JP2013186673A - Machine translation device and machine translation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To realize translation with high accuracy.SOLUTION: A machine translation device which performs machine translation from a source language to a target language using examples includes example-based machine translation means for performing example-based translation using prestored translation example data on a clause or phrase basis in the target language corresponding to the source language for each word contained in input data of the source language, and partial translation synthesizing means for synthesizing pieces of partial translation of a word or a plurality of words translated by the example-based machine translation means.

Description

  The present invention relates to a machine translation apparatus and a machine translation program, and more particularly, to a machine translation apparatus and a machine translation program for realizing highly accurate translation.

  Conventionally, a method of machine-translating the source language into the target language using examples, or a method of statistically obtaining the word that has the highest probability of following a certain word, and machine-translating the source language into the target language using that information Etc. are known (see, for example, Patent Documents 1 to 5).

  For example, the technique disclosed in Patent Document 1 includes an example translation candidate in which an input sentence is translated into a target language based on an example of a target language corresponding to an example of a source language that matches or is similar to the input sentence, and an example translation The first likelihood representing the likelihood of the candidate is obtained, the input sentence is translated into the target language by a process different from the above process, and the second likelihood is the first of the translation result candidates for each word of the input sentence. A translation candidate representing a candidate that is equal to or greater than the threshold is generated, and for each example translation candidate, if the translation included in the example translation candidate does not exist in the translation candidate, the first likelihood is decreased by a predetermined value, A machine translation apparatus for selecting an example translation candidate having the maximum one likelihood is shown.

  Further, the technique disclosed in Patent Document 2 matches a source language (SL) input fragment with an example SL fragment in the example base, and matches all matched blocks in the SL input to the one in the example. Identify as a term block in the SL input that matches one or more SL fragments, select a block combination of the matched blocks to cover one or more fragments of the SL input, and each in the selected block combination For a block, the example associated with the block was identified, and the target language (TL) portion of the identified example was aligned and aligned with the SL portion of the identified example that matched one or more fragments of the SL input An example-based machine translation system that provides translation output based on parts is shown.

  Further, the technique disclosed in Patent Document 3 creates a plurality of subtree groups from a syntax tree of a text in a source language, and for each subtree included in the plurality of subtree groups, the corresponding part from the example database. Search for a group of examples that have a source language syntax tree that matches the tree, and the translation probabilities for each of the examples included in the retrieved group of examples in the partial examples group that have a context similarity greater than or equal to the context similarity of each example In this case, the maximum likelihood subtree group is selected based on the translation probability, and the selected subtree group and examples searched for the subtrees included in the subtree group are selected. Based on this, a machine translation device for generating text in a target language is shown.

  Further, the technique disclosed in Patent Document 4 acquires an input sentence to be translated, and uses the phrase fuzzy matching technique from a phrase table constructed in advance, and the same or most similar parallel phrase pair for each phrase in the input sentence. To obtain the exact translation of each phrase, and correct the target language for the input sentence based on the above-mentioned parallel translation phrase pair and a pre-built language model. 1 shows a phrase-based statistical machine translation method that detects all translated sentences of the above, and selects and outputs a translated sentence having the highest score as a correct target language translated sentence of an input sentence using a statistical model.

  Further, the technique disclosed in Patent Document 5 includes means for determining a vector of probabilities representing class membership of the source sentence, the element of the vector being one of a set of classes in which the source sentence is predetermined. And further comprising a plurality of class specific statistical sub-decoders provided for each of the classes of the predetermined set of classes, the decoder statistically depending on each set of training data of each class. Each of the decoders outputs, for each word or word sequence in the source sentence, the probability of the translated word or word sequence in the target language, and according to the probability of the possible word sequence in the target language, the target language of the source sentence Has a means to estimate the most likely translation hypothesis in The probability of a possible word sequence in the target language is calculated by a statistical machine translation device calculated by interpolating the probabilities output by a plurality of subdecoders according to the probability vector for each of the target language words or word sequences. Has been.

JP 2008-305167 A JP 2008-262587 A JP 2006-252290 A JP 2010-061645 A JP 2009-294747 A

  However, the above-described conventional methods, such as example-based translation and statistical machine translation, are directed to the whole sentence, and there is no one that targets only a partial word string such as a clause or a phrase. In other words, with the conventional method, when example-based translation and statistical machine translation are performed on a sentence-by-sentence basis, the translation result is the entire sentence, and the translation result for partial word strings cannot be merged. there were. In addition, using conventional methods, instead of sentences, it is possible to input clauses or phrase units, use example-based translation or statistical machine translation, and integrate the translation results as one sentence. Since it is unclear what length should be used for input, high-precision translation could not be realized.

  The present invention has been made in view of the above-described problems, and an object thereof is to provide a machine translation device and a machine translation program for realizing highly accurate translation.

  In order to solve the above problems, the present invention employs means for solving the problems having the following characteristics.

  The present invention provides a machine translation device for machine-translating a source language into a target language using an example, and for each word included in the input data of the source language, a clause of the target language corresponding to the source language stored in advance Example-based machine translation means for performing example-based translation using phrase-based translation example data, and partial translation synthesis means for synthesizing partial translations for one or more words translated by the example-based machine translation means It is characterized by.

  The present invention also provides a machine translation program for causing a computer to function as each unit included in the machine translation apparatus described above.

  In addition, what applied the component, expression, or arbitrary combination of the component of this invention to a method, an apparatus, a system, a computer program, a recording medium, a data structure, etc. is also effective as an aspect of this invention.

  According to the present invention, highly accurate translation can be realized.

It is a figure which shows an example of a function structure of the machine translation apparatus in this embodiment. It is a flowchart which shows an example of the machine translation process sequence in this embodiment. It is a figure for demonstrating the example translation example using a CYK table. It is a flowchart which shows an example of the example translation procedure using a CYK table. It is a figure which shows the specific example of the example translation technique using a CYK table. It is a figure which shows an example of the example data for translation containing appearance frequency. It is a figure which shows an example of the production | generation screen of a sign language CG translation system.

<About the present invention>
The present invention efficiently translates a word string that can be translated using, for example, a long example in an input sentence, and translates a word string that could not be translated by statistical machine translation. Example-based translation and statistical machine translation By fusing the translation results of each, high-precision translation is realized. Hereinafter, preferred embodiments of a machine translation apparatus and a machine translation program according to the present invention having the above-described features will be described in detail with reference to the drawings.

  In recent years, not only translation between multilingual languages but also machine translation from primitive language to sign language has attracted attention. Sign language is an important means of communication for the hearing impaired and the like. In particular, sign language is a native language for deaf persons who have been deaf or congenital, and are easier to understand than Japanese. However, there are few information provided by sign language, and some studies have been conducted on automatic translation from Japanese to sign language. However, the scale of the vocabulary handled there is small, and there are many cases where it is an experimental system. Therefore, in the following description, an example of translating from Japanese into sign language will be described as an example of machine translation, but the present invention is not limited to this. Further, in the following description, translating into sign language means converting a sign language operation into a transcript of each word, but the present invention is not limited to this.

<Basic concept of language translation>
First, the basic concept of language translation will be explained. Language translation refers to conversion from a character string to a character string. Language translation methods are roughly classified into rule translation (rule-based translation), example translation (example-based translation), and statistical translation (statistical machine translation).

  Here, since sign language is a visual language, it is ultimately necessary to convert it into a motion video, but it is difficult to directly convert a Japanese single character string into a sign language video. For this reason, in the present embodiment, a Japanese input sentence that is a source language is converted into a sign language word string (a sign language action transcribed into a word string). Then, a sign language image corresponding to each sign language word is extracted from the converted sign language word string, and the extracted sign language images are connected to output a sign language image corresponding to the Japanese input sentence. The sign language video may be, for example, a motion video by CG (Computer Graphics) or a motion video by live action.

  In addition, rule translation requires translation knowledge constructed manually. The construction of translation knowledge requires linguistic knowledge of the translation target, but the sign language (Japanese sign language) has not been fully elucidated until machine translation can be developed. Therefore, at present, it is not realistic to develop a translation system at an early stage using only rule translation. On the other hand, statistical translation has the advantage that knowledge of the target language is not required. However, statistical translation requires a large-scale bilingual corpus and currently has a scale of 1 to 10 million sentences. Current Japanese-Japanese sign language bilingual corpora are not so large, and relatively large bilingual corpora are much smaller, about 20,000 to 30,000 sentences. In that respect, the example translation, which is the same corpus-based translation method, directly uses the parallel translation, so that it is possible to obtain a certain degree of translation accuracy even if it is not as large as the statistical translation.

  Furthermore, for example, if the target of translation is limited to specific information (for example, weather information), the language phenomenon that appears there is limited to some extent. In particular, in terms of sections and phrases, meteorological information has many typical expressions. Note that example translation, unlike statistical translation, requires linguistic knowledge of the translation target, but it may not be as detailed as regular translation.

  Therefore, in the present embodiment, a method that mainly uses example translation and also uses statistical translation is adopted. Specifically, first, example translation based on perfect matching is performed in units of clauses (phrases). Since this translation is completely identical to the example, there are some clauses (phrases) that do not match the example. In that case, translation is performed by statistical translation.

<Machine translation device: functional configuration example>
Here, a functional configuration example of the machine translation apparatus according to the present embodiment will be described with reference to the drawings. FIG. 1 is a diagram illustrating an example of a functional configuration of a machine translation apparatus according to the present embodiment. The machine translation apparatus 10 shown in FIG. 1 includes a morpheme analysis unit 11, an example base machine translation unit 12, a clause / phrase unit translation example storage unit 13, an example base partial translation storage unit 14, and an untranslated word storage unit 15. And statistical machine translation means 16 and partial translation synthesis means 17.

  The morpheme analyzing means 11 inputs the document data to be translated that the user wants to translate in sign language, analyzes the morpheme included in the input sentence, and divides the word unit included in the input sentence from the analysis result. Here, the morphological analysis means to divide the input text data into meaningful words. For example, the part of speech or the content may be determined using a preset corpus or the like. It is not limited to this.

  In the present embodiment, tea bowl (http://chasen.naist.jp/) or the like can be used as an example of morphological analysis, but the present invention is not limited to this. Further, the morpheme analyzing unit 11 outputs the word string of the sentence obtained from the analysis result to the example-based machine translation unit 12.

  The example-based machine translation unit 12 performs example-based translation on the word string obtained by the morpheme analysis unit 11 using the translation examples stored in advance in the clause / phrase unit translation example storage unit 13. The example-based translation method in this embodiment will be described later. The example-based machine translation unit 12 outputs a translation result for a word that has been example-based translated to the example-based partial translation storage unit 14. Further, the example-based machine translation unit 12 outputs the word string to the untranslated word storage unit 15 if there is a word string that could not be translated based on the example.

  The clause / phrase unit translation example storage means 13 stores preset clause or phrase unit translation examples. Moreover, the example base partial translation storage means 14 stores the translation result (for example, translated sign language word string) in the example base machine translation means 12. The untranslated word storage unit 15 stores words that could not be translated by the example-based machine translation unit 12.

  The statistical machine translation means 16 performs statistical machine translation using a translation dictionary or the like that has been machine-learned based on the translation examples stored in the clause / phrase unit translation example storage means 13. Note that the translation dictionary is not limited to this, and a translation dictionary or the like statistically generated by general machine learning can be used. The statistical machine translation unit 16 outputs the statistical machine translation result to the partial translation synthesis unit 17.

  The partial translation synthesis unit 17 translates the partial word string obtained by the example-based partial translation storage unit 14 (sign language word string) and the translation result of the partial word string obtained by the statistical machine translation unit 16 (sign language word string). Are combined in the order corresponding to the word string of the input sentence (word order), and the combined translation result is output as output data. Thereby, in this embodiment, highly accurate translation is realizable.

<Machine translation processing procedure in this embodiment>
Here, the machine translation processing procedure in the present embodiment will be described using a flowchart. FIG. 2 is a flowchart showing an example of a machine translation processing procedure in the present embodiment.

  In FIG. 2, in the machine translation process, when source language data (for example, Japanese sentence data) is input (S01), the input sentence data is subjected to morphological analysis and divided into predetermined words. (S02).

  Next, the machine translation process performs example-based machine translation for each divided word (S03), and determines whether there is an untranslated word that could not be translated among the divided words (S04). . Here, if there is an untranslated word (NO in S04), the machine translation process performs statistical machine translation (S05).

  Machine translation processing is performed in S04 when there is no untranslated word (YES in S04) or after completion of processing in S05, partially translated words (for example, sign language words) are synthesized in correspondence with the order of input data. Then, the synthesized result is output as a final translation result (for example, a sign language word string) (S08). The output data of S08 outputs the above-mentioned sign language word string as a translation result. However, the present invention is not limited to this. For example, motion video data corresponding to a sign language word is prepared in advance. The sign language image may be generated by connecting the motion image data corresponding to the sign language word, and the generated sign language image may be output.

  Further, the machine translation process determines whether or not to end the process (S09). If the process is not ended (NO in S09), the process returns to S01 to input other data and perform the above-described process. . Further, the machine translation process is terminated when the process is terminated due to an end instruction from the user or the like or when there is no data to be input (YES in S09).

<Specific example>
Here, a specific example of machine translation in the present embodiment will be described. For example, consider translating Japanese sentence data “Kyushu and Okinawa are likely to be thunderstorms in the evening”. It is assumed that the data of [Example 1] shown below is stored in advance in the clause / phrase unit translation example storage means 13.
[Example 1] (section / phrase unit)
“Kyushu and” “Kyushu N”
“Okinawa” ⇔ “Okinawa”
"Kyushu and Okinawa are" ⇔ "Kyushu N Okinawa"
“It seems to be a thunderstorm” ⇔ “Thunderstorm dream N”
“Thunderstorm” ⇔ “Thunderstorm”
“Yes” ⇔ “Dream N”
Here, the left side of the arrow (⇔) indicates Japanese, and the right side indicates the sign language. In addition, as an example, the sign language is “Japanese-Sign Language Dictionary” published by the All Japan Deaf Federation (Akihiko Yonekawa (supervised), Japan Sign Language Research Institute (edition), All Japan Samurai Federation Publishing Bureau, 2006. ), But the notation is not limited to this. In addition, “N” included in the sign language notation indicates “whispering” which is a non-finger movement, and indicates a break of a meaningful unit.

  The machine translation apparatus 10 according to the present embodiment first receives a morpheme analysis unit 11 for an input sentence “Kyushu and Okinawa are likely to be thunderstorms in the evening” (hereinafter referred to as “Japanese sentence 1” if necessary). Analysis is performed and divided into predetermined words. Specifically, Japanese sentence 1 is divided into 11 words “Kyushu / and / Okinawa / ha / in the evening / from / thunderstorm / and / be / so / is” according to the result of morphological analysis (“/ "Indicates a word break).

  Next, the example-based machine translation means 12 performs example translation. That is, the example of clauses / phrases is applied in the input sentence, and the example is translated at the part that can be translated. For example, if the above-mentioned [Example 1] is applied to the Japanese sentence 1 described above, the Japanese sections “Kyushu and Okinawa are likely to be thunderstorms” and “Kyushu N Okinawa” and “Thunderstorm Yume N” respectively. And can be translated.

  However, “from the evening” that is not in the example cannot be translated. Therefore, in this embodiment, a word string that could not be translated as an example is translated by statistical translation. In addition, in the translation of “Japanese-sign language”, since the word order is not changed, it can be expected that the translation can be performed with high accuracy even by statistical translation.

<Specific example of example translation in example-based machine translation means 12>
Next, as a specific example of the example translation in the example-based machine translation means 12, an example translation example and a translation technique will be described. Examples are usually divided into sections and phrases by hand. When dividing, consider breaks in Japanese-side clauses and phrases, and breaks in semantic units on the sign language side. Also, the break on the sign language side is determined mainly by “whisper (N)”. If it is difficult to determine the cut on the sign language side, do not cut it forcibly.

Further, in the example translation, the example can be expected to be translated with high accuracy when the data is as long as possible. Therefore, in this embodiment, all the combinations are automatically generated from the clauses (phrases) divided in one sentence and added to the example. For example, in the input sentence “Today will be a thunderstorm” (hereinafter referred to as “Japanese sentence 2” if necessary), when manually divided into clauses and phrases, [Example 2] is as follows: Suppose that it was obtained. When all combinations of [Example 2] are automatically generated, [Example 3] shown below is obtained. Therefore, in this embodiment, all of them are added to the examples.
[Example 2] (Manual section (phrase) division)
“Today” ⇔ “N today”
“Thunderstorm” ⇔ “Thunderstorm”
“Well” ⇔ “Dream N”
[Example 3] (Additional example)
“Today is a thunderstorm” ⇔ “Today N thunderstorm”
"It will be a thunderstorm" ⇔ "Thunderstorm dream N"
Here, when [Example 2] and [Example 3] are used, the example translation result of the Japanese sentence 2 is “today N thunderstorm dream N”. In the present embodiment, by using the above-described method, for example, about 10,000 clauses (phrases) can be obtained from about 3,500 parallel translations of weather information.

  Here, the example translation by the example-based machine translation unit 12 in the present embodiment can be performed by, for example, filling a CYK (Cocke-Younger-Kasami) table, but is not limited thereto.

<Example translation using CYK table>
Here, an example translation example using the CYK table will be described. FIG. 3 is a diagram for explaining an example translation example using the CYK table. FIG. 4 is a flowchart showing an example of an example translation procedure using the CYK table. Note that the arrangement of the CYK table 20 shown in FIG. 3 is an example, and the present invention is not limited to this. Further, the example of FIG. 4 also shows a part in which example-based translation is performed using an example of translation in section / phrase units, and statistical machine translation is performed on an untranslated word.

For example, as a result of morphological analysis of certain input data, it is assumed that “w ₁ , w ₂ ,..., W _i ,..., W _j , ..., w _n ” (where “w” indicates each word, “,” Indicates a delimiter for each word, but is treated as a single sentence (word string) without considering “,” as necessary). In this example, it is assumed that syntax analysis is not performed.

In the example translation process, first, an example that matches the input sentence is found. The input word string “w _i ,..., W _j ” is collated with a bilingual corpus (example) for each clause / phrase, and if it matches, the sign language word string is stored in the storage area of the CYK table 20. Store in t (j−i + 1, j). That is, as shown in FIG. 3A, the translation result for the word string “w _i ,..., W _j ” is stored in the storage area t (j−1 + 1, j) of the CYK table 20.

If there are a plurality of translation results, all of them are registered in the CYK table 20. That is, as shown in FIG. 3B, the translation result of “w _i ,..., W _{j + k−1} ” is stored in the corresponding storage area of the CYK table 20 as “W _{i1 ,.} The translation result of “w _{i + k} ,..., W _j ” is stored in the corresponding storage area of the CYK table 20 as “W _i2 ,..., W _j2 ”. However, at the time of registration, the translation result registered so far is checked and the same is not registered.

In the translation processing in the example-based machine translation means 12, the translation result of a certain Japanese word string “w _i ,..., W _j ” can be obtained, for example, by concatenating the translation results of two consecutive partial word strings. Can do. Therefore, in this embodiment, as described above, when there are a plurality of partial word string translation results, they are concatenated and registered in the corresponding storage area of the CYK table 20. For example, _"w i, ..., _{w j"} of the translation result (t (j-i + 1 , j)) is shown in FIG. 3 (b) _{"w i, ..., w i +} k-1 " of the translation result (t ( k, i + k−1)) and the translation result (t (j−i + 1−k, j)) of “w _{i + k} ,..., w _j ” are both “empty”. Can be obtained by concatenating them. Note that the combinations to be linked are partial word strings corresponding to each other in sequence, such as two storage areas shown in (1) of FIG. 3B and two storage areas shown in (2). Generate translations.

  Further, in the translation processing in the example-based machine translation means 12, the translation result of the entire sentence is registered in t (n, n). If the content stored in the area of t (n, n) is “empty”, it means that the example translation as a whole sentence could not be translated. In that case, a partial translation is requested. In order to obtain a partial translation, a portion that is not “empty” in the CYK table 20 may be found. In that case, as shown in FIG.3 (c), a longer partial translation can be obtained by examining sequentially from the upper table.

Here, when focusing on a certain word string “w _p ,..., W _i ,..., W _j ,..., W _q ”, if t (j−i + 1, j) is not “empty”, This is a translation of the partial word string “w _i ,..., W _j ”. In this case, the word strings “w _p ,..., W _i−1 ” and “w _{j + 1} ,..., W _q ” before and after that are word strings that could not be translated by the example translation. By repeating the same processing as described above for each of these two word strings, it is possible to obtain a word string that has not been converted into an example-translated word string as shown in FIG.

Here, the above-described processing contents will be specifically described using the translation processing procedure of FIG. First, the translation processing procedure is divided into words “w ₁ ,..., W _n ” by the morpheme analysis means 11 when source language sentence data (for example, Japanese sentences) is input (S 11). Next, in the translation process, all partial word strings “w _i ,..., W _j (i ≦ j, i = 1,..., N, j = 1,..., N)” are extracted from the words divided in S11. Generate.

Specifically, in the translation process, i = 1 is set (S12), j = i is set (S13), and it is determined whether the word string “w _i ,..., W _j ” is in the translation example. (S14). Specifically, for the word string “wi,..., W _j ”, examples stored in the clause / phrase unit translation example storage means 13 (“wi,..., W _j ” ⇔ “W _I ,... W _J ]) And the source language side word string “wi,..., W _j ”.

Translation processing word string "wi, ..., _{w j"} (in S14, YES) if there is the translation example, object-side word string translation example _"W I, ..., _{W J"} the example-based partial translation storage means The translation result is stored as t (i, j) = W _I ,..., W _J for the 14 CYK tables 20 (S15).

  Next, if the translation process is not in the translation example (NO in S14), 1 is added to j (S16), and it is determined whether j is greater than n (j> n?) (S17), If j is not larger than n (NO in S17), the process returns to S14. That is, in the above-described processing, each word of the input word string is connected one word at a time, and each of the connected word strings is compared with the above-described example, and if there is a corresponding translation result, it corresponds. The translation result is stored in the storage area of the CYK table 20.

  In the translation process, when j is larger than n (YES in S17), 1 is added to i (S18), and it is determined whether i is larger than n (i> n?) (S19). If i is not larger than n (NO in S19), the process returns to S13. That is, in the above-described processing, the first word in the input word string is shifted one word at a time, collated with the above-described example, and if there is a corresponding translation result, the translation is performed in the storage area of the corresponding CYK table 20. Store the result.

  In the translation process, when i is larger than n (YES in S19), i = 1 is set again (S20) and j = i is set (S21) as the next process. In the translation process, 0 is set to a new variable k (S22).

Next, the translation process generates k = 1,..., J-i for t (i, j) (i = 1,..., N, j = 1,..., N), and example-based partial translation. Referring to the storage means 14, it is determined whether or not both t (k, i + k−1) and t (j−i + 1−k, j) are “empty” (S23). Here, in the process of S23, if both of the translation processes are not “empty” (YES in S23), t (k, i + k−1) and t (j−i + 1−k, j) of the CYK table 20 are used. translation result _{t (k, i + k-} 1) stored = W _I1, ···, and _{W J1, t (j-i} + 1-k, j) = W I2, ···, and a _{W J2} linked , T (j−i + 1−k, j) = W _I1 ,..., W _J1 , W _I2 ,..., W _J2 (additionally registered in the example base partial translation storage unit 14) (S24) .

  Also, in the translation processing, if both are not “empty” in the processing of S23 (NO in S23), 1 is added to k (S25), and whether k is larger than j−1 (k> j−1). ?) Is determined (S26). If k is not larger than j−1 (NO in S26), the process returns to S23. That is, in the above process, it is confirmed whether or not the translation result is stored for each storage area of the CYK table 20, and if the translation result exists in the two storage areas to be linked, the linkage process is performed. The registration result is additionally registered (stored) in the corresponding storage area.

  In the translation process, when k is larger than j-1 (YES in S26), 1 is added to j (S27). Next, the translation processing determines whether j is greater than n (j> n?) (S28). If j is not greater than n (NO in S28), the process returns to S22. In the translation process, when j is larger than n (YES in S28), 1 is added to i (S29), and it is determined whether i is larger than n (i> n?) (S30). If i is not greater than n (NO in S30), the translation process returns to S21. That is, in the above-described processing, the above-described concatenation processing is performed on all the storage areas while shifting each storage area of the CYK table 20 until both i and j exceed n.

  In the translation process, if i is larger than n (YES in S30), it is determined whether or not the storage area t (n, n) of the CYK table 20 is “empty” (S31).

  In the translation process, if t (n, n) is not “empty” (YES in S31), an example translation can be performed as a whole sentence. Therefore, the translation result stored in t (n, n) ( The sign language word string is output and the process ends (S32).

  Also, in the translation process, when t (n, n) is “empty” (NO in S31), variables p = 1 and q = n are set (S33), and k = q−p−1 is set ( S34), h = 0 is set, and it is determined whether t (k + 1, p + h + k) is not “empty” (S36). In the translation process, when t (k + 1, p + h + k) is “empty” (NO in S36), 1 is added to h (S37), and whether h is equal to or less than q−p−k (h ≦ q− pk?) is determined (S38), and if it is not less than qpk (NO in S38), the process returns to S36. Also, in the translation process, when h is equal to or less than q−p−k (YES in S38), 1 is subtracted from k (S39), and whether or not k is smaller than 0 (k <0?). Is determined (S40). Also, the translation process returns to S35 if k is not smaller than 0 (NO in S40). That is, in the above-described processing, as shown in FIG. 3C described above, the translation result is extracted from the CYK table 20 (a word string having a large number of connections).

  Further, in the translation process, when k is smaller than 0 (YES in S40), it means that the example translation as a whole sentence could not be performed, so the whole sentence is statistically machine translated (S41).

In the translation process, if t (k + 1, p + h + k) is not “empty” in the process of S36 (YES in S36), the translation result stored in t (k + 1, p + h + k) is extracted (S42), w _p, ..., and _{w p + h-1, w} p + h + k + 1, ..., w q determine the partial translation result and the untranslated word against (S43), to translate a statistical machine translation of the word strings that could not be example-based translation ( S44).

  Thereafter, the translation result (sign language word string) obtained by the processing of S41 and S44 is output (S45). In addition, when outputting the translation result obtained by the process of S44, the translation result which synthesize | combined the translation result translated by the example base and the translation result translated by statistical machine translation is output, and a process is carried out. finish.

  Here, FIG. 5 is a diagram showing a specific example of the example translation technique using the CYK table. 5A shows an example of translation example data, FIG. 5B shows an example of the CYK table, FIG. 5C shows a specific example of the example translation result, and FIG. d) shows a specific example of partial translation.

Here, the example of the CYK table 20 in FIG. 3 gives i and j by paying attention to the word, whereas the example of the CYK table in FIG. 5B pays attention to the storage area of the table. i and j are assigned. That is, the CYK table 20 of FIG. 3 shows “translation result of w _i ,..., W _j = t (j−i + 1, j)” (hereinafter referred to as equation (1)), whereas FIG. The CYK table shows “translation results of t (i, j) = w _{j−i + 1} ,..., W _j ” (hereinafter referred to as equation (2)), and both can be converted into each other.

Specifically, when “x = j−i + 1, y = j” in the above-described equation (1), “j = y, i = j−x + 1 = y−x + 1” is obtained. Therefore, the expression (1) is expressed as “translation result of left side = w _i ,..., W _j = translation result of w _{y−x + 1} ,..., W _y ” (hereinafter referred to as expression (3)), “right side = t ( j−i + 1, j) = t (x, y) ”(hereinafter referred to as equation (4)). Here, if x → i and y → j are set again, “(3) expression = w _{j−i + 1} ,..., W _j translation result”, (4) = t (i, j) ”, which is described above. Equation (2) is obtained. In other words, in the present embodiment, i and j can be given to the CYK table by paying attention to the word, for example, and i and j can be given by paying attention to the storage area of the table.

A specific example of the example translation method using the CYK table shown in FIG. 5 is a specific example of example translation corresponding to the above-described Japanese sentence 1 (Kyushu and Okinawa are likely to be thunderstorms in the evening). Japanese sentence 1 is divided into “Kyushu / and / Okinawa / ha / evening / from / thunderstorm / to / nari / so / is” based on the morphological analysis results, and when all partial word strings are generated from the divided words, It looks like the following example.
"Kyushu"
"Kyushu /"
"Kyushu / and / Okinawa"
"Kyushu / and / Okinawa / ha"
...
"When"
"To / Okinawa /"
"To / Okinawa / Ha"
...
"That's right"
"is"
Next, a certain word string generated like the partial word string described above is collated with the source language side word string of the translation example stored in the clause / phrase unit translation example storage means 13. It is assumed that the example for translation is given as an example for translation in sections / phrases as shown in FIG.

  In the present embodiment, when the result of collation is the same, the target side word string of the example is stored in the corresponding storage area of the CYK table stored in the example base partial translation storage unit 14. Here, when the partial word string is collated with the Japanese side of the translation example, “Kyushu / to”, “Okinawa / ha”, “Thunderstorm / to / nari / so / is”, “Thunderstorm / to / nari”, “ “So / Is” matches, “t (2,2) = Kyushu N”, “t (2,4) = Okinawa”, “t (5,11) = Thunderstorm Dreams” shown in FIG. N ”,“ t (3, 9) = thunderstorm ”, and“ t (2, 11) = dream N ”are registered (FIG. 5B).

Next, the CYK table of the example base partial translation storage unit 14 is referred to and it is determined whether or not t (k, k) and t (4-k, 4) are both empty. as a result,
When k = 1, t (1,1) = empty, t (3,4) = empty k = 2, t (2,2) = Kyushu N, t (2,4) = Okinawa k = 3 Since t (3,3) = empty and t (1,4) = empty, the storage areas to be linked are not empty when k = 2.

  Therefore, when both t (k, k) and t (4-k, 4) are not empty, the translation result “t (k, k) and t (4-k, 4) stored in“ t (2,2) = Kyushu N ”and“ t (2,4) = Okinawa ”are concatenated as a partial translation in the CYK table of the example base partial translation storage means 14 as shown in FIG. “t (4, 4) = Kyushu N Okinawa” is additionally registered (FIG. 5B). However, if the same translation result is already stored, no additional registration is made. For example, “t (3,9) = Thunderstorm” and “t (2,11) = Dream N”, so they are not both empty, but “Thunderstorm Dream N” connecting the two is already stored in t (5,11). Therefore, in the above case, the partial translation is not additionally registered.

  Next, it is determined whether t (11, 11) is not empty. In the example of FIG. If t (11,11) is not empty, the translation result stored in t (11,11) is output and the process ends.

Next, for the input sentence “w ₁ ,..., W _i ,..., W _j ,..., W ₁₁ ”, i = 1 + h, j = 1 + h + k (k = 9, 8,..., 0, h = 0, 1,..., 10-k), the following processing is performed. It is determined whether or not t (k + 1, 1 + h + k) in the example base partial translation storage unit 14 storing the translation result of the word string is not empty. For example, when k = 4, h = 0, 1,..., 6; “when h = 0, t (5,5) = empty”; when “h = 1”, t (5,6 ) = Empty ”,“ h = 2, t (5,7) = empty ”,“ h = 3, t (5,8) = empty ”,“ h = 4, t (5 9) = sky ”,“ when h = 5, t (5,10) = sky ”, and when h = 6, t (5,11) = thunderstorm dream N”. Therefore, it is not empty only when h = 6.

  In the present embodiment, when no translation result is stored in the CYK table of the example-based partial translation storage unit 14 (when all t (i, j) are empty), the entire sentence is stored. Perform statistical machine translation. This is a process in the case where even partial translation cannot be obtained by example-based translation. In this case, the entire sentence is translated by the statistical machine translation means 16. The statistical machine translation means 16 performs statistical machine translation using a translation dictionary that is learned in advance using translation examples stored in the clause / phrase unit translation example storage means 13.

  Next, since t (5,11) in the CYK table of the example base partial translation storage means 14 is not “empty”, the translation result “Thunderstorm dream N” stored in t (5,11) is extracted.

Then, in the example of FIG. 5, the input sentence _{"w 1, ..., w i,} ..., w j, ..., w 11 " in the, the translation results are obtained word sequence _{"w 7,} ..., _{w 11} For the word string “w ₁ ,..., W ₆ ” before “,” a method similar to the above is recursively repeated with p = j + 1 and q = 6. As a result, if the CYK table is not empty, the translation result is stored as a partial translation. In the example of FIG. 5, the partial translation “Kyushu N Okinawa” of “Kyushu / and / Okinawa / ha” is stored in the corresponding storage area of the CYK table stored in the example-based partial translation storage means 14. The words that could not be translated are stored in the untranslated word storage means 15. In the example of FIG. 5, “evening” and “kara” are stored in the untranslated word storage means 15.

  Next, in the present embodiment, the words “evening” and “from” stored in the untranslated word storage unit 15 are arranged in order to form the word string “evening / from”, and each word string is processed by the statistical machine translation device. Translate and obtain the translation result “Yugata from” as a partial translation.

  Next, in the present embodiment, the translation results of the partial translation described above are output in the input data input order, that is, “Kyushu N Okinawa Yugata to Thunder Rain Yume N” is output as a translation result (sign language word string). .

<Statistical translation in statistical machine translation means 16>
Next, the statistical translation in the statistical machine translation means 16 mentioned above is demonstrated concretely. In the present embodiment, in the example-based machine translation in the example-based machine translation means 12 described above, statistical translation is performed for each partial word string for a Japanese partial word string that could not be example-translated. In this case, the unit of the partial word string is not necessarily a linguistic unit such as a clause or a phrase.

  In statistical translation, the example of clause or phrase used in example translation is used for learning a translation model, and a sign language sentence is used for learning a language model. In the statistical translation in the statistical machine translation means 16, statistical translation can be performed using a predetermined translation model (for example, GIZA ++), decoding (for example, mosaics), language model (for example, SRILM), and the like. However, the present invention is not limited to this.

<About weighting>
Here, in the example translation method in the present embodiment, the example data for translation stored in the section / phrase unit translation example storage unit 13 described above is used, but a plurality of translation results are extracted for one word. There is. In such a case, the appearance frequency for each translation result is set in advance in the example data for translation, and when the example translation is performed, a plurality of translation results are extracted with this appearance frequency as a weight. By outputting a large (heavy weight) one, a more accurate translation result can be acquired.

  Here, FIG. 6 is a diagram illustrating an example of translation example data including the appearance frequency. Appearance frequencies are set in advance in the bilingual data of Japanese and sign language shown in FIG. 6 (numerical values on the left side of FIG. 6). For example, as shown in FIG. 6, when “Okinawa is” ⇔ “Okinawa N” and “Okinawa is” ⇔ “Okinawa” are stored in the example data for translation, the word string “Okinawa is The two translation results are extracted, and “Okinawa” having a high appearance frequency is output as a translation result (sign language word).

<Outline of sign language CG translation system>
In the present embodiment, a sign language image corresponding to input data can be output using the translation result (sign language word string) synthesized by the partial translation synthesis unit 17.

  Here, FIG. 7 is a diagram illustrating an example of a generation screen of the sign language CG translation system. The screen 30 shown in FIG. 7 includes a sign language video display area 31, a Japanese input area 32, a translation candidate display area 33, a sign language word display area 34, and a motion video link display area 35. Yes.

  For sign language translation motion generation, for example, the above-described morphological analysis is performed on a Japanese sentence input to the Japanese input area 32, and each word of the analysis result is converted using translation example data, a translation dictionary, or the like. Then, a smooth motion video for the input data is generated by connecting the motion video preset to each sign language word by linear interpolation.

  The generated video is displayed in the sign language video display area 31 and provided to the user or the like. At this time, in the sign language video display area 31, as shown in FIG. 6, a sentence (input data) corresponding to the sign language video can be displayed. The image displayed in the sign language video display area 31 includes motion data acquired by optical motion capture and a human body CG model having a skeleton structure. In addition, since the CG drawing uses a video content description language TVML (TV program making language), operation editing can be easily performed.

<Execution program>
Here, the machine translation device 10 described above includes, for example, a volatile storage medium such as a CPU (Central Processing Unit) and a RAM (Random Access Memory), a nonvolatile storage medium such as a ROM (Read Only Memory), a mouse and a keyboard. Further, it can be constituted by a computer including an input device such as a pointing device, a display device for displaying images and data, and an interface device for communicating with the outside.

  Accordingly, the above-described functions of the machine translation apparatus can be realized by causing the CPU to execute a program describing these functions. These programs can also be stored and distributed in a recording medium such as a magnetic disk (floppy disk, hard disk, etc.), optical disk (CD-ROM, DVD, etc.), semiconductor memory, or the like.

  That is, an execution program (machine translation program) for causing a computer to execute the processing in each configuration described above is generated, and the machine translation processing is realized by installing the program in, for example, a general-purpose personal computer or server Can do. In addition, about the process by the execution program in this invention, each process mentioned above is realizable, for example.

  As described above, according to the present invention, highly accurate translation can be realized. Specifically, the present invention can provide machine translation that fuses example-based translation and statistical machine translation. Further, in the present invention, the word string that can be translated by using a long example in the input sentence is efficiently translated, and the word string that cannot be translated is translated by statistical machine translation, so that example-based translation and statistical Machine translation can be fused.

  The present invention automatically applies proper nouns when translating Japanese into sign language, and can be widely applied to natural language processing used to support sign language interpretation and conversion to CG. Therefore, the input data is not limited to the weather information described above, and can be applied to, for example, a lot of standard expressions such as news manuscripts and sports news, and any other natural language.

  In the machine translation described above, an example of translation from Japanese into sign language has been shown. However, the present invention is not limited to this. For example, between the languages in which the word order is the same between the source language and the translated language. This can also be applied to cases (for example, Japanese and Korean, English and French).

  Although the preferred embodiment of the present invention has been described in detail above, the present invention is not limited to the specific embodiment, and various modifications, within the scope of the gist of the present invention described in the claims, It can be changed.

DESCRIPTION OF SYMBOLS 10 Machine translation apparatus 11 Morphological analysis means 12 Example base machine translation means 13 Example / phrase unit translation example storage means 14 Example base partial translation storage means 15 Untranslated word storage means 16 Statistical machine translation means 17 Partial translation synthesis means 20 CYK table 30 screen 31 sign language video display area 32 Japanese input area 33 translation candidate display area 34 sign language word display area 35 motion video link display area

Claims (5)

In a machine translation device that translates a source language into a target language using an example,
Example-based machine translation means for performing example-based translation for each word included in the input data of the source language, using example data for translation in a section / phrase unit of the target language corresponding to the source language stored in advance,
A machine translation apparatus comprising: a partial translation synthesis unit that synthesizes partial translations for one or a plurality of words translated by the example-based machine translation unit. Statistical machine translation means that performs statistical machine translation using a translation dictionary that has been machine-learned in advance for words that could not be translated by the example-based machine translation means,
The partial translation synthesis means synthesizes the partial translation translated by the example-based machine translation means and the partial translation translated by the statistical machine translation means in correspondence with the word order of the input data. The machine translation apparatus according to claim 1. The example-based machine translation means includes:
The machine translation apparatus according to claim 1, wherein partial translation using each word included in the input data is performed using a CYK table. When the target language is sign language,
4. The partial translation synthesizing unit converts a sign language word string obtained by the synthesis into a sign language video using video data for a pre-stored sign language word, and outputs the sign language video. The machine translation device according to claim 1.   The machine translation program for functioning a computer as each means which the machine translation apparatus of any one of Claims 1 thru | or 4 has.