following embodiment, it is assumed that an English claim is translated into a Japanese claim. The English pattern is determined by decomposing the word string of the English claim, which is the original language, into a plurality of structural parts and determining the grammatical characteristics of each structural component. From this result, the contents of the texts constituting each structural component can also be obtained. As in Patent Document 1, the delimiter pattern in the claim can be used for disassembling the structural parts. In particular, the transition phrase is peculiar to the claim expression and its variation is small, so it can be identified relatively easily. After the transition phrase is specified, the structural part corresponding to the subject of the invention is located either before or after the transition phrase, and one or more structural parts corresponding to the description are located on the other side. Separation between these structural parts can be performed by discovering a specific delimiter pattern, or a statistical model for separation is created in advance to identify the part that is likely to be the delimiter position. Can also be done.

In this embodiment, separate translation models are learned for each grammatical characteristic of structural parts (for example, noun phrase, verb phrase type, etc.), and a model corresponding to the grammatical characteristics of each structural component is used. Automatic translation is performed for each structural part. That is, when translating a structural component consisting of a noun phrase, a pre-learned translation model using only the noun phrase is used. When translating a structural component consisting of a verb phrase, a pre-learned translation model using only the verb phrase is used. As a result, a Japanese translation that is easier to understand for each structural part can be obtained.

The original English pattern is associated with a certain Japanese pattern, and which structural part of the English pattern corresponds to which structural part of the Japanese pattern is also recorded in advance. Using this correspondence, the Japanese translation of the claim can be obtained by substituting the Japanese translation of each structural part into the position of the structural part corresponding to the original structural part of the Japanese pattern corresponding to the original English pattern. Obtainable.

The "translation model" here is a translation model from the original language to the target language used in the statistical translation device, a language model of the target language, and a phrase table generated in the process of learning them. Refers to a set of models consisting of etc. In the following embodiments, the original language is English and the target language is Japanese. Further, in the following embodiment, a noun phrase and a verb phrase are adopted as grammatical characteristics.

FIG. 4 schematically shows the learning process of the noun phrase translation model 210 and the verb phrase translation model 214. Here, the English pattern and the Japanese pattern are collected in advance, and the correspondence between them is also established. This correspondence is one-to-one as described above. That is, once the English pattern is determined, the corresponding Japanese pattern is also determined, and the correspondence between the structural parts is also determined on a one-to-one basis.

With reference to FIG. 4, a plurality of claim translations 170 for learning are prepared. Each translation 170 includes an English claim 172 and a Japanese claim 174, which is a translation of the English claim. For example, there are many application documents that have a bilingual relationship with each other, such as a Japanese application translated into English and filed in the United States. The bilingual translation 170 may be selected from the set of published applications.

The English pattern of the English claim 172 is determined, the claim is decomposed into structural parts according to the English pattern, and the grammatical characteristics of each structural part are classified according to the English pattern (process 176). Similarly, Japanese claim 174 is decomposed into structural parts according to the Japanese pattern corresponding to the English pattern, and the grammatical characteristics of each structural part are classified according to the Japanese pattern (process 180). This process can be executed manually, or can be executed by a machine by preparing a machine-readable pattern collection.

As a result of this pattern classification, for example, the English claim 172 is determined to match the English pattern 178, and the Japanese claim 174 is determined to match the Japanese pattern 182. The structural parts of the English pattern and the Japanese pattern are associated in advance, and each structural part can be paired (process 184). For example, in the example shown in FIG. 4, the subject of the English claim is associated with the subject of the Japanese claim. The transition phrase of the English claim is associated with the transition phrase of the Japanese claim. The explanation of the English claim is associated with the explanation of the Japanese claim.

Among the structural parts combined in this way, in the present embodiment, attention is paid to the grammatical characteristics of the structural parts, and the pairs of the structural parts are classified into separate sets for each grammatical characteristic (grammatical type). That is, if the text of the pair of structural parts is a noun phrase, it is classified into a set of noun phrase pairs 198, and if it is a verb phrase, it is classified into a set of verb phrase pairs 200 (process 196). The set of noun phrase pairs 198 thus obtained contains many pairs of noun phrases of English claims and noun phrases of Japanese claims, and the set of verb phrase pairs 200 includes verb phrases of English claims and Japanese. Includes many pairs with claim verb phrases.

Finally, the noun phrase translation model 210 is trained using the noun phrase pair set 198 as training data, and the verb phrase translation model 214 is trained using the verb phrase pair set 200 as training data.

With reference to FIG. 5, the following processing is performed when translating from English to Japanese. By collating the English claim 70 with the English pattern, the claim pattern of the English claim 70 is specified. In this example, it is assumed that the English pattern 74 is specified. The English pattern 74 includes three structural parts in this example: the subject, the transition phrase, and the description. Each of these has an area for storing text. In these areas, the English text corresponding to the subject, the English text corresponding to the transition phrase, and the English text corresponding to the explanation, which are obtained by decomposing the English claim 70 into structural parts, are stored (process 72).

Next, the Japanese pattern associated with the specified English pattern is identified and generated (process 76). Specifically, information specifying the structure of the Japanese pattern 78 paired with the specified English pattern 74 is read from the claim pattern storage unit, and a new area for storing this new Japanese pattern is newly created. Reserved in the storage area. In this example, the read Japanese pattern 78 includes a description, a transitional phrase, and a subject. These are arranged in this order, and each has an area in the storage device for storing the corresponding text. These areas are empty when the Japanese pattern is generated.

Subsequently, the subject, transitional phrase, and explanatory texts of the English pattern are automatically translated using the model selected according to each grammatical characteristic (process 240). For example, since the subject is a noun phrase, it is translated using a noun phrase translation model. Of the explanations, those that are verb phrases are translated using the verb phrase translation model, and those that are noun phrases are translated using the noun phrase translation model. Since the transition phrase is a verb phrase, it is translated using a verb phrase translation model. The result of translating the subject is in the area of the subject of the Japanese pattern 78, the result of translating the transition phrase is in the area of the transition phrase of the Japanese pattern 78, and the result of translating the explanation is in the area of the explanation of the Japanese pattern. Each is stored. As a result, the translated Japanese pattern 242 is obtained. By concatenating the texts in each area of the obtained Japanese pattern 242 in the order of each area of the Japanese pattern 242, a Japanese claim which is a translation of the English claim 70 can be obtained.

<Structure>
The claim translation system 300 according to the present embodiment, which realizes the above-mentioned automatic translation process with reference to FIG. 6, is for translating an English claim into a Japanese claim. The claim translation system 300 learns the nose phrase translation model 320 used when performing English-Japanese translation of nomenclature, the verb phrase translation model 322 used when performing English-Japanese translation of verb phrases, and learning these. The English claim 324 is syntactically analyzed using the structural component translation model learning unit 318, the verb phrase syntactic analysis model 312, and the noun phrase syntactic analysis model 314, and the English claim 334 after the syntactic analysis is output. A claim for translating the input syntactically analyzed English claim 334 and outputting a Japanese claim 326 using the syntactic analysis unit 332 for Includes translation unit 328.

The structural component translation model learning unit 318 has a claim translation corpus 310 that stores a large number of translations of claims between English and Japanese in order to learn the nose phrase translation model 320 and the verb phrase translation model 322, and English. Model 312 for verb phrase syntax analysis for syntactic analysis of verb phrases, model 314 for nose phrase syntax analysis for syntactic analysis of English nomenclature, English patterns related to claims and corresponding Japanese patterns Pattern classification data 316, which stores a plurality of pairs of and, and syntax analysis tree conversion rule 330 for pre-sorting English structural parts, which will be described later, are used. The verb phrase parsing model 312 is a parsing model in which only English verb phrases are learned as training data. The noun phrase parsing model 314 is a parsing model in which only English noun phrases are learned as learning data.

With reference to FIG. 7, the structural component translation model learning unit 318 reads the phrase translations in order from the phrase translation corpus 310, and collates the English claims with the pattern classification data 316 to obtain a matching English pattern. Based on the pattern classification unit 350 for reading the Japanese pattern paired with the English pattern, the phrase read from the claim translation corpus 310, and the English pattern read by the pattern classification unit 350, the English claim is made. A structural part classification unit 352 that classifies into structural parts and classifies them into nomenclatures and verb phrases and outputs them, a nomenclature data storage unit 354 that stores English nomenclatures output by the structural parts classification unit 352, and structural parts. It includes a verb phrase data storage unit 356 that stores a verb phrase output by the classification unit 352.

The model learning unit 318 for structural component translation further performs a noun phrase pre-arrangement process, which is called pre-sorting, for each of the noun phrase data stored in the noun phrase data storage unit 354. It is stored in the rearrangement unit 358, the noun phrase learning data storage unit 362 for storing the pre-sorted noun phrase data output by the noun phrase pre-sorting unit 358, and the noun phrase learning data storage unit 362. It includes a noun phrase model learning unit 364 for learning the noun phrase translation model 320 by using the pair of the English phrase and the Japanese claim after the pre-sorting as learning data.

The model learning unit 318 for structural component translation further includes a verb phrase pre-sorting unit 360 that pre-sorts words in English claims for each of the verb phrase data stored in the verb phrase data storage unit 356, and a verb. After pre-sorting, the verb phrase learning data storage unit 366 for storing the pre-sorted verb phrase data output by the phrase pre-sorting unit 360 and the verb phrase learning data storage unit 366 for storing the pre-sorted verb phrase data. It includes a verb phrase model learning unit 368 for learning the verb phrase translation model 322 using a pair of an English claim and a Japanese claim as learning data.

The nomenclature pre-sorting unit 358 performs parsing for each of the nomenclatures stored in the nomenclature data storage unit 354 using the parsing model 314, and outputs a parsing tree. 380 and the parsing tree output by the parser 380 are transformed into a parsing tree with the same structure as the corresponding Japanese parsing tree using the conversion rules stored in the parsing tree conversion rule 330. To rearrange English words in the order specified by the parsing tree after conversion output by the tree structure conversion unit 382 and the tree structure conversion unit 382, and store them in the learning data storage unit 362 for nomenclature. Includes a word rearrangement unit 384. In this embodiment, pre-sorting is used to sort the words in the original text (English) before translation, but this is just an example, and other sorting methods can also be used. .. For example, it is possible to adopt a so-called post-sorting method in which words are translated according to the order of words in the original text and then the translated words are rearranged so as to be in the word order of the target language.

The verb phrase pre-sorting unit 360 performs parsing for each of the verb phrases stored in the verb phrase data storage unit 356 using the verb phrase parsing model 312, and is a parsing unit for outputting a parsing tree. The parsing tree 400 and the parsing tree output by the parsing unit 400 are transformed into a parsing tree having the same structure as the corresponding Japanese parsing tree by using the conversion rule stored in the parsing tree conversion rule 330. To the verb phrase learning data storage unit 366, the words in the English verb phrase are rearranged in the order specified by the converted parsing tree output by the tree structure conversion unit 402 and the tree structure conversion unit 402. Includes a word sort unit 404 for storage.

In the present embodiment, the verb phrase syntax analysis model 312 and the noun phrase syntax analysis model 314 are models that have been trained to perform syntactic analysis of an English verb phrase and an English noun phrase, respectively. These learnings are performed in advance by the parsing model learning unit 442 having the configuration shown in FIG.

With reference to FIG. 8, the syntactic analysis model learning unit 442 extracts a nose phrase from the original language corpus 440 that stores a large amount of English documents, and decomposes the nose phrase and the nose phrase into various obtained ones. To memorize the nose phrase extraction / decomposition unit 460 for extracting words and phrases (hereinafter, these are collectively simply referred to as "noun phrase") and the nose phrase extracted by the nose phrase extraction / decomposition unit 460. Nose phrase for learning the nose phrase syntax analysis model 314 using the nose phrase syntax analysis learning data storage unit 462 and the nose phrase syntax analysis learning data storage unit 462 stored in the nose phrase syntax analysis learning data storage unit 462. Includes a model learning unit 464 for syntax analysis. The noun phrase parsing model 314 is a statistical model, and when an English noun phrase is given, it learns to output the most probable parsing tree as the noun phrase parsing tree.

The model learning unit 442 for syntax analysis further includes a verb phrase extraction unit 470 for extracting verb phrases from the original language corpus 440 and a verb phrase syntax analysis learning unit for storing the verb phrases extracted by the verb phrase extraction unit 470. Verb phrase syntax analysis learning unit for learning the verb phrase syntax analysis model 312 using the data storage unit 472 and the verb phrase syntax analysis learning data stored in the data storage unit 472. Includes 474 and. The verb phrase parsing model 312 is a statistical model, and when an English verb phrase is given, learning is performed so as to output the most probable parsing tree as the parsing tree for the verb phrase.

With reference to FIG. 9, the claim translation unit 328 shown in FIG. 6 receives the input English claim 334 after parsing, searches the pattern classification data 316, and matches the English claim 334 after parsing. A pattern search unit 500 that reads out a pattern pair having a pattern, a Japanese pattern generation unit 504 that extracts a Japanese pattern from the pattern pair read by the pattern search unit 500, and generates an empty Japanese pattern, and Japanese. It includes a Japanese pattern storage unit 506 for storing the Japanese pattern generated by the pattern generation unit 504.

The claim translation unit 328 further receives the parsed English phrase 334 with the English pattern from the pattern search unit 500, and divides the parsed English phrase 334 into structural parts according to the structure of the English pattern. The output input sentence dividing unit 502, the structural part storage unit 508 for storing the structural parts output by the input sentence dividing unit 502 in association with the corresponding parts of the English pattern, and the structure stored in the structural part storage unit 508. The values of the judgment signal output by the structural part type judgment unit 510 and the structural part type judgment unit 510 that read the parts in order, identify whether the read structural part is a nomenclature phrase or a verb phrase from the English pattern and output a judgment signal. In response to the selector 512 that selects either the verb phrase parsing model 312 or the nomenclature parsing model 314, and in response to the value of the determination signal output by the structural component type determination unit 510, the verb. The selector 522 for selecting either the phrase translation model 322 or the nose phrase translation model 320 and the structural component whose type was determined by the structural component type determination unit 510 were read from the structural component storage unit 508, and the selector 512 was selected. It includes a parsing unit 514 that performs parsing using a parsing model and outputs a parsing tree.

The claim translation unit 328 further converts the parsing tree output by the parsing unit 514 into the corresponding Japanese parsing tree according to the parsing tree conversion rule stored in the parsing tree conversion rule 330. Includes a tree conversion unit 518.

The claim translation unit 328 further includes a word rearrangement unit 520 that outputs an English word string in which English words are rearranged according to the order in which each word appears in the Japanese syntactic analysis output by the syntactic analysis tree conversion unit 518. The translation model selected by the selector 522 after receiving the English word string output by the word sorting unit 520 (the verb phrase translation model 322 when the target input sentence is a verb phrase, and the nose phrase translation when the target input sentence is a nomenclature phrase). The Japanese pattern storage unit 506 stores the Japanese sentences output by the statistical automatic translator 524 that performs statistical automatic translation from English to Japanese using the model 320) and the statistical automatic translator 524. Includes a Japanese pattern substitution unit 526 that assigns to the area of the pattern corresponding to the structural component to be translated.

<Operation>
The claim translation system 300 whose configuration has been described above operates as follows. The operation of the claim translation system 300 is roughly divided into three stages. The first step is the learning of the verb phrase parsing model 312 and the noun phrase parsing model 314 shown in FIG. The second step is the learning of the noun phrase translation model 320 and the verb phrase translation model 322 shown in FIG. When these two processes are completed, the parsing unit 332 and the claim translation unit 328 can translate the English claim 324 into Japanese and output the Japanese claim 326 as the third step. Prior to these processes, it is assumed that the pattern classification data 316 and the parsing tree conversion rule 330 are prepared in advance by some means. Hereinafter, these will be described in order.

<Phase 1: Learning a model for parsing>
With reference to FIG. 8, an English original language corpus 440 is prepared in advance. For each sentence stored in the original language corpus 440, morphological analysis and syntactic analysis are performed in advance, and information on the syntactic analysis tree is attached.

The noun phrase extraction / decomposition unit 460 extracts a noun phrase, a word contained in the noun phrase, and a partial analysis tree showing the syntactic structure of the noun phrase from the original language corpus 440, and stores the noun phrase parsing learning data. Accumulate in unit 462. The noun phrase syntax analysis model learning unit 464 reads the information stored in the noun phrase syntax analysis learning data storage unit 462 and learns the noun phrase syntax analysis model 314. As a result of this learning, the noun phrase parsing model 314 can be used for calculations that output the most probable syntactic analysis tree given an English noun phrase.

On the other hand, the verb phrase extraction unit 470 extracts a verb phrase and a partial analysis tree showing the syntactic structure of the verb phrase from the original language corpus 440, and stores the verb phrase in the verb phrase parsing learning data storage unit 472. The verb phrase syntax analysis model learning unit 474 reads the verb phrase stored in the verb phrase syntax analysis learning data storage unit 472 and learns the verb phrase syntax analysis model 312. As a result of this learning, the verb phrase parsing model 312 can be used for calculations that output the most probable syntactic analysis tree given an English verb phrase.

<Second stage: Learning of translation model>
With reference to FIG. 7, the pattern classification unit 350 identifies the pattern of the English claim by reading the Japanese-English translation of the claim from the claim translation corpus 310 and collating it with the pattern included in the pattern classification data 316. The pattern classification unit 350 reads the specified English pattern from the pattern classification data 316. The pattern classification unit 350 divides an English claim into structural parts according to an English pattern and substitutes them into a text area of each structural part of the English pattern. The structural component classification unit 352 stores each structural component of the English pattern to which the text is assigned in the noun phrase data storage unit 354 if it is a noun phrase, and in the verb phrase data storage unit 356 if it is a verb phrase. Claim Biling When this process is completed for all claim translations stored in the corpus 310, or in parallel with this process, the noun phrase pre-sorting unit 358 and the verb phrase pre-sorting unit 360 can operate.

The parsing unit 380 of the noun phrase pre-sorting unit 358 reads the noun phrase data stored in the noun phrase data storage unit 354 in order. The parsing unit 380 outputs a parsing tree to the tree structure conversion unit 382 by parsing the read noun phrase using the noun phrase syntactic analysis model 314. When a parse tree is given, the tree structure conversion unit 382 uses one of the tree structure conversion rules stored in the parse tree conversion rule 330 to convert the parse tree of the nomenclature of the English claim into the corresponding Japanese parse tree. It is converted into a parsing tree structure and given to the word sorting unit 384. The word sorting unit 384 rearranges English words in the order determined by the Japanese parsing tree and stores them in the noun phrase learning data storage unit 362. The noun phrase model learning unit 364 learns the noun phrase translation model 320 using the noun phrase learning data accumulated in the noun phrase learning data storage unit 362 in this way.

The syntactic analysis unit 400 of the verb phrase pre-sorting unit 360 reads out the verb phrase data stored in the verb phrase data storage unit 356 in order. The parsing unit 400 outputs the parsing tree to the tree structure conversion unit 402 by parsing the read verb phrase using the verb phrase parsing model 312. Given a parsing tree, the tree transforming unit 402 uses one of the tree parsing rules stored in the parsing rule 330 to translate the parsing tree of the verb phrase in the English claim into the corresponding Japanese parse tree. It is converted into a parsing tree structure and given to the word sorting unit 404. The word sorting unit 404 rearranges English words in the order determined by the Japanese parsing tree and stores them in the verb phrase learning data storage unit 366. The verb phrase model learning unit 368 learns the verb phrase translation model 322 using the verb phrase learning data accumulated in the verb phrase learning data storage unit 366 in this way.

When the processing for all the claim translations stored in the claim translation corpus 310 is completed as described above, the English-Japanese nomenclature and the verb phrase are used by using the nose phrase translation model 320 and the verb phrase translation model 322. Can be translated.

<Third stage: Complaint translation>
When translating a claim, the claim translation unit 328 operates as follows. Prior to the processing of the claim translation unit 328, the syntax analysis unit 332 shown in FIG. 6 performs a syntactic analysis on the English claim 324. As a result, the English claim 334 after parsing is input to the claim translation unit 328.

With reference to FIG. 9, when the English claim 334 after the parsing is given to the claim translation unit 328, the pattern search unit 500 collates the pattern stored in the pattern classification data 316 with the English claim 334 after the parsing. And search for a matching English pattern. The pattern search unit 500 reads the English pattern that best matches the English of the English claim 334 after parsing from the pattern classification data 316 together with the corresponding Japanese pattern, and sends the Japanese pattern to the Japanese pattern generation unit 504 in English. The pattern and the English claim 334 after parsing are given to the input sentence division unit 502, respectively.

Based on the given Japanese pattern, the Japanese pattern generation unit 504 creates a storage area for the text of the Japanese pattern in the Japanese pattern storage unit 506 for each structural component of the Japanese pattern. Each of these regions is empty at this point.

The input sentence division unit 502 divides the English claim 334 after parsing according to the English pattern read from the pattern classification data 316, substitutes it into the text storage area of each structural component, and stores it in the structural component storage unit 508. .. When the structural component type determination unit 510 stores these structural components in the structural component storage unit 508, the structural component type determination unit 510 reads one of them, determines whether the structural component is a noun phrase or a verb phrase, and selects a determination signal 512 and a determination signal. Give to selector 522. In response to this determination signal, the selector 512 selects the verb phrase syntax analysis model 312 when the structural component is a verb phrase, and the noun phrase syntax analysis model 314 when the structural part is a noun phrase. Connect to the syntax analysis unit 514. The selector 522 selects a verb phrase translation model 322 when the structural part is a verb phrase, and a noun phrase translation model 320 when it is a noun phrase, and connects it to the statistical automatic translator 524. ..

The parsing unit 514 receives the structural parts from the structural part type determination unit 510, and uses the verb phrase parsing model 312 or the noun phrase parsing model 314 connected via the selector 512 to parse the structural parts. To do. As a result, the parsing tree is given from the parsing unit 514 to the parsing tree conversion unit 518. In this parsing, the parsing unit 514 uses the noun phrase parsing model 314 if the target structural part is a noun phrase, and uses the verb phrase parsing model 312 if it is a verb phrase. , Parsing the target structural parts respectively. The verb phrase syntax analysis model 312 was trained by the syntactic analysis model learning unit 442 shown in FIG. 8 using the verb phrase as training data, and the noun phrase syntax analysis model 314 was used for syntactic analysis using the noun phrase as training data. It was learned by the model learning unit 442. Therefore, the syntactic analysis performed by the noun phrase parsing model 314 is optimal according to the type of structural component.

When the parsing tree conversion unit 518 receives the parsing tree from the parsing unit 514, it refers to the parsing tree conversion rule 330, and in the conversion rules, the parsing rule for the parsing tree received from the parsing unit 514. Select the most suitable one as. The parsing tree conversion unit 518 uses the selected conversion rule to translate the parsing tree given by the parsing unit 514 into the English represented by the parsing tree in Japanese. Convert to a parse tree that shows the syntax structure. In this converted parsing tree, the structure of the tree is the structure of a Japanese sentence, but the words assigned to each leaf are English words. The parse tree conversion unit 518 gives the converted parse tree to the word sorting unit 520.

The word rearranging unit 520 uses the converted syntactic analysis tree given by the parsing tree conversion unit 518, and rearranges the English words appearing in the parsing tree in the order determined by the structure of the parsing tree. The obtained English word string is given to the statistical automatic translator 524 as an input.

The statistical automatic translator 524 performs statistical translation from English to Japanese using the translation model connected by the selector 522 on this English word string, and the Japanese word string obtained as a result. Is given to the Japanese pattern substitution unit 526. The translation model used at this time is a verb phrase translation model 322 when the structural component to be translated is a verb phrase, and a noun phrase translation model 320 when it is a noun phrase. Therefore, in the English-Japanese translation of each structural part, there is a high possibility that a noun phrase will be correctly translated as a noun phrase and a verb phrase will be translated as a verb phrase.

The Japanese pattern substitution unit 526 uses the correspondence information between the original English structural parts and the Japanese pattern structural parts for the Japanese word string output by the statistical automatic translator 524 to store the Japanese pattern. The translation result is substituted into the text storage area of the corresponding structural component of the Japanese pattern stored in the part 506.

When the above processing is performed on all the structural parts stored in the structural part storage unit 508, each structural part of the Japanese pattern stored in the Japanese pattern storage unit 506 has a corresponding English structure. The word string that translates the parts into Japanese will be stored. These translation results were performed using the verb phrase translation model 322, which was optimized for verb phrases if the original English structural part was a verb phrase, and optimized for noun phrases if it was a noun phrase. This was done using a noun phrase translation model 320. Therefore, it is expected that the individual structural parts that make up the Japanese claim will be accurate, reflecting the type of the original English structural part. Further, the structural parts of the Japanese pattern storage unit 506 are arranged according to the appearance order of each structural part in the Japanese claim. Therefore, by reading out each structural component stored in the Japanese pattern storage unit 506 in order from the beginning and displaying it appropriately, a Japanese claim 326 which is an accurate translation of the English claim 324 can be obtained.

[Second Embodiment]
In the first embodiment, in the learning of the translation model by the structural component translation model learning unit 318 shown in FIG. 7, noun phrase and verb by the noun phrase pre-sorting unit 358 and the verb phrase pre-sorting unit 360. Pre-sorting phrases. Further, when the English claim is translated by the claim translation unit 328 shown in FIG. 9, the words are pre-sorted by the parsing unit 514, the parsing tree conversion unit 518, and the word rearrangement unit 520. In this way, by performing pre-sorting of words at the time of learning and translation, the learning efficiency is improved and the translation accuracy is also improved. Therefore, it is desirable to rearrange the words in this way. However, the present invention is not limited to such embodiments. Separate translation models are learned in advance according to the grammatical characteristics of the structural parts of the English pattern (types such as verb phrases or nomenclature phrases), and at the time of translation, according to the grammatical characteristics of the structural parts to be translated. Anything can be used as long as it automatically translates using an appropriate translation model. The second embodiment omits the pre-sorting.

FIG. 10 shows the overall configuration of the claim translation system 550 according to the second embodiment in a block diagram format. The claim translation system 550 differs from the claim translation system 300 shown in FIG. 6 in that it does not use the verb phrase parsing model 312, the nomenclature parsing model 314, and the parsing tree conversion rule 330, and FIG. Instead of the structural component translation model learning unit 318 and the claim translation unit 328 of 6, each includes the structural component translation model learning unit 570 and the claim translation unit 572, which do not use pre-sorting. The verb phrase parsing model 312 and the noun phrase parsing model 314 are used in the parsing by the parsing unit 332, as in the first embodiment.

FIG. 11 shows a schematic block diagram of the structural component translation model learning unit 570. The structural component translation model learning unit 570 differs from the structural component translation model learning unit 318 shown in FIG. 7 in that the nomenclature data storage unit 354, the verb phrase data storage unit 356, and the nomenclature pre-arrangement for pre-sorting. The nomenclature output from the structural part classification unit 352 is directly stored in the nomenclature learning data storage unit 362 without including the replacement unit 358 and the verb phrase pre-sorting unit 360, and the verb phrase is directly learned for the verb phrase. This is a point accumulated in the data storage unit 366. In other respects, each part of the structural part translation model learning unit 570 is the same as the corresponding part of the structural part translation model learning unit 318.

FIG. 12 shows a schematic block diagram of the claim translation unit 572 shown in FIG. The claim translation unit 572 does not include the selector 512, the parsing unit 514, the parsing tree conversion rule 330, the parsing tree conversion unit 518, and the word sorting unit 520, which are necessary for pre-sorting, as compared with FIG. It differs from the claim translation unit 328 of FIG. 9 in that the output of the structural component storage unit 508 is directly given to the statistical automatic translator 524. In other respects, each part of the claim translation unit 572 has the same configuration as each part of the claim translation unit 328.

This second embodiment is different from the first embodiment in that pre-sorting is not performed at the time of training the translation model and at the time of translation. Even with such a configuration, the translation of the claim can be performed with higher accuracy than the conventional one. In particular, in translation between languages in which the word orders of the two languages are relatively close, it can be expected that the translation accuracy will be sufficiently improved without using pre-sorting.

<Modification example>
The first embodiment uses pre-sorting, and the second embodiment does not use pre-sorting. The present invention is not applicable only to such embodiments, and for example, it may be possible to select whether or not to perform pre-sorting. By doing so, it is possible to perform both translations that are pre-sorted according to the target language and translations that are not.

Further, in the above embodiment, a model for verb phrase translation and a model for noun phrase translation are used as translation models. However, the present invention is not limited to such an embodiment, and other models, such as a model for adverbial phrase translation, can also be used.

[Realization by computer]
The claim translation system 300 according to the first embodiment, the claim translation system 550 according to the second embodiment, and other modifications are computer hardware and a computer program executed on the computer hardware. It can be realized by. FIG. 13 shows the appearance of the computer system 930, and FIG. 14 shows the internal configuration of the computer system 930.

With reference to FIG. 13, the computer system 930 includes a computer 940 having a memory port 952 and a DVD (Digital Versaille Disc) drive 950, a keyboard 946, a mouse 948, and a monitor 942.

With reference to FIG. 14, the computer 940 has a memory port 952 and a DVD drive 950, as well as a CPU (central processing unit) 956, a CPU 956, a bus 966 connected to the memory port 952 and the DVD drive 950, and a boot program. It includes a read-only memory (ROM) 958 that stores the data, a random access memory (RAM) 960 that is connected to the bus 966 and stores program instructions, system programs, work data, and the like, and a hard disk 954. The computer system 930 further includes a network interface (I / F) 944 that provides a connection to a network 968 that allows communication with other terminals.

A computer program for causing the computer system 930 to function as each functional unit of the claim translation system 300 and the claim translation system 550 according to the above-described embodiment is provided in the DVD 962 or the removable memory 964 mounted on the DVD drive 950 or the memory port 952. It is stored and further transferred to the hard disk 954. Alternatively, the program may be transmitted to the computer 940 via network 968 and stored on the hard disk 954. The program is loaded into RAM 960 at run time. Programs may be loaded directly into RAM 960 from DVD 962, from removable memory 964, or via network 968.

This program includes an instruction sequence including a plurality of instructions for causing the computer 940 to function as each functional unit of the claim translation system 300 and the claim translation system 550 according to the above embodiment. Some of the basic functions required to force the computer 940 to perform this operation are operating systems or third-party programs running on the computer 940 or various dynamically linkable programming toolkits or programs installed on the computer 940. Provided by the library. Therefore, the program itself does not necessarily include all of the object code for realizing the functions required to realize the system and method of this embodiment. This program, as described above, by dynamically invoking the appropriate function or the appropriate program in the programming toolkit or program library at runtime in a controlled manner to obtain the desired result within the instructions. It is only necessary to include the instruction that realizes the function of. Of course, the program alone may provide all the necessary functions.

In addition, each functional part of the claim translation system 300 and the claim translation system 550 may be distributed and processed on different computers, or may be distributed and processed by different computers existing in different regions via a network. You may.

The embodiments disclosed this time are merely examples, and the present invention is not limited to the above-described embodiments. The scope of the present invention is indicated by each claim of the scope of claims, taking into consideration the description of the detailed description of the invention, and all changes within the meaning and scope equivalent to the wording described therein. include.

300, 550 Claim translation system 310 Claim translation corpus 312 Model for verb phrase parsing 314 Model for nomenclature parsing 316 Pattern classification data 318 570 Model for structural parts translation Learning unit 320 Model for verb phrase translation 322 Model for verb phrase translation 328, 572 Claim translation department 330 Parsing tree conversion rule 358 Nomenclature pre-sorting section 360 Verb phrase pre-sorting section 442 Parsing model learning section 524 Statistical automatic translator

Claims (4)

An automatic translation device that translates a sentence in the first language into a sentence in a second language different from the first language.
A dividing means for identifying the sentence pattern of the first language and dividing the sentence of the first language into structural parts, and
Correspondence relationship storage means for memorizing the correspondence between the sentence pattern of the first language and its structural parts and the sentence pattern of the second language and its structural parts, and the grammatical characteristics of each structural part. there are, the pattern of said pattern and sentence of the second language sentence of the first language are each, and correlation storage means you express in a tree structure,
By referring to the correspondence storage means, the pattern specifying means for specifying the pattern of the sentence of the second language, which is previously associated with the pattern of the sentence of the first language specified by the dividing means. ,
Using the correspondence relationship stored in the correspondence storage means, the pattern of the sentence of the first language divided by the division means and the pattern of the second language specified by the pattern identification means. Correspondence identification means for specifying the correspondence relationship of structural parts and the grammatical characteristics of each structural part,
For each of the structural parts of the first language, for the word string constituting the structural part, for translation from the first language to the second language according to the grammatical characteristics of the structural part. A translation means that performs automatic translation using a model to generate a translation of the second language,
For each of the structural parts of the first language, the translation of the second language obtained by the translation means is subjected to the pattern identification means according to the correspondence relationship of the structural parts specified by the correspondence identification means. By substituting into any of the structural components of the pattern of the sentence in the second language specified by, the substitution means for generating the sentence in the second language, which is a translation of the sentence in the first language. Including
The dividing means
Predetermined delimiters used to delimit structural components of a sentence in the first language contained in the sentence in the first language, based on the result of parsing the sentence in the first language. A sentence dividing means for dividing a sentence in the first language into a plurality of structural parts by detecting a pattern, and a sentence dividing means.
A grammatical characteristic determining means for determining the grammatical characteristics of each structural component divided by the sentence dividing means based on the result of parsing for the structural component,
Automatic translation including means for specifying the sentence pattern of the first language based on the appearance order of the structural parts divided by the sentence dividing means and the grammatical characteristics determined by the grammatical characteristic determining means for each structural part. Device. Further, it receives the structural component of the first language provided between the corresponding identification means and the translation means and whose correspondence relationship and grammatical characteristics are specified by the corresponding identification means, and receives the structural component of the first language. For each of the above, the order of the word strings constituting the structural component is rearranged according to the order in which the translated words of the word strings appear in the second language, and the word rearranging means is given to the translation means. , The automatic translation apparatus according to claim 1. The word rearrangement means
A parsing means for generating a parsing tree for the first language by performing a parsing for each of the structural parts of the first language according to the grammatical characteristics of the structural parts.
A conversion means for converting the parsing tree of the first language generated by the parsing means into a parsing tree of the second language according to a conversion rule prepared in advance.
A claim including a rearrangement means for rearranging and outputting the words of the structural parts of the first language according to the order of appearance of the words in the parsing tree of the second language obtained by the conversion by the conversion means. Item 2. The automatic translation device according to item 2. The translation means is pre-optimized for the translation of the word string having the grammatical characteristics of the structural part with respect to the word string constituting the structural part for each of the structural parts of the first language. Claims 1 to 3 include a grammatical characteristic-specific translation means that performs automatic translation using a model for translation from the first language to the second language and generates a translation of the second language. The automatic translation device described in any of.

Images