JP2006004366A - Machine translation system and computer program therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a machine translation device capable of expanding a range of sentences that can be translated.
A machine translation system (32) is a machine translation system for translating an input sentence (30), and includes a corpus (46) composed of translatable sentences and a first machine translation device (42) for translating the input sentence. The machine translation device outputs an index of the translation quality of the translation. The machine translation system 32 further includes a similar sentence search unit 48 that searches the corpus 46 for a sentence similar to the input sentence 30 when the translation quality is poor, and a second sentence that translates the sentence searched by the similar sentence search unit 48. Machine translation device 50.
[Selection] Figure 1

Description

  The present invention relates to a machine translation technique, and more particularly to a technique for improving translation performance for an original sentence including many sentences that are difficult to machine translate such as spoken language.

  The purpose of speech translation is to translate speech utterances in real time and output them as speech. The constituent technologies include speech recognition, machine translation, and speech synthesis.

  The machine translation unit in the speech translation uses the utterance sentence output from the speech recognition unit as a translation target. However, because these sentences have spoken language specific properties, their translation is difficult. For example, spoken language utterances often have ineligibility such as stagnation, rephrasing, repetition, and omission of particles. These phenomena lead to a decrease in accuracy of processing such as morphological analysis and syntax analysis. In addition, the language phenomenon is complicated by the fact that various treatment expressions are frequently used in spoken language. For example, in spoken language, various expressions appearing at the end of the sentence, such as “Please do”, “Can you please do it”, “Please do it”, etc.

  Further, as a problem derived from voice recognition, it is mentioned that the utterance obtained from the voice recognition unit is not necessarily a “sentence” suitable for translation processing. In speech recognition, utterances are divided by silent intervals. However, in spoken language, a long silent section often appears in a sentence, or a short silent section often appears between sentences. Therefore, the utterance obtained from the voice recognition unit is not necessarily a sentence that is semantically organized.

  One means for solving these problems is a so-called pretreatment technique. This is a technique for rewriting an input sentence so as to become a sentence advantageous to machine translation before being given to machine translation. This is also called pre-editing.

  Regarding preprocessing, Non-Patent Document 1 attempts to manually create conversion rules such as complement of omitted elements, simplification of redundant expressions, and syntax recombination, and apply them to input sentences. Non-Patent Document 2 attempts to improve the accuracy of machine translation by dividing a long input sentence given to machine translation into short sentences. Non-Patent Document 3 takes up inversion, omission, insertion, and emphasis as phenomena that lead to failure in parsing, and attempts to improve the success rate of parsing by applying rewrite rules that eliminate these phenomena. Furthermore, Non-Patent Document 4 applies local rewriting rules to the input sentence, such as simplification of expression and deletion of unimportant words. Furthermore, Non-Patent Document 5 attempts to join and divide utterance units in order to obtain semantically coherent sentences using silent section information and recognized utterance n-grams.

Satoshi Shirai and three others, "Automatic Rewriting Translation Method and Its Effects in Japanese-English Machine Translation", IPSJ Journal, 36 (1), 12-21, 1995 Kim Jin-Hae et al., "Automatic short sentence segmentation and subject complementation for Japanese-English machine translation", Transactions of Information Processing Society of Japan, 35 (6), 1018-1028, 1994 Yasuhiko Yoshimi and two others, "Automatic text pre-editing for the realization of a robust English-Japanese machine translation system", Natural Language Processing, 7 (4), 99-118, 2000 Yamamoto, K. et al. "Machine Translation by Interaction between Parafraser and Transfer", 19th International Computer Language Convention (COLING-2002), pp. 1107-1113, 2002 (Yamamoto, K., "Machine Translation by Interaction between Paraphraser and Transfer." In Proc. Of the 19th International Conference on Computational Linguistics (COLING-2002), pp. 1107-1113, 2002.) Toshiyuki Takezawa and 1 other, "Conversion method to language processing units by dividing or joining speech units", natural language processing, 6 (2), 80003-95, 1999

Takezawa, T. Four others, “Towards a broad bilingual corpus for language translation of travel conversations in the real world”, 3rd LREC-2002 Proceedings, pp. 147-152, 2002 (Takezawa, T. et al., "Toward a Broad-coverage Bilingual Corpus for Speech Translation of Travel Conversations in the Real World", In Proc. Of the 3rd LREC, pp. 147-152, 2002 ) Takezawa, T. Another one, “Collecting bilingual dialogues using machine translation for corpus-based speech translation”, Eurospech-2003, pp. 2757-2760, 2003 (Takezawa, T. et al., "Collecting Machine-Translation-Aided Bilingual Dialogues for Corpus-Based Speech Translation." In Eurospeech-2003, pp. 2757-2760, 2003) Kikui, G. 3 other authors, “Creating a corpus for inter-utterance translation”, Eurospeech-2003 Proceedings, pp. 381-384, 2003 (Kikui, G. et al., "Creating Corpora for Speech-to-Speech Translation." In Eurospeech-2003, pp. 381-384, 2003)

  However, the above-described conventional technique has a problem that a large cost is required to prepare a rewrite rule for pre-editing. In addition, since most of the prior art also requires syntax analysis information, there is a problem that correct preprocessing cannot be performed unless the syntax analysis can be performed correctly. In spoken language, the majority of input sentences with ineligibility are already explained, and it is difficult to improve the accuracy of parsing, and in that respect it is difficult to effectively apply the above-mentioned conventional technology to spoken language translation. It is.

  Therefore, an object of the present invention is to provide a machine translation system that can expand the range of sentences that can be translated even when an input sentence including many incompatibility such as spoken language is targeted.

  The machine translation system according to the first aspect of the present invention is a machine translation system for translating an input sentence in a first language into a second language, and is capable of translation into the second language. A pre-selected first corpus including a plurality of sentences in a first language, a first machine translation means for translating an input sentence into a second language, and a translated sentence by the first machine translation means An index output means for outputting an index of translation quality, and a sentence in a predetermined relationship with the input sentence in response to the index output by the index output means meeting a predetermined condition indicating that the translation quality is poor Search means for searching from the first corpus, and second machine translation means for translating the sentence searched by the search means into the second language.

  If the translation quality is poor due to the translation quality index of the first machine translation means, a sentence having a predetermined relationship with the input sentence is searched from the first corpus, and the sentence is translated by the second machine translation means. This is a translation for the input sentence. Since the sentence in the first corpus is a sentence selected as translatable, there is a high possibility that it can be translated by the second machine translation means. Even when the translation quality of the first machine translation means is poor, the translation is replaced by translation for a sentence having a predetermined relationship with the input sentence, so that it is possible to avoid outputting a translation with a poor translation quality as it is. As a result, the range of input sentences that can be translated by the machine translation system can be expanded compared with the range of sentences that can be translated by the first machine translation means.

  Preferably, in response to the fact that the index output by the index output means matches a predetermined condition, the search means determines that the input sentence is the most similar to the input sentence according to the similarity calculated between the input sentence and the input sentence according to a predetermined calculation method. Similar sentence search means for searching the sentence determined to be similar from the first corpus is included.

  When the translation quality by the first machine translation means is poor, a sentence that is most similar to the input sentence is retrieved from the first corpus and translated by the second machine translation means. Since the sentence most similar to the input sentence is translated by the second machine translation means, the translation obtained as a result represents the same content as the translation for the input sentence. A translation having poor translation quality by the first machine translation means can be output as it is, and a translation having almost the same content as the input sentence obtained by the second machine translation can be output.

  More preferably, the similar sentence search means includes a similarity calculation means for calculating a similarity defined based on a common part between each of the plurality of sentences included in the first corpus and the input sentence. And means for extracting, from the first corpus, a sentence determined to be most similar to the input sentence based on the similarity calculated by the similarity calculation means.

  Whether or not the input sentence and each sentence of the first corpus are similar is represented by a similarity defined based on a common part between the input sentence and the first sentence. From the experimental results, it is known that a sentence having a content very similar to the input sentence can be searched from the first corpus with the similarity defined as described above.

More preferably, the similarity calculation means includes a word number calculation means for calculating the number of words of the input sentence, the number of words of the candidate sentence, and the number of words common to the input sentence and the candidate sentence, the input sentence, The similarity between the candidate sentence that is the target of the calculation of the following formula: relevance rate = (number of words common to the input sentence and candidate sentence) / number of words in the candidate sentence recall = (common to input sentence and candidate sentence) The number of words to be calculated) / the number of words in the input sentence.

  By using the function of the relevance ratio and the recall ratio, it is possible to calculate a similarity that accurately represents the size of the common part between the input sentence and the candidate sentence.

  The word number calculation means includes means for calculating the number of words in the input sentence, the number of words in the candidate sentence, and the number of words common to the input sentence and the candidate sentence by multiplying by a predetermined weight according to the type of each word. But you can.

  Depending on the type of word, the weight for expressing the content of the sentence is changed. As a result, for example, it is possible to distinguish the weights of the types of words that are important in expressing the content of the sentence and the words that are not so, and to calculate the similarity for searching for an appropriate candidate sentence according to the purpose.

  Preferably, the means for calculating includes the number of words in the input sentence, the number of words in the candidate sentence, the number of words common to the input sentence and the candidate sentence, a predetermined first weight for the content word, and a function word Means for multiplying each of the predetermined second weights by calculation is included, and the second weight is a positive value smaller than the first weight.

  Content words are considered more important for expressing the content of sentences compared to function words. Thus, by making the first weight for the content word larger than the second weight for the function word, candidate sentences similar to the input sentence can be searched with emphasis on the content of the sentence.

  More preferably, the word number calculating means calculates the number of words in the input sentence, the number of words in the candidate sentence, and the number of words common to the input sentence and the candidate sentence from the number of n-grams in the sentence (n> 0). Including means. The value of n is preferably determined by experiment. For example, n = 1 may be used, and n = 2 may be used.

  Experiments have shown that good results can be obtained by calculating the number of words using n-grams.

  Preferably, the word number calculating means shares the number of words common to the input sentence and the candidate sentence with respect to the means for calculating the number of words in the input sentence and the number of words in the candidate sentence, respectively. And means for calculating by the number of words appearing in the matched order.

  Experiments have shown that relatively good results can be obtained even when the number of common words is calculated in this way.

More preferably, the similarity calculation means calculates the similarity between the input sentence and the candidate sentence by the following expression: similarity = 2 × matching rate × recall rate / (matching rate + recall rate)
Means for calculating according to:

  Experiments have shown that sentences searched using the similarity determined in this way often represent contents similar to the input sentence. Therefore, by using this similarity, even when the first machine translation means cannot obtain a good translation, a translation representing the same content as the input sentence can be obtained by translation of the second machine translation means.

  More preferably, the similar sentence search means further includes means for excluding the candidate sentence including the content word not included in the input sentence from the extraction target by the means for extracting.

  Thus, content words that are not included in the input sentence often add unnecessary limitations to the meaning of the sentence. By excluding candidate sentences including such content words, it is possible to increase the probability that the finally obtained translated sentence accurately conveys the contents of the input sentence.

Preferably, the means for extracting is a candidate sentence that is determined to be most similar to the input sentence based on the similarity calculated by the similarity calculation means and does not include a content word that is not included in the input sentence. And the following conditions: (1) Among the content words included in the input sentence, there are no more than one word in the candidate sentence, or (2) there are two or more content words in common with the input sentence ,
Means for extracting from the first corpus a candidate sentence that satisfies any of the above.

  In the experiment, the best results were obtained under such conditions.

  More preferably, the first machine translation means is in accordance with a bilingual example corpus including a plurality of examples including a sentence in the first language and a sentence in the second language, which are favorable translations, and a predetermined similarity criterion. Means for retrieving an example including a sentence in the first language most similar to the input sentence from the bilingual example corpus, and translation of the sentence in the first language retrieved by the means for retrieving An example translation means for translating an input sentence by correcting a sentence in the second language that is based on a difference between the sentence in the first language searched by the means for searching and the input sentence And the index output means determines whether or not a predetermined similarity defined between the sentence in the first language searched by the means for searching and the input sentence satisfies a predetermined criterion. And means for outputting the judgment result as an index Including.

  By using the example translation unit as the first machine translation unit, it is possible to use the similarity obtained in the example sentence search process in the example translation as an index of translation quality. No separate functional module is required to assess translation quality.

  More preferably, the first machine translation unit and the second machine translation unit are realized by the same machine translation unit.

  By realizing the first machine translation unit and the second machine translation unit using the same machine translation unit, it is possible to expand the range of sentences that can be translated while preventing an increase in resources necessary for translation.

  When executed by a computer, the computer program according to the second aspect of the present invention causes the computer to operate as one of the machine translation systems described above.

  The machine translation system according to the present embodiment is capable of machine translation by searching for a sentence similar to the sentence from a corpus of translatable sentences prepared in advance for an utterance sentence determined to be difficult to translate in speech translation. It is a system that expands the scope of simple sentences. Hereinafter, the configuration and operation of the machine translation system according to this embodiment will be described in order. In the following description, the same parts are denoted by the same reference numerals. Their names and functions are also the same. Therefore, detailed description thereof will not be repeated.

  In the following description of the system configuration of the embodiment, parameter values determined in advance according to the results of preliminary experiments are used. These preliminary experiments were conducted in Japanese, and two types of Japanese corpus were used in the experiments. An utterance corpus, which is a set of input sentences, and a candidate sentence corpus, which is a set of candidate sentences that serve as a basis for searching similar sentences to the input sentence.

  As the utterance corpus, 437 different sentences extracted from the utterance corpus (Non-patent Document 7) created by writing a dialogue performed assuming a certain situation while traveling (first utterance corpus) are used. . Of the utterance corpus of Non-Patent Document 7, the part other than the first utterance corpus (second utterance corpus) is used in the experiment described last. As the candidate sentence corpus, a corpus (described in Non-Patent Document 8) containing basic expressions of travel conversation was used (first basic expression corpus). The first basic expression corpus is different and includes 176,145 sentences.

[Constitution]
FIG. 1 is a block diagram of a machine translation system 32 according to an embodiment of the present invention. Referring to FIG. 1, this machine translation system 32 attempts to translate a Japanese input sentence 30 into English, outputs an output sentence 34 if it can be translated, and whether or not further translation is possible. The translation availability signal 36 indicating that is output. As will be described later, the machine translation system 32 can be realized by a computer system and a computer program executed thereon.

  The machine translation system 32 is an example-based machine translation system (example translation system), and is computer-readable including a large number of pairs of Japanese sentences (referred to as example sentences) and corresponding English translations. A bilingual example corpus (hereinafter simply referred to as “example corpus”) 40 and a Japanese sentence similar to the input sentence 30 are searched in the example corpus 40, and an English translation of the searched Japanese sentence is referred to as an input sentence 30. The input sentence 30 is translated by correcting the difference based on the difference from the searched example sentence, and the translation result 58 is output. The similarity between the searched example sentence and the input sentence 30 exceeds a predetermined value. And a machine translation device 42 for outputting a determination signal indicating whether or not there is. The similarity used in the machine translation device 42 is a positive value, and the smaller the value, the more similar the example sentence and the input sentence 30 are.

  Since the example translation is used in the present embodiment, it is highly possible that a translation with high translation quality can be obtained if an example sentence very similar to the input sentence is found. If the example sentence does not resemble the input sentence, the translation quality is generally low. Therefore, this similarity is used as an index of the translation quality of the translation result.

  In this search, the machine translation device 42 sets the similarity between the input sentence 30 and the example sentence to an edit distance in units of words (deletion of a word necessary for converting the input sentence 30 into an example sentence, Calculate based on the number of insertions and substitutions). Accordingly, in the present embodiment, the similarity between matching sentences is zero. The determination signal 60 is at a logic 1 level when the similarity of the searched example sentence exceeds a prescribed standard, that is, when the searched example sentence and the input sentence 30 are not very similar, otherwise In this case, the logic level is zero. Furthermore, this edit distance is corrected by the semantic distance of the word. In other words, replacement of two words that are close in meaning is corrected so that the edit distance decreases as the semantic distance decreases. The machine translation system 32 includes a computer-readable thesaurus 44 for performing this correction process. In other words, replacement of two words that are in a semantically close relationship reduces the weight according to the magnitude of the difference between the levels of each word on the thesaurus.

  The machine translation system 32 is further connected to receive a computer-readable translatable sentence corpus 46 comprising a set of translatable sentences prepared in advance, an input sentence 30 and a decision signal 60. When the logic level is 1 (that is, when it is determined that translation by the machine translation device 42 is impossible), the similar sentence 62 which is the most similar to the input sentence 30 and whose similarity is smaller than a predetermined value is translated. And a similar sentence search unit 48 for outputting a search result signal 64 indicating whether or not the similar sentence 62 satisfying the above-described conditions has been searched. The search result signal 64 takes a logic 0 level when there is a sentence satisfying the above condition, and takes a logic 1 level when there is no sentence.

  The machine translation system 32 is further connected to the example corpus 40 and the Japanese thesaurus 44, and receives the determination signal 60 from the machine translation device 42, and the similar sentence 62 and the search result signal 64 from the similar sentence search unit 48, respectively. When the search result signal 64 is logic 0 level (that is, when a similar sentence 62 satisfying a predetermined condition is searched), the example corpus 40 and the Japanese thesaurus 44 are set in the same manner as the machine translation device 42. And a machine translation device 50 for performing an example translation on the similar sentence 62 and outputting a translation result 66 and a translation availability signal 68 indicating whether or not the translation is completed. The translation enable / disable signal 68 takes a logic 1 level when translation by the machine translation apparatus 50 is impossible, and takes a logic 0 level when translation is possible. Since there is a possibility that the translatable sentence corpus 46 contains a sentence that cannot be translated by the machine translation apparatus 50, the translation possibility signal 68 displays whether or not the translation by the machine translation apparatus 50 has actually been performed.

  The machine translation system 32 further includes an AND gate 54 having two inputs connected to receive the translation enable / disable signal 68 and the search result signal 64, the translation result 58 from the machine translation device 42, and the machine translation device 50, respectively. A selection unit for selecting the translation result 58 when the determination signal 60 is at a logic 0 level, and selecting the translation result 66 when the determination signal 60 is at a logic 0 level and outputting it as the output sentence 34. 52.

  FIG. 2 is a schematic diagram for explaining processing by the similar sentence search unit 48 shown in FIG. Referring to FIG. 2, generally, an input sentence to the machine translation system 32 includes a translatable sentence set 80 composed of sentences that can be used for example translation using the example corpus 40 (see FIG. 1), and cannot be translated. It is divided into a non-translatable sentence set 82 composed of sentences. When it is determined that the input sentence 30 cannot be translated by the machine translation device 42 shown in FIG. 1, that is, when the input sentence 30 belongs to the untranslatable sentence set 82, the similar sentence search technique 48 of the similar sentence search unit 48 performs in advance. A sentence similar to the input sentence 30 is searched from the prepared translatable sentence corpus 46. That is, the similar sentence search technique 84 replaces the non-translatable sentences 90, 92, 94, 96, 98, etc. with the translatable sentences 100, 102, or 104 similar to them (if any), thereby originally translating them. This is a technique that enables translation of an input sentence 30 that was an impossible sentence. As a result, the range of sentences that can be translated by the machine translation system 32 is expanded.

  FIG. 3 is a diagram showing the configuration of the similar sentence search unit 48 shown in FIG. Referring to FIG. 3, the similar sentence search unit 48 selects the input sentence 30 and all the sentences in the translatable sentence corpus 46 in order, and the sentence output from the selection unit 110. A morpheme analysis unit 112 for performing morpheme analysis, a computer-readable dictionary 114 for storing word information used when the morpheme analysis unit 112 performs morpheme analysis, and an analysis of the input sentence 30 by the morpheme analysis unit 112 The branching unit 116 for branching the result to the first output and the analysis result of each sentence of the translatable sentence corpus 46 to the second output are connected to the first output of the branching unit 116. Each sentence (hereinafter “candidate sentence”) of the translatable sentence corpus 46 is connected to an input sentence storage unit 118 for storing the morphological analysis result of the input sentence 30 output from 116 and the second output of the branching unit 116. Comprising a morphological analysis result, the similarity calculating unit 120 for calculating a similarity between the morphological analysis result of the input sentence 30, which is stored in the input sentence storage section 118 for called.) And.

  In the morpheme analysis performed by the morpheme analysis unit 112, the numeric string is generalized using a special symbol. Since nouns are given attribute information such as place names, organization names, and personal names, these nouns are also generalized using the attribute information. In order to determine that two words match in the similar sentence search process, it is necessary and sufficient that the basic form or attribute information of the words match and the parts of speech also match.

  In the present embodiment, the similarity calculation by the similarity calculation unit 120 is performed based on the ratio of the common part between the input sentence 30 and the candidate sentence to both the input sentence 30 and the candidate sentence. The higher the ratio of the common part to both sentences, the higher the similarity of the candidate sentence with respect to the input sentence 30. In this embodiment, an F value is used as the similarity. The F value is defined by the following equation (1).

F value = 2PR / (P + R) (1)
However, P (relevance rate) = number of words common to input sentence and candidate sentence / number of words of candidate sentence R (reproduction rate) = number of words common to input sentence and candidate sentence / number of words of input sentence Common between two sentences In general, three methods of n-gram, word string, and word set are often used as the definition of the portion. In this embodiment, the common part is defined using n-grams. This method will be described later.

  Referring to FIG. 3, in the present embodiment, the weight of the function word with respect to the content word is changed in calculating the number of common words. This weight is a positive value. It is desirable that this weight is variable. For this purpose, the similar sentence search unit 48 includes a function word weight storage unit 128 that is connected to the similarity calculation unit 120 and stores weights for the function words and gives them to the similarity calculation unit 120.

  The similar sentence search unit 48 further stores a similarity calculated between each sentence included in the translatable sentence corpus 46 and the input sentence 30 by the similarity calculation unit 120 together with a morphological analysis result. And the input sentence 30 based on the morphological analysis result of the input sentence 30 stored in the input sentence storage unit 118 and the morpheme analysis result of each sentence of the translatable sentence corpus 46 stored in the storage unit 122. Of the candidate sentence that has not been excluded by the exclusion processing unit 124 for performing the process of excluding the candidate sentence including the content word, the candidate sentence that is most similar to the input sentence 30, and (1) Among the content words included in the input sentence, the condition words that the content word not included in the candidate sentence is one word or less, or (2) the content words common to the input sentence are two or more words are satisfied. Things are similar 6 And a selection processing unit 126 for outputting as a search result signal 64 whether or not there is a similar sentence satisfying the above conditions, and a similar sentence search when the value of the determination signal 60 is a logic 1 level. And a sequence control unit 130 for controlling each functional unit of the unit 48 to operate to search the translatable sentence corpus 46 for a sentence most similar to the input sentence 30.

  In the similar sentence search by the similar sentence search unit 48, various conditions are imposed as described above, or a specific method is adopted. This is because there is a possibility that a favorable translation for the input sentence 30 can be finally obtained from similar sentences obtained by imposing the above-mentioned conditions and adopting the method as a result of actually performing various experiments. This is because it has been found that is high. Hereinafter, each condition will be described.

  In the following description, various evaluations are made for similar sentences. The evaluation criteria are shown in FIG. The evaluation criteria were determined from the viewpoint of how much to serve as a substitute sentence for the input sentence in the dialogue scene. As shown in FIG. 4, the evaluation rank has four levels of A1, A2, B1, and B2 in descending order of the degree of substitution. Of the retrieved sentences, sentences of rank A1 and A2 are appropriate as similar sentences, and sentences of B1 and B2 are inappropriate.

  In FIG. 4, “evaluation as a substitute sentence” represents an indication of a role to be played when a searched sentence is used as a substitute sentence of an input sentence. Appropriateness as a substitute sentence is mainly based on “semantic differences” and does not evaluate secondary information such as politeness.

  An evaluation example is shown in FIG. The candidate sentence of Example 1 represents the same meaning as the input sentence although there is a difference in expression, and the evaluation is A1. The candidate sentence of Example 2 is the main sentence part of the input sentence composed of compound sentences, and can be said to capture the main part. However, since the sub sentence indicating the cause of the main sentence request is missing, the evaluation is A2. The candidate sentence of Example 3 is a sub-sentence part of the input sentence composed of compound sentences, and does not represent the main part, so the evaluation is B1. Although the candidate sentence of Example 4 captures the main sentence, it is evaluated as B1 because an important object is missing. The candidate sentence of Example 5 adds a condition not included in the input sentence “Tomorrow is”. Such a condition is a misunderstanding that is serious and difficult to find in conversation. Therefore, the evaluation of Example 5 is B2. Since the candidate sentence of Example 6 is different from the input sentence at the basic level of modality, the evaluation is B2.

<Similarity calculation method>
As described above, the F value is used as a basic method of similarity calculation by the similarity calculator 120 of the similar sentence search unit 48. As a definition of the common part between the input sentence and the candidate sentence of the difference, there is one that uses an n-gram, a word string, or a word set. In this embodiment, n-grams are used. Hereinafter, a similarity calculation method using n-grams will be described.

  In the n-gram method, the similarity is calculated based on the n-gram common to the input sentence and the candidate sentence. In this calculation, each n-gram is weighted. In the present embodiment, the same formula as that adopted in the translated sentence automatic evaluation method called BLEU is adopted as the weight. That is, in this embodiment, for example, the precision P is calculated by the following equation.

ただしｐ_nは各ｎにおける適合率を表しており、下式で表される。

However p _n represents the suitability index for each n, represented by the following formula.

Count(x)は候補文ｘ中の頻度、Count_clip(x)は、入力文中のｘの頻度と候補文中のｘの頻度のいずれか少ない方を表す。再現率もこれと同様の考え方で算出される。

Count (x) represents the frequency in the candidate sentence x, and Count _clip (x) represents the lesser of the frequency of x in the input sentence or the frequency of x in the candidate sentence. The recall is also calculated using the same concept.

  In this embodiment, up to bigrams are used as n-grams. The size of n varies depending on the nature (number of constituent words) of the translation domain to be applied. In the case of travel conversations and the like, up to n = 2, and in the case of newspaper articles, it is considered that n = 4 is good.

  For comparison, as a similarity calculation method, a method using a word string as a common part definition (a method based on the longest common word string) will be described. In this method, the similarity is calculated using the longest common word string obtained by performing DP (Dynamic Programming) matching between the input sentence and the candidate sentence. In short, it is a method of extracting a common word in consideration of the word order.

  In the method using DP matching, a method using an edit distance is often used, but here, since the “similarity based on the common part” is the basis, the longest common word string is considered. The edit distance and the longest common word string have a complementary relationship, and the candidate sentence having the longest edit distance from the input sentence has the property that the longest common word string becomes shorter. In preliminary experiments, it has been found that there is almost no performance difference between the method based on the edit distance and the method based on the longest common word string.

  Of the three basic methods, a method based on a word set is a method in which a sentence is regarded as a word set and the number of words common to both the input sentence and the candidate sentence is a common part. This method corresponds to a case where n = 1 in the n-gram method.

  FIG. 6 shows an F value calculation example according to each basic method. In the word string method of FIG. 6, “Da” is excluded as a common word, but this is because the position of “Da” is greatly different between the input sentence and the candidate sentence. It was because it was not adopted.

<Exclusion of candidate sentences containing content words that are not in the input sentence>
In the exclusion processing unit 124, candidate sentences including content words not included in the input sentence are excluded without being adopted as similar sentences. This is because preliminary experiments have shown that such candidate sentences often do not substitute for input sentences. In many cases, candidate sentences including surplus content words further limit the meaning of the input sentence. In this case, there is a high risk of misunderstanding if the input sentence is replaced with the candidate sentence.

  FIG. 7 shows an example of a candidate sentence that becomes inappropriate by including surplus content words. In Example 1, the candidate sentence includes the content word “cash” and lacks “credit card”. In this example, the meanings of the input sentence and the candidate sentence are completely different. In example 2, the content word “seven o'clock” is added to the candidate sentence, which gives a serious constraint to the meaning of the input sentence. In the case of Example 3 as well, the content word “Chinese Chinese” added to the candidate sentence imposes inappropriate restrictions on the meaning of the input sentence.

  For each of the basic methods described above, similar sentences were searched and evaluated by a method (with surplus content words) and a method (with no surplus content words) in which candidate sentences including surplus content words are to be searched. Here, the content words are defined as nouns, verbs, adjectives, numbers, Roman letters, and the like, and the function words are defined as judgment words, particles, auxiliary verbs, conjunctions, adverbs, impression verbs, and the like. The sub-verb “suru” is considered to be a function word because it is considered to have little concrete meaning. The experimental results are shown in FIG.

  Referring to FIG. 8, no matter which method is used, an improvement of 8% can be seen in the search accuracy by imposing a restriction that no surplus content word is included.

<Weighting of content words and function words>
When measuring the similarity between two sentences using a common word for spoken language, it is considered that the value of the function word is low compared to the content word. The reason is that the spoken language has a large variety of function words, as indicated by the lack of particles and various end-of-sentence expressions. When function words representing the same meaning take various expressions, the magnitude of the degree of coincidence between the function words between the input sentence and the candidate sentence is not an effective index. Also, in a domain where conversation contents are roughly determined, such as travel conversations, the relations, case relations, and modification relations are often almost uniquely determined by the contained content words. For example, a sentence containing the content word (thief, me, purse, steal) can theoretically have various meanings, but it is actually a sentence other than "the thief stole my purse" Can hardly happen. That is, if the relationship is naturally limited by the content word set, the function word plays a smaller role.

  From the above, in calculating the similarity for searching for similar sentences, it is considered desirable to reduce the weight of the function word as compared with the weight for the content word. Actually, an experiment was conducted in which the retrieval accuracy was compared between the case where the content word weight is 1 and the function word weight is 1 which is the same as the content word, and 0.4. In the bigram of the n-gram system, the weight is set to 0.4 only when the two words constituting the bigram are both function words, and 1 is set otherwise. The experimental results are shown in FIG.

  Referring to FIG. 9, the search accuracy is improved by about 1 to 2% in any method. In short, reducing the weight of function words has an effect of giving priority to candidate sentences that share a lot of main information. An example in which this effect appears is shown in FIG.

  In FIG. 10, common words in search sentences are shown in bold. By reducing the weight of the function word, it is possible to output a sentence including all the main information although the sentence end portion is different.

  Through experiments, the highest accuracy rate was obtained when the n-gram method was adopted as the basic method, and the exclusion of candidate sentences including content words not included in the input sentence and the weight reduction of function words were adopted. The configuration of the present embodiment corresponds to that case. Note that the function word weight value (0.4 in this embodiment) in this process is stored in the function word weight storage unit 128 shown in FIG.

<Selection of candidate sentence by selection processing unit 126>
The translatable sentence corpus 46 used as a set of candidate sentences does not cover all utterances that appear as the input sentence 30. It is considered that there are many cases where the input sentence 30 that cannot be substituted by the sentence in the translatable sentence corpus 46 is given. Therefore, it is necessary to provide a condition for recognizing the retrieved sentence as a similar sentence and not select it as a similar sentence if the condition is not satisfied even if the sentence has a high degree of similarity. As described above, the selection processing unit 126 selects only candidate sentences satisfying one of the following two conditions as similar sentences. Hereinafter, the reason for adopting these conditions will be described.

(1) The number of content words missing in the candidate sentence compared to the input sentence is one word or less. (2) The number of words common to the input sentence and the candidate sentence is two words or more. Introduced a heuristic to exclude candidate sentences with words. Therefore, the content word set included in the retrieved candidate sentence is always a subset of the content word set of the input sentence 30. Considering the similarity based on the content word set of the retrieved candidate sentences, the highest degree of coincidence is when the content word sets of both sentences match. Then, the similarity decreases as the number of content words lacking in the retrieved sentence increases.

  FIG. 11 shows the relationship between the number of content words that are deficient compared to the input sentence and the accuracy rate of the retrieved candidate sentences. When the content word sets match between the input sentence and the retrieved candidate sentence, the correct answer rate is a high value of 89.1%. As the number of shortage content words increases, the correct answer rate decreases greatly. When the number of deficient content words is 2 or more, the correct answer rate is less than 50%, which is not sufficient accuracy. Therefore, the above condition (1) is imposed.

  Next, condition (2) will be considered. Considering from the viewpoint that the similar sentence may be a sentence that expresses the meaning of the input sentence, if the content words for the parts that express the meaning of the retrieved candidate sentence and the input sentence are common, It is considered sufficient that the content words are not common. Therefore, it is considered that a similar sentence can be determined based on the number of content words common to the candidate sentence and the input sentence (hereinafter, “number of common content words”).

  FIG. 12 shows the relationship between the number of common content words and the correct answer rate of the retrieved candidate sentences. From FIG. 12, as a large tendency, the correct answer rate improves as the number of common content words increases. However, the inclination is gentle. Under condition (1) regarding the number of content words that are insufficient, the accuracy rate of the boundary was 58.8%. Therefore, from FIG. 12, as a condition close to the correct answer rate, only candidate sentences with two or more common content words are selected as similar sentences. The correct answer rate when the number of common content words is 2 is 63.0%.

[Operation]
The machine translation system 32 according to the present embodiment whose configuration has been described above operates as follows. Referring to FIG. 1, it is assumed that an example corpus 40, a thesaurus 44, and a translatable sentence corpus 46 are prepared in advance. When the Japanese input sentence 30 is given, the machine translation device 42 refers to the thesaurus 44 to search for an example sentence having a Japanese sentence most similar to the input sentence 30 among the examples in the example corpus 40. In this search, an example sentence having the closest edit distance between the input sentence 30 and the Japanese sentence of each example sentence is searched. However, when calculating the edit distance, the word replacement is weighted by the semantic distance between the two words to be replaced, which is obtained by referring to the thesaurus 44. The edit distance calculated in this way becomes the similarity between the input sentence 30 and the example sentence.

  The machine translation device 42 selects the searched example sentence having the smallest similarity. At this time, if the similarity of the selected example sentence exceeds a predetermined value, that is, if the selected example sentence and the input sentence 30 are not very similar, the machine translation device 42 The value is a logic 1 level, otherwise it is a logic 0 level.

  If the degree of similarity between the selected example sentence and the input sentence 30 is equal to or less than a predetermined value, the machine translation device 42 determines the English part of the example sentence based on the difference between the input sentence 30 and the Japanese part of the example sentence. As a result, the translated sentence of the input sentence 30 is generated and the translation result 58 is output. The translation result 58 is given to the selection unit 52.

  When the value of the determination signal 60 is logic 0 level, the selection unit 52 selects the translation result 58 and outputs it as the output sentence 34.

  When the value of the determination signal 60 is logic 1 level, the following processing is performed. The similar sentence search unit 48 searches the translatable sentence corpus 46 for candidate sentences similar to the input sentence 30. That is, referring to FIG. 3, selection unit 110 first selects input sentence 30 according to the control of sequence control unit 130, and provides it to morphological analysis unit 112. The morpheme analysis unit 112 refers to the dictionary 114 to analyze the morpheme of the input sentence 30, decomposes it into a word string, and gives it to the branch unit 116. At this time, various information obtained by referring to the dictionary 114 is assigned to each word. The branching unit 116 provides the word string output from the morphological analysis unit 112 to the input sentence storage unit 118 according to the control of the sequence control unit 130. The input sentence storage unit 118 stores this word string.

  Next, the selection unit 110 reads the first Japanese part of the example sentences included in the translatable sentence corpus 46 and gives it to the morpheme analysis unit 112 under the control of the sequence control unit 130. The morpheme analysis unit 112 performs morpheme analysis on this Japanese part with reference to the dictionary 114, and gives the obtained word string to the branch unit 116. Also in this case, attribute information is given to each word. The branching unit 116 gives this word string to the similarity calculating unit 120 this time under the control of the sequence control unit 130.

  Based on the morphological analysis result of the example sentence given from the branching unit 116 and the morphological analysis result of the input sentence 30 stored in the input sentence storage unit 118, the similarity calculation unit 120 calculates the n-gram shown in the equation (1). According to the similarity calculation method used, the similarity between the input sentence 30 and the first candidate sentence of the translatable sentence corpus 46 is calculated and given to the storage unit 122. At this time, out of the number of common words between the input sentence 30 and the candidate sentence in the similarity calculation, the number of function words is multiplied by the value stored in the function word weight storage unit 128 as a weight. The storage unit 122 stores this similarity together with the morphological analysis result of the first candidate sentence.

  Thereafter, each candidate sentence stored in the translatable sentence corpus 46 is morphologically analyzed by the morpheme analyzer 112 under the control of the sequence controller 130, and the similarity with the input sentence 30 is calculated by the similarity calculator 120. Is done. The similarity of each candidate sentence obtained as a result is stored in the storage unit 122 together with the morphological analysis result of the candidate sentence.

  When the similarity is calculated for all candidate sentences, the exclusion processing unit 124 refers to the morphological analysis result of the input sentence 30 stored in the input sentence storage unit 118, and the content words that are not in the input sentence 30 among the candidate sentences Are excluded from the Japanese part, and other candidate sentences and similarities are given to the selection processing unit 126. The selection processing unit 126 includes (1) content words that are not included in the candidate sentence among the content words included in the input sentence among the given candidate sentences, or (2) common with the input sentence. A candidate sentence that satisfies the above-mentioned condition that the content word is two or more words and that is most similar to the input sentence 30 is output as the similar sentence 62, and a search is performed to determine whether or not a similar sentence that satisfies the above condition exists. The result signal 64 is output. The search result signal 64 takes a logic 0 level when there is a sentence satisfying the above condition, and takes a logic 1 level when there is no sentence.

  Referring to FIG. 1 again, the machine translation device 50 uses the example corpus 40 and the thesaurus 44 for the similar sentence 62 from the similar sentence search unit 48 when the search result signal 64 is at the logic 0 level. Translate. This example translation process is the same as that performed by the machine translation device 42. When the machine translation device 50 cannot retrieve an appropriate example sentence from the example corpus 40, it sets the signal 68 to logic 1 level and ends the process. When an appropriate example sentence can be searched from the example corpus 40, the machine translation device 50 corrects the English part of the example sentence based on the difference between the Japanese part of the example sentence and the similar sentence 62, thereby producing a similar sentence. 62 translations are performed. Then, the result of this translation process is given to the selection unit 52 as a translation result 66.

  When the determination signal 60 is at the logic 1 level, the selection unit 52 selects the translation result 66 given from the machine translation device 50 to the selection unit 52 in this way, and outputs it as the output sentence 34.

  As described above, the machine translation system 32 outputs the translation result as the output sentence 34 when the machine translation device 42 can machine translate the input sentence 30. When the input sentence 30 is a sentence that cannot be translated by the machine translation device 42, the untranslatable sentences 90, 92, 94, 96, 98 shown in FIG. 2 are replaced with the sentences 100, 102, 104, etc. in the translatable sentence corpus 46. Similarly to this, this input sentence 30 is replaced with any candidate sentence in the translatable sentence corpus 46. Since the translatable sentence corpus 46 is prepared by collecting sentences that can be translated in advance, the machine translation apparatus 50 has a high possibility of translating this candidate sentence. As a result, there is an effect that the range of sentences that can be translated by the machine translation system 32 is wider than that in the case where the similar sentence search unit 48 does not search for similar sentences.

  Note that there may be an input sentence 30 in which a similar sentence cannot be searched from the translatable sentence corpus 46 by the similar sentence search unit 48 as described above. In that case, the search result signal 64 becomes a logic 1 level and the translation enable / disable signal 36 becomes a logic 1 level.

  In addition, as long as the translatable sentence corpus 46 is composed of translatable sentences, if the similar sentence search unit 48 can search for similar sentences, it can be considered that translation by the machine translation device 50 is possible. However, since there may be a case where the contents of the translatable sentence corpus 46 are incomplete, the machine translation device 50 outputs the translation availability signal 68. That is, if the translation enable / disable signal 68 is at a logic 0 level, translation by the machine translation device 50 is possible, and if the translation enable / disable signal 68 is at a logic 1 level, translation by the machine translation device 50 is impossible. .

  Since the AND gate 54 takes the AND of the search result signal 64 and the translation enable / disable signal 68, it can be understood that the translation cannot be performed if the output AND gate 54 is at the logic 1 level, otherwise the translation is performed. It was found that was possible.

[Experimental result]
An experiment was conducted for Japanese-English translation using the similar sentence search unit 48 of the machine translation system 32 according to the embodiment described above. In this experiment, two types of corpus are used. An untranslatable sentence corpus, which is a set of untranslatable sentences, and a translatable sentence corpus 46 shown in FIG.

  Referring to FIG. 13, untranslatable sentence corpus 146 gives each sentence of second utterance corpus 140 referred to in the introduction of the embodiment of the invention to machine translation 142, and collects sentences that cannot be translated. It was created by that. The second utterance corpus 140 includes 1,698 sentences, among which the corpus 144 composed of translatable sentences is 1393 sentences, and the untranslatable sentence corpus 146 composed of sentences that cannot be translated is 305 sentences.

  The translatable sentence corpus 46 includes 70,671 sentences determined to be translatable by machine translation from the first basic expression corpus mentioned at the beginning of the description of the embodiment of the invention.

  When each sentence of the untranslatable sentence corpus 146 is given to the similar sentence search unit 48, the retrieved similar sentence corpus 150 including the obtained similar sentences is 164 sentences. That is, similar sentences could be searched for 164 untranslatable sentences. When the similarities of the retrieved similar sentences 164 were manually evaluated, it was found that 81 sentences were correct similar sentences as shown in the correct similar sentence 154 of FIG.

  Furthermore, the translated sentence obtained by giving the retrieved similar sentence 164 sentence to the machine translation apparatus 50 and the input sentence were presented to the evaluator and evaluated as a translated sentence. The translated sentences are evaluated by English native speakers in four ranks: Good, Fair, Acceptable, and Bad. Among these, the evaluation sentence of Good, Fair, and Acceptable is set as “appropriate translation”. This evaluation criterion is determined for evaluating the translation quality of machine translation, and is different from the similar sentence evaluation criterion shown in FIG.

  As shown in FIG. 13, 61 sentences were obtained as the correct translation 156 as a result. The breakdown is 12 sentences for Good, 10 sentences for Fair, and 39 sentences for Acceptable.

  FIG. 14 shows the correct answer rate for similar sentences and translated sentences, and the relief ratio for untranslatable sentences, respectively. Referring to FIG. 14, an appropriate similar sentence can be searched by similar sentence search for 26.6% of similar sentences, that is, untranslatable sentences at the same language stage as the input sentence. In addition, it was possible to obtain appropriate translations for 20% of untranslatable sentences even at the translation stage.

  As described above, according to the machine translation system 32 according to the present embodiment, an existing machine translation can be translated by combining a similar sentence search technique using a language resource called a monolingual corpus with easy translation. The range of possible sentences can be expanded. Without the time-consuming task of setting rules for pre-editing, it is possible to increase the translatability of input sentences that have many variations in the same meaning as spoken language.

  In the above-described embodiment, a method using n-grams common to the input sentence and the candidate sentence is adopted as the similarity calculation method in the similar sentence search. However, the present invention is not limited to such a method. For example, a similarity calculation method based on the longest common word string between the input sentence and the candidate sentence may be used, or a similarity calculation method based on the common part of the word set of the input sentence and the candidate sentence may be used. Good. Also, other similarity calculation methods may be used, but even in that case, it is desirable to employ a similarity that effectively indicates how similar the input sentence and the candidate sentence are in terms of content.

  In the above embodiment, in the similarity calculation method based on n-grams, the weight of the function word when the weight of the content word is 1 is 0.4. However, this weight is not limited to such a value, and any weight may be given to the function word as long as it is less than the weight of the content word.

  Furthermore, the value of each parameter described above can vary depending on the target language and the target domain. It is desirable to determine them based on experiments conducted in accordance with the environment where the present invention is actually implemented.

  In the above-described embodiment, the example translation is used as the machine translation device. Then, whether or not an example sentence very similar to the input sentence is obtained in the process of example translation is used as an index of translation quality. In this case, translation quality can be evaluated in the process of example translation. However, the present invention is not limited to such an embodiment. For example, an arbitrary machine translation device may be used, the translation quality of the translation to be output may be evaluated according to some criteria, and the result may be used as an index for determining whether or not to search for a similar sentence. For example, translation quality may be evaluated based on a comparison result with a plurality of reference translations prepared in advance, or a translation may be evaluated using a language model and / or a translation model. In this case, they can be provided in the machine translation system as function modules independent of the machine translation apparatus. In other words, an independent function module for evaluating the translation quality is not required when a translation quality index equivalent to that in the translation process is obtained as in the first machine translation apparatus in the above-described embodiment. It is.

[Realization by computer]
The machine translation system 32 according to the above-described embodiment can be realized by a computer system. FIG. 15 is an external view showing the overall configuration of a computer system 250 that implements the machine translation system 32 according to the above-described embodiment. The system 250 includes a pair of a microphone 264 and a speaker 278, a computer 260 having a CD-ROM (Compact Disc Read-Only Memory) drive 270 and an FD (Flexible Disk) drive 272, and a monitor 262 connected to the computer 260. A keyboard 266 and a mouse 268 are included.

  The microphone 264 and the speaker 278 are used for input and output of speech translation if necessary, and do not constitute a part of the present invention. Accordingly, the details of the portion of the system relating to the microphone 264 and speaker 278 will not be described here.

  FIG. 16 is a hardware block diagram of the computer 260. Referring to FIG. 16, a computer 260 includes a CPU (Central Processing Unit) 340, a bus 342 connected to the CPU 340, a read only memory (ROM) 344 connected to the bus 342, and a bus 342. A CD-ROM drive 270 that is loaded with a random access memory (RAM) 346 connected to the CD-ROM, a hard disk 348 connected to the bus 342, and a CD-ROM (compact disk ROM) 360 and reads data from the CD-ROM. An FD (flexible disk) 362 is mounted, reads data from the FD, writes data, a sound board 350 to which the microphone 264 and the speaker 278 are connected, and a local area network connected to the bus 342. Network board 352 that provides a function of connecting to a data communication network such as a network (LAN).

  The machine translation system 32 according to the embodiment described with reference to FIGS. 1 to 14 is stored in the hardware of the computer system 250, the computer program executed thereon, and the hard disk 348, the RAM 346, and the like of the computer system 250. It can be realized by data such as various corpora. The configuration of the computer program will be described later. Such computer programs and data such as a corpus (hereinafter referred to as “programs”) are stored in a storage medium such as a CD-ROM 360 for distribution. Those programs and the like are read into the hard disk 348 from such a storage medium. When the system is started up, the program is read from the hard disk 348 and loaded into the RAM 346 and read and executed by the CPU 340. A program read address is designated by a program counter (not shown). The contents of the program counter are rewritten as the program is executed. Data read and write addresses are specified by the calculation results according to the program.

  FIG. 17 is a flowchart of a program that implements the machine translation system 32 according to the above-described embodiment. Details of the contents of each step are as described with reference to FIGS. Here, a preferable configuration of the entire program when the machine translation system 32 is realized by the program is shown.

  Referring to FIG. 17, in step 400, machine translation is executed on the input sentence. This machine translation is based on the example translation, and the similarity of the example sentence used as the basis for the example translation is obtained together with the translation result. In step 402, it is determined whether or not the similarity value is greater than a predetermined value. As described above, in the example translation used in this embodiment, the similarity is smaller as the two sentences are more similar, and the similarity is 0 when the two sentences are completely matched. If they are similar (that is, if the condition of similarity> predetermined value is not satisfied), the control proceeds to step 414, and the translated sentence obtained in step 400 is output as the entire translation result and processed. finish. If the two are not similar, control proceeds to step 404.

  In step 404, it is determined whether the result of the machine translation process is obtained by a machine translation process for the input sentence or obtained by a machine translation process after replacing the input sentence with a similar sentence. If it is after the replacement, the control proceeds to step 410, an output (display) indicating that the translation is impossible is performed in step 410, and the process is terminated. If it is before replacement, control proceeds to step 406.

  In step 406, a process for retrieving a similar sentence for the input sentence from the translatable sentence corpus is performed. The conditions to be satisfied by the similar sentence searched here are as described above. Thereafter, in step 408, it is determined whether or not a translatable sentence (similar sentence) satisfying the condition of being similar to the input sentence in step 406 has been searched. If it is determined here that there is no similar sentence, it is displayed that translation is impossible in step 410 and the process is terminated. If there is a similar sentence, the process proceeds to step 412.

  In step 412, the input sentence is replaced with the searched similar sentence. Control returns to step 400. Thereafter, the processes of steps 400, 402, and 414 are executed for the replaced sentence, or the processes of steps 400, 402, 404, and 410 are executed, and the entire machine translation process is completed.

  As described above, FIG. 1 shows the machine translation device 42 and the machine translation device 50 as separate units, but these may be realized by the same unit. Thus, by realizing the machine translation device 42 and the machine translation device 50 with the same thing, it is possible to widen the range of sentences that can be translated while preventing an increase in resources necessary for translation. Of course, both may be different. When the two are separate, the machine translation principles of both may or may not be the same.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim in the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

It is a block diagram of the machine translation system 32 concerning one embodiment of the present invention. It is a figure which shows the concept of the similar sentence search technique in the machine translation system. 4 is a block diagram of a similar sentence search unit 48 of the machine translation system 32. FIG. It is a figure explaining the evaluation criteria at the time of evaluating a similar sentence. It is a figure which shows the example of evaluation of the candidate of various similar sentences with respect to an input sentence. It is a figure which shows the example of F value calculation by three basic methods of similarity calculation. It is a figure which shows the example of the candidate sentence which becomes inappropriate by including a surplus content word. It is a figure which shows the influence of presence of a surplus content word in three basic methods of similarity calculation. It is a figure which shows the influence of the weight of a function word in tabular form in three basic methods of similarity calculation. It is a figure which shows the influence on the search result by reducing the weight of a function word in a table form. It is a figure which shows the relationship between the number of the content words which are in short compared with an input sentence in the similar sentence searched, and the correct answer rate. It is a figure which shows the relationship between the number of common content words with an input sentence, and the correct answer rate of the searched candidate sentence. It is a figure which shows the corpus used in experiment, and an experimental result. It is a figure which respectively shows the correct rate in the similar sentence obtained in experiment, and the relief rate of the untranslatable sentence in a translation sentence. It is an external view of the computer system 250 which implement | achieves the machine translation system by one embodiment of this invention. It is a hardware block diagram of the computer 260 shown in FIG. It is a flowchart of the program for implement | achieving the machine translation system which concerns on one embodiment of this invention with a computer system.

Explanation of symbols

  32 machine translation system, 36, 68 translation enable / disable signal, 40 example corpus, 42 machine translation device, 44 thesaurus, 46 translatable sentence corpus, 48 similar sentence search unit, 50 machine translation device, 52, 110 selection unit, 54 AND gate 58 translation result, 60 judgment signal, 62 similar sentence, 64 search result signal, 66 translation result, 80 translatable sentence set, 82 untranslatable sentence set, 84 similar sentence search technology, 112 morpheme analysis unit, 114 dictionary, 116 branch 118, input sentence storage unit, 120 similarity calculation unit, 122 storage unit, 124 exclusion processing unit, 126 selection processing unit, 128 function word weight storage unit, 130 sequence control unit

Claims (7)

A machine translation system for translating an input sentence in a first language into a second language,
A first corpus including a plurality of sentences in the first language, preselected as being translatable into the second language;
First machine translation means for translating the input sentence into the second language;
Index output means for outputting an index of the translation quality of the translation by the first machine translation means;
To search the first corpus for a sentence having a predetermined relationship with the input sentence in response to the index output by the index output means meeting a predetermined condition indicating that the translation quality is poor. Search means,
A machine translation system comprising: second machine translation means for translating the sentence retrieved by the search means into the second language. In response to the fact that the index output by the index output means matches the predetermined condition, the search means is the most based on the similarity calculated according to a predetermined calculation method with the input sentence. The machine translation system according to claim 1, further comprising a similar sentence search unit for searching the first corpus for a sentence determined to be similar to the input sentence. The similar sentence search means includes:
Similarity calculating means for calculating a similarity defined based on a common part between each of the plurality of sentences included in the first corpus and the input sentence;
And means for extracting, from the first corpus, a sentence determined to be most similar to the input sentence based on the similarity calculated by the similarity calculation means. Machine translation system. The similarity calculation means includes:
A word number calculating means for calculating the number of words of the input sentence, the number of words of the candidate sentence, and the number of words common to the input sentence and the candidate sentence;
The similarity between the input sentence and the candidate sentence for which the similarity is to be calculated is expressed by the following formula: relevance ratio = (number of words common to the input sentence and candidate sentence) / number of words in the candidate sentence reproduction ratio = 4. The machine translation system according to claim 3, comprising means for calculating as a function of both the precision and the recall defined by (number of words common to input sentence and candidate sentence) / number of words of input sentence. . The word number calculation means calculates the number of words in the input sentence, the number of words in the candidate sentence, and the number of words common to the input sentence and the candidate sentence by multiplying by a predetermined weight according to the type of each word. The machine translation system according to claim 4, comprising means for: The first machine translation means includes:
A bilingual example corpus that includes a plurality of examples of sentences of the first language and sentences of the second language that are good translations of each other;
Means for retrieving an example from the bilingual example corpus that includes a sentence in the first language that is most similar to the input sentence according to a predetermined similarity criterion;
The sentence of the first language searched by the means for searching and the input of the sentence of the second language that is a parallel translation of the sentence of the first language searched by the means for searching And example translation means for translating the input sentence by correcting based on the difference between sentences,
The index output means determines whether or not a predetermined similarity defined between the sentence in the first language searched for by the searching means and the input sentence satisfies a predetermined criterion. The machine translation system according to any one of claims 1 to 5, further comprising: means for determining a determination result and outputting a determination result as the index. A computer program that, when executed by a computer, causes the computer to operate as the machine translation system according to any one of claims 1 to 6.

Images