JP2006024114A - Mechanical translation device and mechanical translation computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a mechanical translation device capable of verifying obtained translated sentences in example-based machine translation and outputting a translated sentence determined to be correct according to a predetermined condition. <P>SOLUTION: The mechanical translation device 130 comprises an example base 106 including a plurality of parallel translation sentences consisting of English sentences and Japanese sentences, an example-based translation section 140 that receives an English input sentence 120, retrieves a parallel translation sentence from the example base 106, and generates a plurality of candidates 132A to 132M of the Japanese translated sentence corresponding to the input sentence 120, and a statistical selection section 150 for selecting, as a translated sentence corresponding to the input sentence 120, one of the plurality of candidates 132A to 132M where a statistical probability score calculated using a language model 102 and a translation model 104 satisfies the predetermined condition. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a machine translation apparatus that performs example translation using a bilingual corpus, and more particularly, to a machine translation apparatus that overcomes the weaknesses of example translation using a statistical model and performs high-quality translation.

  Currently, in machine translation, a translation system that performs translation using a bilingual corpus is producing high results. A typical example of such machine translation is called example translation (see Non-Patent Document 1). For example translation, machine translation is performed using a bilingual corpus as a kind of database (hereinafter referred to as “example base”).

  FIG. 11 schematically shows a configuration of a typical conventional machine translation apparatus that performs example translation. Referring to FIG. 11, a conventional machine translation apparatus 300 uses an input sentence 310 of a first language (hereinafter referred to as “source language”) as a second language (hereinafter referred to as “object”). The example base 304 consisting of a bilingual corpus 302 including a source language sentence and a target language translation corresponding to the sentence is most similar to the input sentence 310. The search unit 320 for searching the translated text (hereinafter referred to as “example”) from the example base 304, and the difference between the source language word string of the example searched by the search unit 320 and the input sentence 310 The difference identification unit 321 for identifying a column, the bilingual dictionary 306 between the source language and the target language, and the bilingual dictionary 306 are identified by the difference identifying unit 321 among the examples obtained by the search unit 320. Source language Based on the difference of the sentence and the input sentence 310, and a correction unit 322 for generating an output sentence 312 and correct the statement in the target language of the examples.

  When the input sentence 310 is given to the machine translation apparatus 300, the search unit 320 searches the example base 304 for an example that is most similar to the input sentence 310. The searched example is given to the difference identification unit 321. An input sentence 310 is also given to the difference identification unit 321. The difference identification unit 321 specifies a different part between the input sentence 310 and the sentence on the source language side of the searched example. The difference identification unit 321 gives the example to the correction unit 322 together with information specifying the input sentence 310 and the difference between the examples. The correcting unit 322 refers to the bilingual dictionary 306 and corrects the part specified in the sentence on the target language side of the example using the translation of the word at the corresponding part of the input sentence 310 and outputs the corrected sentence to the output sentence 312 is output.

  For example, consider the case where the source language is English and the target language is Japanese. In this case, as the example base 304, one containing many parallel translations of English and Japanese is used. Assume that an English sentence “where is the cheap hotel” is input as the input sentence 310. The search unit 320 searches the example base 304 for an example in which the English sentence is most similar to the input sentence 310. Here, it is assumed that an example “where is the cheapest restaurant / where is the cheapest restaurant” is retrieved from the example base 304. The search unit 320 gives this example and the input sentence 310 to the difference identification unit 321.

  The input sentence 310 includes the word “hotel”, whereas the sentence on the English side of the example differs from the input sentence and the example in that there is a word “restaurant” instead of “hotel”. The difference identification unit 321 identifies this difference between the English sentence of the example and the input sentence 310. The difference identification unit 321 further identifies the word “restaurant” corresponding to the word “restaurant” identified as being different from the input sentence 310 in the sentence on the English side of the example among the sentences on the Japanese side of the example. . The difference identification unit 321 gives the modification example 322 the retrieved example and the different word “hotel” in the input sentence 310 for the example. At this time, the difference identification unit 321 instructs the correction unit 322 of the word “restaurant” corresponding to the word “restaurant” in the example. The correcting unit 322 refers to the bilingual dictionary 306 and obtains “hotel” which is a translated word of the different word “hotel” in the input sentence 310. The correction unit 322 replaces the word “restaurant” instructed by the difference identification unit 321 with “hotel” obtained from the bilingual dictionary 306. The machine translation apparatus 300 outputs the Japanese sentence “Where is the cheapest hotel” obtained as a result of these series of processes as an output sentence 312?

  As an example used in example translation, a parallel translation in units of phrases or sentences is often used. Therefore, example translation has the advantage that conventional expressions and the like can be appropriately translated.

Nagao, M.M. (1984). "Machine Translation Mechanism between Japanese and English by Similar Principles", Artificial and Human Intelligence, pages 173-180, Amsterdam: North Holland (Nagao, M. (1984). "A framework of mechanical translation Japan and English by. analog principal ". In Artificial and Human Intelligence, pages 173-180, Amsterdam: North-Holland.) Imamra, K.M. (2002). “Application of translation knowledge acquired by hierarchical phrase alignment for pattern-based MT”. 9th Conference on Theoretical and Methodological Issues in Machine Translation (TMI-2002) Proceedings, pp. 74-84. (Imamura, K. (2002). "Application of translation knowledgwe acquired by hierarchical phrase alignment for pattern-based MT" .In Proceedings of the 9th Conference on Theoretical and Methodological Issues in Machine Translation (TMI-2002), pages 74-84 .)

  In the example translation, for example, when the similarity is measured by the number of different words, it is assumed that an example having the smallest number of different words from the input sentence can be searched from the example base 304. However, such an example cannot always be uniquely searched, and a plurality of examples having similar word differences may compete. When a plurality of examples conflict, the conventional example translation measures the similarity between the input sentence and the source language side of those examples, and selects the example having the highest similarity. However, even when the most similar example is selected, an appropriate output sentence is not always generated from the selected example.

  For example, it is assumed that a sentence “where is the nearest resturant” is input. As examples similar to this input sentence, there are an example of “where is the nearest subway / where is the nearest subway station” and an example of “where is the cheapest restaurant / where is the cheapest restaurant”. Suppose you compete. Assume that the former is selected as an example most similar to the input sentence. In this case, since “subway” in the example is different from “restaurant” in the input sentence, “subway” is replaced with the translated word “restaurant” of “restaurant”. Therefore, “where is the nearest restaurant station” is output as the output sentence 312. However, this output sentence is a mistranslation.

  If the latter is selected, the example “cheapest” is different from the input sentence “nearest”. Therefore, the output sentence 312 is generated by replacing “cheapest” in the example with the translated word “nearest”. Here, by referring to the bilingual dictionary 306, if “nearest” is obtained as the translation of “nearest”, the output sentence 312 becomes “where is the nearest restaurant” and becomes a correct translation. However, this is not always the case. When “nearest” is obtained as the translation of “nearest”, the output sentence 312 is “where is the nearest restaurant”. This is a grammatically incorrect sentence.

  As described above, in the example translation, the example is corrected without considering what result is obtained by the translation. Therefore, it is impossible to verify whether the final output sentence is correct as a translation of the input sentence, and there is a possibility that an incorrect sentence is output as a mistranslated sentence or a target language sentence.

  SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide a machine translation apparatus capable of verifying an obtained translation sentence and outputting a translation sentence determined to be correct according to a predetermined condition in an example machine translation.

  Another object of the present invention is to determine whether or not a translated sentence after machine translation is a correct sentence as a sentence in a target language in an example machine translation, and to output a translated sentence that can be determined as a correct sentence Is to provide.

  Still another object of the present invention is to determine whether a translated sentence after machine translation is a correct sentence as a sentence in a target language by using a statistical model used in statistical machine translation in an example machine translation, An object of the present invention is to provide a machine translation apparatus that can output a translated sentence that can be determined as a correct sentence.

  A machine translation device according to a first aspect of the present invention includes a predetermined example base including a plurality of examples including pairs of a sentence in a first language and a sentence in a second language, and an input sentence in the first language. Then, referring to the example base, an example translation unit for generating a plurality of candidates for the translated sentence of the second language for the input sentence, and a predetermined probability statistic among the plurality of candidates generated by the example translation unit Statistical selection means for selecting and outputting a probability score calculated using a model that satisfies a predetermined condition.

  Preferably, the example translation means receives an input sentence, searches an example base having a sentence in a first language satisfying a predetermined similarity condition with the input sentence, and searches the sentence in the first language searched for. Retrieval means for extracting a plurality of examples included therein, and correction means for correcting each sentence in the second language of the plurality of examples and generating translation sentence candidates for the input sentence from each of the plurality of examples Including.

  More preferably, the search means receives an input sentence, searches for a plurality of sentences in the first language having the smallest number of words unlike the input sentence on an example basis, and includes each searched sentence in the first language. Means for obtaining a plurality of examples are included.

  More preferably, the machine translation device further includes a bilingual dictionary between the first language and the second language. The correcting means compares the input sentence with sentences in the first language of each of the plurality of examples, and for each of the plurality of examples, difference identifying means for identifying a difference from the input sentence, and a plurality of examples Candidate generation for correcting a sentence in the second language with reference to the bilingual dictionary based on the difference identified by the difference identifying means, and generating a translation sentence candidate for the input sentence from each of a plurality of examples Means.

  The bilingual dictionary may include a plurality of words in the second language as translations for one word in the first language. The candidate generating means includes one or a plurality of second language words obtained by referring to the bilingual dictionary based on the differences identified by the difference identifying means for the sentences in the second language of the plurality of examples. Means for generating one or a plurality of translation sentence candidates for the input sentence by modifying each of them using.

  The means for generating the sentence of the second language of each of the plurality of examples is obtained by referring to the bilingual dictionary based on each of the differences identified by the difference identifying means. Means may be included for generating one or more translation sentence candidates for the input sentence by modifying according to all possible combinations using words in the language.

  The search means receives the input sentence, searches the example base for a plurality of first language sentences that have the smallest editing distance from the input sentence, and includes a plurality of examples each including the searched first language sentence. Means for obtaining may be included.

  The search means receives an input sentence, searches the example base for sentences in a plurality of first languages that have the smallest editing distance from the input sentence calculated in consideration of the semantic distance between words, and is searched. A means for acquiring a plurality of examples each including a sentence in the first language may be included.

  Preferably, the statistical selection means includes means for selecting and outputting a candidate having the highest probability score calculated using a predetermined probability statistical model from among a plurality of candidates generated by the example translation means.

  More preferably, the machine translation device further includes language model storage means for storing a language model of the second language. The means for outputting includes, for each of the plurality of candidates, a language probability calculation means for calculating a language probability using a language model stored in the language storage means, and a language probability calculated by the language probability calculation means Means for selecting and outputting the highest candidate.

  More preferably, the machine translation device further includes a translation model storage unit for storing a translation model from the second language to the first language. The means for outputting includes, for each of the plurality of candidates, a translation probability calculation means for calculating a translation probability using a translation model stored in the translation model storage means, and a translation calculated by the translation probability calculation means Means for selecting and outputting a candidate having the highest probability.

  More preferably, the machine translation apparatus further includes language model storage means for storing a language model of the second language, and translation model storage means for storing a translation model from the second language to the first language. Including. The means for outputting, for each of a plurality of candidates, a language probability calculating means for calculating a language probability using a language model stored in the language storage means, and a translation model for each of the plurality of candidates A translation probability calculation means for calculating a translation probability using a translation model stored in the storage means, a language probability calculated by the language probability calculation means, and a translation probability calculated by the translation probability calculation means as a predetermined function. Score calculating means for calculating a probability score; and means for selecting and outputting a candidate having the highest probability score calculated by the score calculating means.

  The machine translation computer program according to the second aspect of the present invention, when executed by a computer, causes the computer to operate as the machine translation apparatus according to the first aspect of the present invention.

  A machine translation system according to a specific embodiment of the present invention described below is realized by computer hardware, a program executed by the computer hardware, and data stored in the computer hardware. FIG. 1 shows the external appearance of a computer system 30 that implements this machine translation system, and FIG. 2 shows the internal configuration of the computer system 30.

  Referring to FIG. 1, a computer system 30 includes a computer 40 having an FD (flexible disk) drive 52 and a CD-ROM (compact disk read only memory) drive 50, a keyboard 46, a mouse 48, and a monitor 42. including.

  2, in addition to the FD drive 52 and the CD-ROM drive 50, the computer 40 includes a CPU (central processing unit) 56 and a bus 66 connected to the CPU 56, the FD drive 52, and the CD-ROM drive 50. And a read only memory (ROM) 58 for storing a boot-up program and the like, and a random access memory (RAM) 60 connected to the bus 66 for storing a program command, a system program, work data and the like. The computer system 30 further includes a printer 44.

  Although not shown here, the computer 40 may further include a network adapter board that provides a connection to a local area network (LAN).

  A computer program for causing the computer system 30 to realize the functions of the machine translation system according to the present embodiment is stored in the CD-ROM 62 or FD 64 inserted into the CD-ROM drive 50 or FD drive 52 and further stored in the hard disk 54. Transferred. Alternatively, the program may be transmitted to the computer 40 through a network (not shown) and stored in the hard disk 54. The program is loaded into the RAM 60 at the time of execution. The program may be loaded directly into the RAM 60 from the CD-ROM 62, from the FD 64, or via a network.

  The machine translation program described below includes a plurality of instructions that cause the computer 40 to realize the functions of the machine translation apparatus according to the present embodiment. Some of the basic functions necessary to implement this device are provided by operating system (OS) or third-party programs running on the computer 40 or various toolkit modules installed on the computer 40. Therefore, this program does not necessarily include all functions necessary for realizing the machine translation apparatus according to this embodiment. This program need only contain instructions that implement the functions of the machine translation device by calling the appropriate functions or “tools” in a controlled manner to obtain the desired results. The operation of computer system 30 is well known and will not be repeated here.

-Overview-
As a method of machine translation using a bilingual corpus, there is a method called statistical translation in addition to example translation. In this method, the problem of translating a sentence in one language into a sentence in another language is formulated as a problem of maximizing conditional probability. If the source language sentence is F and the target language sentence is E, statistical translation searches the source language sentence F for a sentence that satisfies the following expression.

ここで、Ｐ（Ｅ）とＰ（Ｆ）とは、目的言語Ｅの文の尤度と原言語Ｆの文の尤度とをそれぞれ表わし、その文の流暢さの指標となる。Ｐ（Ｆ｜Ｅ）は、目的言語Ｅの文から原言語Ｆの文が得られる確率を表わし、翻訳の正確さの指標となる。上にあげた式のうちＰ（Ｆ）は定数である。したがってこの問題は、次の式を満足する文Ｅを探索する問題となる。

Here, P (E) and P (F) represent the likelihood of the sentence in the target language E and the likelihood of the sentence in the source language F, respectively, and serve as indices of the fluency of the sentence. P (F | E) represents a probability that a sentence in the source language F can be obtained from a sentence in the target language E, and is an index of translation accuracy. Of the equations listed above, P (F) is a constant. Therefore, this problem is a problem of searching for a sentence E that satisfies the following expression.

統計翻訳においてＰ（Ｅ）は言語Ｅの言語モデル確率と呼ばれる。言語モデル確率Ｐ（Ｅ）は、対訳コーパスをもとに計算される。あるコーパスに含まれる言語Ｅの全ての語に
ついて言語モデル確率を算出したものを言語モデルと呼ぶ。また、Ｐ（Ｆ｜Ｅ）は、翻訳
モデル確率と呼ばれる。翻訳モデル確率Ｐ（Ｆ｜Ｅ）もまた、対訳コーパスをもとに算出される。なおここで使用される翻訳モデルが目的言語から原言語へのものであり、翻訳の向きとは逆であることに注意が必要である。

In statistical translation, P (E) is called the language model probability of language E. The language model probability P (E) is calculated based on the bilingual corpus. What calculated the language model probability about all the words of the language E contained in a certain corpus is called a language model. P (F | E) is called a translation model probability. The translation model probability P (F | E) is also calculated based on the bilingual corpus. It should be noted that the translation model used here is from the target language to the source language and is opposite to the direction of translation.

  The machine translation apparatus according to the present embodiment selects and outputs an optimum output sentence by introducing a sentence search using the language model and the translation model into the example translation.

−Configuration−
FIG. 3 shows a block diagram of a machine translation system 90 according to the present embodiment. Referring to FIG. 3, the machine translation system 90 includes a parallel corpus 100 including a large number of parallel translations composed of a source language sentence (English) and a translation of a target language (Japanese) for the sentence, and a parallel corpus 100. In addition, a language model creation device 110 for creating a Japanese language model 102 as a target language and a translation model creation device 112 for creating a translation model 104 from the target language to the source language based on the parallel translation corpus 100. And a machine translation device 130 that translates the input sentence 120 in the source language using the example base 106 including the bilingual corpus 100, the language model 102, and the translation model 104, and generates an output sentence 122 in the target language.

  FIG. 4 shows the configuration of the example base 106. Referring to FIG. 4, the example base 106 includes a translation sentence number, a word string (source language word string) on the source language (English) side of the translation sentence, and a word string (objective translation) of the target language with respect to the source language sentence. Language word string) and information indicating word correspondence in the parallel translation. In FIG. 4, “/” indicates a word break. The word correspondence indicates a set of a word in the source language word string and a word or a word string in the target language word string corresponding to the word. Usually, the bilingual corpus 100 includes such information, and the bilingual corpus 100 can be used as the example base 106.

FIG. 5 shows the configuration of the language model 102. The language model 102 shown in FIG. 5 is a word bigram model. In the language model 102, the word e _i is generated under the condition that the ordered set of the words e _i-1 and e _i in the target language sentence of the bilingual corpus 100 and the word e _i-1 are placed immediately before. It contains a large number of entries with probability p (e _i | e _i-1 ). In the language model 102 shown in FIG. 5, “<s>” is a symbol indicating the start position of a sentence, and “</ s>” is a symbol indicating the end position of the sentence.

FIG. 6 shows the configuration of the translation model 104. The translation model 104 shown in FIG. 6 is a Lexicon (vocabulary) model. The translation model 104 includes a pair of a target language (Japanese) word e _i and a source language word f _j in the parallel corpus 100, and a word e _i in the parallel translation sentence in the parallel corpus 100 as a source language word f _j . And a value t (f _j | e _i ) representing the probability of being translated. In the translation model 104 shown in FIG. 6, “NULL” is a special word indicating that the source language word is not translated. The translation model is a model that represents the accuracy of translation of words between the source language and the target language.

  Referring to FIG. 3, in addition to the example base 106, the machine translation device 130 includes a bilingual dictionary 108 including a number of bilingual translations composed of a source language word and one or a plurality of target language words or word strings corresponding to the word. When the input sentence 120 is given, the example base 106 and the bilingual dictionary 108 are used, and a plurality of Japanese output sentence candidates 132A,. .., 132M are statistically determined to be statistically optimal using the language model 102 and the translation model 104. And a statistical selection unit 150 for selecting and outputting the output sentence 122.

  FIG. 7 shows the configuration of the bilingual dictionary 108. Referring to FIG. 7, bilingual dictionary 108 includes a large number of entries each consisting of a set of a source language word and a target language word string which is a translation for that word. When there are a plurality of translations for the source language word, the plurality of translations are listed in the parallel translation dictionary. For example, for the source language word “nearest”, the target language word string lists the translation of “nearest” and the translation of “most / closest”. In the bilingual dictionary 108 shown in FIG. 7, “/” indicates a word break as in the example base 106 (see FIG. 4).

  FIG. 8 shows a detailed block diagram of the example translation unit 140 (see FIG. 3). Referring to FIG. 8, the example translation unit 140 searches the example base 106 for a predetermined number (N) of examples having source language sentences (English word strings) similar to the input sentence 120. A difference identification unit 202 for identifying a difference word between the source language word string of each of the N examples searched for by the search unit 200 and the input sentence 120, and a difference identification in the target language word string of each example The translation corresponding to the different word in the source language identified by the unit 202 is identified by the word correspondence information in the example included in the example base 106, and the translation is retrieved from the parallel translation dictionary 108 using the different word in the input sentence 120. And a modification unit 204 for generating output sentence candidates 132A,..., 132M by replacement or modification. Here, since a plurality of translated words may be obtained from the bilingual dictionary 108, M ≧ N.

  FIG. 9 shows a detailed block diagram of the statistical selection unit 150 (see FIG. 3). Referring to FIG. 9, the statistical selection unit 150 is based on the language model probability calculation unit 220 that calculates the language model probability of each of the output sentence candidates 132A,. A translation model probability calculation unit 222 that calculates a translation model probability between the input sentence 120 and each of the output sentence candidates 132A,. The statistical selection unit 150 further multiplies the language model probability and the translation model probability for the output sentence candidates 132A,..., 132M, and uses the obtained values for the statistical probabilities for the output sentence candidates 132A,. A selection unit for selecting the one with the highest statistical probability score given by the multiplication unit 224 from the multiplication unit 224 given as a score and the output sentence candidates 132A,. Part 226.

Language model probability calculation unit 220, the language model probability for the word bigram model language model 102 used consisting, output sentence candidates E consisting of the word e _i of the number of words _{W E (1 ≦ i ≦ W} E) of FIG. 5 P (E) is calculated by the following formula.

ここで、ｐ（ｅ_i｜ｅ_i-1）は、直前に単語ｅ_i-1が置かれているという条件下で単語ｅ_iが生起する確率を表わす。

Here, p (e _i | e _i-1 ) represents the probability that the word e _i will occur under the condition that the word e _{i -1} is placed immediately before.

Translation model probability calculation unit 222, using a translation model 104 consisting Lexicon model shown in FIG. 6, the number of words W _E word _{e i (1 ≦ i ≦ W} E) of the output sentence candidates E and the number of words W _F consisting of The translation model probability P (F | E) for conversion with the input sentence F (120) consisting of the word f _j (1 ≦ j ≦ W _F ) is calculated by the following equation.

ここで、ｔ（ｆ_j｜ｅ_i）は、目的言語の単語ｅ_iが原言語の単語ｆ_jに翻訳される確率を表わし、ｔ（ｆ_j｜ＮＵＬＬ）は，原言語単語ｆ_jが翻訳されない確率を表わす。翻訳モデルは原言語と目的言語との間の単語訳の正確性を表わすモデルである。したがって上の式により算出された翻訳モデル確率Ｐ（Ｆ｜Ｅ）が高いほど、より正確な翻訳であると考えられる。

Here, t (f _j | e _i ) represents the probability that the target language word e _i will be translated into the source language word f _j , and t (f _j | NULL) represents the source language word f _j translated. Represents the probability of not being A translation model is a model that represents the accuracy of a word translation between a source language and a target language. Therefore, it is considered that the higher the translation model probability P (F | E) calculated by the above equation, the more accurate the translation.

-Processing structure-
FIG. 10 is a flowchart showing the structure of processing executed in the machine translation apparatus 130. Referring to FIG. 10, when input sentence 120 (see FIG. 3) is given to machine translation device 130, in step S <b> 252 (hereinafter simply referred to as “S”), the number of words is different from input sentence 120. The smallest N examples are searched from the example base 106.

  Subsequently, the processing of S256 and S258 surrounded by S254A and S254B is executed for all N examples retrieved in S252. That is, in S256, the input sentence 120 and the retrieved source language word string are compared, and a difference between the two, that is, a different word is identified. In S258, the word corresponding to the different word in the source language word string is identified from the word correspondence information in the example base 106 in the target language word string of the example, and the translated word of the different word in the input sentence is acquired from the parallel translation dictionary Then, the word corresponding to the different word in the target language word string of the example is replaced with the translation acquired from the parallel translation dictionary 108. The target language word string corrected in this way becomes the output sentence candidate 132. When a plurality of translation words are listed in the bilingual dictionary, an output sentence candidate 132 is generated for each of them. If there are a plurality of different words, all possible translations for each of them are examined, and output sentence candidates 132 are generated according to all possible combinations between them. When the output sentence candidate 132 is generated for all the examples, the process proceeds to S260A.

  The processing of S262, S264, and S266 surrounded by S260A and S260B is executed for all output sentence candidates 132 generated in S254A to S254B. In S262, a language model probability is calculated for each of the output sentence candidates 132. In S264, a translation model probability between each of the output sentence candidates 132 and the input sentence is calculated. In S266, a statistical probability score that is the product of the language model probability calculated in S262 and the translation model probability calculated in S264 is calculated for each of the output sentence candidates 132. When the above process is performed for each output sentence candidate 132, the process proceeds to S268.

  In S268, the output sentence candidate 132 having the largest statistical probability score value calculated in S266 is selected as an output sentence, and the process ends.

-Operation-
The machine translation system 90 operates as follows. Referring to FIG. 3, the bilingual corpus 100 includes a large number of bilingual sentences including a source language sentence and a target language translation sentence, and is prepared in advance so that it can be used as an example base 106 (see FIG. 4). It is assumed that The bilingual dictionary 108 (see FIG. 7) is also prepared in advance by some means. The language model creation device 110 creates a language model 102 (see FIG. 5) based on the bilingual corpus 100 and provides it to the machine translation device 130 in advance. Also, the translation model creation device 112 creates a translation model 104 (see FIG. 6) based on the bilingual corpus 100 and provides it to the machine translation device 130 in advance.

  Referring to FIG. 8, input sentence 120 is given to search unit 200 of example translation unit 140. The search unit 200 compares the source language word string in the example base 106 (see FIG. 4) with the input sentence 120, searches N examples in order from the smallest number of words different from the input sentence 120, and performs a difference. This is given to the identification unit 202. The difference identification unit 202 is also given an input sentence 120.

  The difference identification unit 202 compares the source language word string of each example searched by the search unit 200 with the input sentence 120 and identifies a word different from the input sentence 120 in the source language word string. Furthermore, based on the word correspondence in the example, a word string in the target language corresponding to the word identified as a word different from the input sentence 120 is identified in the target language word string in the example. The difference identification unit 202 gives the correction unit 204 the searched examples and the different words in the input sentence 120 for the examples. At this time, the difference identification unit 202 instructs the correction unit 204 to identify a word or word string identified as a different word or word string in the target language word string of the given example.

  The correction unit 204 refers to the bilingual dictionary 108 and acquires a translation of a word given as a different word in the input sentence 120. When a plurality of translations are listed in the bilingual dictionary 108, all listed translations are acquired. The correction unit 204 further generates an output sentence candidate 132 by replacing a different word string on the target language string side of the given example with the translated word acquired from the bilingual dictionary 108. When a plurality of translated words are retrieved from the bilingual dictionary 108, an output sentence candidate 132 is generated for each of the plurality of translated words. When there are a plurality of different words, all of the translated words are used, and output sentence candidates 132 are generated for all possible combinations.

  Thus, the example translation unit 140 generates M (M ≧ N) output sentence candidates 132A,..., 132M for one input sentence 120. The generated output sentence candidates 132A,..., 132M are given to the statistical selection unit 150.

  Referring to FIG. 9, one of output sentence candidates 132A,..., 132M, for example, output sentence candidate 132A is given to language model probability calculation unit 220, translation model probability calculation unit 222, and multiplication unit 226. The language model probability calculation unit 220 calculates the language model probability of the given output sentence candidate 132A according to the following expression, using the language model 102 composed of the word bigram model shown in FIG.

なお、該当するエントリが言語モデル１０２内にないバイグラムが出力文候補１３２Ａ中にある場合、そのバイグラムについてｐ（ｅ_i｜ｅ_i-1）＝１．０×１０^-7であるものとして、言語モデル確率の計算を行なう。

If there is a bigram whose output does not exist in the language model 102 in the output sentence candidate 132A, it is assumed that p (e _i | e _i-1 ) = 1.0 × 10 ⁻⁷ for that bigram. Calculate model probabilities.

  The translation model probability calculation unit 222 receives the input sentence 120 and uses the translation model 104 including the Lexicon model shown in FIG. 6 to calculate the translation model probability between the input sentence 120 and the given output sentence candidate 132A as Calculate according to

なお、該当するエントリが翻訳モデル１０４内にない単語の対応関係が出力文候補１３２Ａと入力文１２０との間にある場合、その単語の組（ｅ_i，ｆ_j）について、ｔ（ｆ_j｜ｅ_i）＝１．０×１０^-7であるものとして、翻訳モデル確率の計算を行なう。

Note that if the correspondence between words for which there is no corresponding entry in the translation model 104 is between the output sentence candidate 132A and the input sentence 120, t (f _j |) for the word set (e _i , f _j ). Translation model probabilities are calculated assuming that e _i ) = 1.0 × 10 ⁻⁷ .

  The multiplication unit 224 multiplies the language model probability P (E) calculated by the language model probability calculation unit 220 and the translation model probability P (F | E) calculated by the translation model calculation unit 222 to obtain a product P (E ) P (F | E) and the given output sentence candidate 132A are given together to the selection unit 226.

  The language model probability calculation unit 220, the translation model probability calculation unit 222, and the multiplication unit 224 further execute the above-described processing for each of the output sentence candidates 132B,. A statistical probability score composed of the product of the language model probability and the translation model probability is assigned to the selection unit 226.

  The selection unit 226 selects the output sentence candidate 132A,..., 132M to which the statistical probability score is assigned, that has the maximum value. The selection unit 226 outputs the selected output sentence candidate as the output sentence 122.

-Translation example-
Taking a case where an English sentence is translated into a Japanese sentence as an example, a specific example of translation by the machine translation system 90 will be shown. For example, it is assumed that an English sentence “where is the nearest resturant” is given to the example translation unit 140 as the input sentence 120. Let N = 2.

  Referring to FIG. 4, the English sentence “where is the nearest subway” of the example number 1 of the example base 106 includes one word different from the input sentence 120. Also, the sentence “where is the cheapest restaurant” on the English side of the example of No. 2 includes one word (cheapest) different from the input sentence 120. In the example “today is free” of number 3, there are two words different from the input sentence 120. In this case, the search unit 200 shown in FIG. 8 searches the example base 106 for the number 1 example and the number 2 example.

  The target language word string of the number 1 example is “nearest / no / subway / no / station / ha / where / is / ka”. In addition, the target language word string of the example of No. 2 is “most / cheap / restaurant / ha / where / is /?”.

  In the example of number 1, “subway”, which is a different word in the source language word string, corresponds to the word “subway” in the target language word string. 8 retrieves the translated word “restaurant” of “restaurant” in the input sentence 120 from the bilingual dictionary 108 shown in FIG. The correcting unit 204 replaces the word “subway” in the target language word string in the example of number 1 with “restaurant” searched from the parallel translation dictionary 108, and generates and outputs an output sentence candidate.

  In the example of number 2, “cheapest”, which is a different word in the source language word string, corresponds to the word string “best / cheap” in the target language word string. Further, the search unit 200 searches for “nearest” and “nearest / closest” as the translation of “nearest” in the input sentence 120 from the bilingual dictionary 108 shown in FIG. The example translation unit 140 replaces the word “most / cheap” in the target language word string in the example number 2 with “nearest” and the word “best / cheap” in the target language word string. Generate and output the output sentence candidate replaced with "Nearest".

In this way, the following three output sentence candidates are created. That is,
(A) “Nearest / No / Restaurant / No / Station / Ha / Where / Is / K”
(B) “The closest / closest / restaurant / ha / where / is / ka”
(C) “Nearest / Restaurant / Ha / Where / Is it /?”
It is. Hereinafter, these output sentence candidates are referred to as “E _a ”, “E _b ”, and “E _c ”, respectively.

The table below shows the language model probabilities P (E) for the output sentence candidates E _a , E _b , and E _c calculated based on the language model 102 shown in FIG. 5, and the translation model 104 shown in FIG. The translation model probabilities P (F | E) for each of the output sentence candidates E _a , E _b , and E _c calculated in the above and the input sentence 120, and a statistical probability score P (E) P ( F | E).

この表より、統計的確率スコアＰ（Ｅ）Ｐ（Ｆ｜Ｅ）が最大のものは、出力文候補Ｅ_bである。よって、出力文候補Ｅ_b、すなわち「一番近いレストランはどこですか」という日本語の文が、選択部２２６により選択され出力文１２２として出力される。

From this table, the output sentence candidate E _b has the largest statistical probability score P (E) P (F | E). Therefore, the output sentence candidate E _b , that is, the Japanese sentence “Where is the nearest restaurant” is selected by the selection unit 226 and output as the output sentence 122.

  As described above, in machine translation apparatus 130 according to the present embodiment, unlike conventional example translation, it is not possible to uniquely determine an example based only on the similarity between the input sentence and the source language side of the example. do not do. Instead, a plurality of examples similar to the input sentence are searched, and translation is performed based on these examples. Therefore, even when a plurality of examples compete, there is no problem caused by selecting one of them. In addition, the machine translation apparatus 130 generates a plurality of output sentence candidates based on the plurality of examples selected as described above. The probability that an appropriate output sentence is included in all of the plurality of output sentence candidates is higher than when only a single output sentence candidate is generated. Therefore, by selecting an appropriate one from a plurality of output sentence candidates, it is possible to increase the possibility that a high-quality output sentence will be output.

  Furthermore, the machine translation apparatus 130 selects an output sentence from the output sentence candidates using a language model and a translation model used in statistical translation. The language model is a model representing the fluency of sentences, and the translation model is a model representing the accuracy of word translation between the source language and the target language. By using these models and comprehensively considering the language model probabilities and translation model probabilities, the output sentence candidates are selected as natural as the target language sentence and accurate as the translation of the input sentence. be able to.

  In the above-described embodiment, the search unit 200 (see FIG. 8) searches for an example similar to the input sentence based on the number of words unlike the input sentence. However, the present invention is not limited to such an embodiment. The edit distance may be used as a reference for searching for an example similar to the input sentence. Also, in calculating the similarity based on the number of different words or the edit distance, the semantic distance obtained using the thesaurus is taken into account, and in calculating the number of different words or the edit distance, if the word is semantically close, the word Similar examples may be searched by reducing the number or editing distance. In this case, it is desirable that the difference identification unit 202 and the correction unit 204 respectively identify the difference between the input sentence and the searched example and correct the searched example by a method according to the search criteria.

  The number N of examples searched by the search unit 200 may be a fixed value given in advance or may be adjusted according to the situation. For example, when the search unit 200 succeeds in retrieving an example having a source language sentence that completely matches the input sentence, the number of examples is set to 1 in response to this, and the search is terminated at that stage. Also good.

  The translation method is not limited to the above-described method, that is, the method of specifying different words in the example using the word correspondence information included in the example base, and the syntax tree of the input sentence using the syntax transfer method (Non-Patent Document 2). May be mapped to the syntax tree of the target language, and when a non-terminal symbol remains in the leaf of the obtained tree structure, a translation word candidate of the leaf may be obtained by referring to the bilingual dictionary. In addition, any type of example translation may be used as long as a plurality of output sentence candidates can be output.

  In the above-described embodiment, the language model 102 is a word bigram model, and the language model probability calculation unit 220 calculates the language model probability using the word bigram model. However, the present invention is not limited to such an embodiment. N-gram models such as word trigram models, part-of-speech N-gram models, combinations of these models, or other languages used in statistical translation It is also possible to use a model. These language models may be created from the bilingual corpus 100 used for example translation as in the present embodiment, or may be created from another corpus.

  In the above-described embodiment, the translation model is a Lexicon model, but the present invention is not limited to such an embodiment. It is also possible to use other translation models, for example, a translation model combined with a Fertility, NULL generation model, and a distortion model. These language models may be created based on a parallel corpus 100 used for example translation as in the present embodiment, or created based on a corpus different from the parallel corpus 100. It may be a thing.

  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 an external view which shows an example of the computer system 30 which implement | achieves the function of the machine translation apparatus which concerns on one embodiment of this invention. 2 is a diagram showing an internal configuration of a computer system 30. FIG. It is a block diagram which shows the functional structure of the machine translation apparatus 130 which concerns on embodiment of this invention. 3 is a diagram illustrating a configuration of an example base 106. FIG. 3 is a diagram illustrating a configuration of a language model 102. FIG. 3 is a diagram illustrating a configuration of a translation model 104. FIG. It is a figure which shows the structure of the bilingual dictionary. 3 is a block diagram showing a functional configuration of an example translation unit 140. FIG. 3 is a block diagram showing a functional configuration of a statistical selection unit 150. FIG. 3 is a flowchart showing a structure of processing executed in the machine translation device 130. It is the schematic which shows the structure of the conventional typical machine translation apparatus which performs example translation.

Explanation of symbols

  30 computer system, 40 computer, 42 monitor, 44 printer, 46 keyboard, 48 mouse, 50 CD-ROM drive, 52 FD drive, 54 hard disk, 56 CPU, 58 ROM, 60 RAM, 62 CD-ROM, 64 FD, 66 Bus, 90 machine translation system, 100 bilingual corpus, 102 language model, 104 translation model, 106 example base, 108 bilingual dictionary, 110 language model creation device, 112 translation model creation device, 120 input sentence, 122 output sentence, 130 machine translation Device, 132A,..., 132M Output sentence candidate, 140 Example translation unit, 150 Statistical selection unit, 200 Search unit, 202 Difference identification unit, 204 Correction unit, 220 Language model probability calculation unit, 222 Translation model probability calculation unit 224 multiplication unit, 226 the selection unit

Claims (13)

A predetermined example base including a plurality of examples of pairs of sentences in a first language and sentences in a second language;
Example translation means for receiving an input sentence in the first language and generating a plurality of translation sentence candidates in the second language for the input sentence with reference to the example base;
Statistical selection means for selecting and outputting a candidate whose probability score calculated using a predetermined probability statistical model satisfies a predetermined condition among the plurality of candidates generated by the example translation means Machine translation device. The example translation means includes:
Receiving the input sentence, the example having the sentence in the first language satisfying a predetermined similarity condition with the input sentence is searched in the example base, and a plurality of sentences each including the searched first language sentence Search means for extracting examples;
The correction means for correcting each sentence of the second language of the plurality of examples and generating a translation sentence candidate for the input sentence from each of the plurality of examples. Machine translation device. The search means receives the input sentence, searches the example base for a plurality of sentences in the first language having the smallest number of words unlike the input sentence, and searches the sentence in the first language. The machine translation apparatus according to claim 2, further comprising means for acquiring the plurality of examples including each. A bilingual dictionary between the first language and the second language;
The correcting means is
A difference identifying means for comparing the input sentence with the sentence in the first language of each of the plurality of examples, and identifying a difference from the input sentence for each of the plurality of examples;
The sentence of the second language of the plurality of examples is corrected by referring to the bilingual dictionary based on the difference identified by the difference identification unit, and the translation of the input sentence from each of the plurality of examples The machine translation apparatus according to claim 3, further comprising candidate generation means for generating sentence candidates. The bilingual dictionary may include a plurality of words in the second language as translations for one word in the first language,
The candidate generating means is configured to obtain one or a plurality of the sentence in the second language of each of the plurality of examples by referring to the bilingual dictionary based on the difference identified by the difference identifying means. The machine translation apparatus according to claim 4, comprising means for generating one or a plurality of translation sentence candidates for the input sentence by correcting each word using a second language word. The means for generating is obtained by referring to the bilingual dictionary based on each of the differences identified by the difference identification means for the sentence in the second language of each of the plurality of examples. 6. A means for generating one or more translation sentence candidates for the input sentence by modifying according to each possible combination using a plurality of words in the second language. The machine translation device described in 1. The search means receives the input sentence, searches the example base for a plurality of sentences in the first language that have the smallest editing distance with the input sentence, and searches for the sentence in the first language. The machine translation device according to claim 2, further comprising means for acquiring a plurality of examples each including The search means receives the input sentence, and selects a plurality of sentences in the first language that have a minimum editing distance from the input sentence calculated in consideration of a semantic distance between words in the example base. The machine translation device according to claim 2, further comprising means for retrieving and acquiring a plurality of examples each including the retrieved sentence in the first language. The statistical selection means includes means for selecting and outputting the highest probability score calculated using a predetermined probability statistical model among the plurality of candidates generated by the example translation means. The machine translation device according to any one of claims 1 to 8. And further comprising language model storage means for storing a language model of the second language,
The means for outputting is:
Language probability calculating means for calculating a language probability using the language model stored in the language storage means for each of the plurality of candidates;
The machine translation apparatus according to claim 9, further comprising: means for selecting and outputting a candidate having the highest language probability calculated by the language probability calculation means. A translation model storage unit for storing a translation model from the second language to the first language;
The means for outputting is:
For each of the plurality of candidates, a translation probability calculation means for calculating a translation probability using the translation model stored in the translation model storage means;
The machine translation apparatus according to claim 9, further comprising: means for selecting and outputting a candidate having the highest translation probability calculated by the translation probability calculation means. A language model storage means for storing a language model of the second language;
Translation model storage means for storing a translation model from the second language to the first language;
The means for outputting is:
Language probability calculating means for calculating a language probability using the language model stored in the language storage means for each of the plurality of candidates;
For each of the plurality of candidates, a translation probability calculation means for calculating a translation probability using the translation model stored in the translation model storage means;
Score calculating means for calculating a predetermined probability score as a function of the language probability calculated by the language probability calculating means and the translation probability calculated by the translation probability calculating means;
The machine translation apparatus according to claim 9, further comprising: means for selecting and outputting a candidate having the highest probability score calculated by the score calculation means. A machine translation computer program that, when executed by a computer, causes the computer to operate as the machine translation device according to any one of claims 1 to 12.

Images