JP5416021B2 - Machine translation apparatus, machine translation method, and program thereof - Google Patents

Description

  The present invention relates to a machine translation apparatus, a machine translation method, and a program for machine translation of an input sentence in one language into a sentence in a different language.

  In statistical machine translation (for example, Patent Documents 1 and 2) for obtaining a translated sentence in a different language (hereinafter referred to as “target language”) from an input sentence in a certain language (hereinafter referred to as “source language”). In order to search for all possible translations, it is known that the calculation amount of the power of the length of the input sentence (number of words) is required. When translating long source language sentences, the search range is generally used. To deal with by narrowing.

  Methods for narrowing the search range include a method called beam search that leaves only solution candidates with a score of a certain level or higher during the search process, and terminates further search of solution candidates that do not satisfy the level. A method is generally used that restricts the rearrangement to a certain range, thereby avoiding a full search for word order factorial word order. However, any of these methods has a problem that when the input sentence is long, the degree of approximation increases and the final translation quality deteriorates. In particular, in the latter method, a correct translation cannot be obtained in a translation between languages in which the word order is largely switched, such as translation from English to Japanese, and the translation quality may be significantly lowered.

  As a conventional technique for dealing with such a problem, there is a technique in which a sentence is divided into a plurality of partial strings, and the partial strings are translated and combined again (for example, Patent Document 3). By dividing the sentence, the above problem caused by the length of the sentence to be translated can be reduced. Patent Document 3 proposes a method of translating divided subsequences with a plurality of translation devices and generating a translated sentence from an optimal combination of subsequence translations. Specifically, sentence division and assembly are performed using rules created in advance.

JP 2006-99208 A JP 2008-15844 A JP 2001-222529 A

  In the method of Patent Document 3, it is necessary to manually create rules for dividing and assembling sentences, and in order to divide and assemble various sentences, it takes effort to define a large number of rules. . It is an object of the present invention to enable translation of input sentences divided into substrings in an appropriate order without using manually created rules, and to translate between languages with largely different word orders. The object is to provide a machine translation device, a machine translation method, and a program thereof that can improve the quality of a translation.

  The machine translation apparatus of the present invention includes a block parallel translation database, a translation training unit, a block translation model database, a block division unit, a translation unit, and a coupling unit.

  The block-divided parallel translation database is based on syntax analysis, each learning block is divided into a plurality of blocks each composed of one or more words and a non-terminal symbol indicating the insertion position of a lower block, and the divided sentence The block divided parallel translation composed of the ideal translation including the non-terminal symbol in the target language for each block is accumulated.

  The translation training unit learns a block translation model using the block division parallel translation read from the block parallel translation database, and writes it into the block translation model database.

  The block translation model database stores the block translation model.

  The block division unit divides an input sentence of the source language that is a translation target into the target language into a plurality of the blocks based on syntax analysis.

  Using the block translation model read from the block translation model database, the translation unit translates each block of the input sentence divided by the block division unit into a translated sentence including the non-terminal symbol in a target language. .

  The combining unit generates a translated sentence for the input sentence by combining the translated sentence of each block based on the block insertion position represented by the non-terminal symbol.

  According to the machine translation device, the machine translation method, and the program thereof according to the present invention, an object of the present invention is to translate a partial sequence without using a manually created rule in translating an input sentence divided into partial sequences. It can be combined in an appropriate order, and the quality of the translated sentence can be improved even in translation between languages with largely different word orders. Therefore, a long input sentence can be divided and efficiently translated, and even when the difference in the word order between the source language and the target language is large, a highly accurate translated sentence with an appropriate word order can be generated.

FIG. 2 is a block diagram illustrating a configuration example of a machine translation apparatus 100. The figure which shows the example of a processing flow of the machine translation apparatus. Example of parsing result. The figure which shows the storage image of the block division parallel translation sentence to the block division parallel translation database 110. FIG. Another example of parsing result. FIG. 2 is a block diagram showing a configuration example of a machine translation device 200. FIG. 2 is a block diagram showing a configuration example of a machine translation device 200. The figure which shows the storage image of the bilingual sentence to the bilingual sentence database 240. FIG. The figure which shows the matching image with each word of a target language, and the block of a source language. The figure which shows the storage image of the word related degree model to the word related degree database 220. FIG. The figure which shows the storage image of the word connection degree model to the word connection degree database 230. FIG.

  Hereinafter, embodiments of the present invention will be described in detail.

  FIG. 1 is a block diagram showing a configuration example of the machine translation apparatus 100 of the present invention, and FIG. 2 shows a processing flow example of the machine translation apparatus 100. The machine translation apparatus 100 includes a block translation database 110, a translation training unit 120, a block translation model database 130, a block division unit 140, a translation unit 150, and a combination unit 160.

  The block-divided parallel translation database 110 includes a learning sentence including a non-terminal symbol of the source language divided into a plurality of blocks based on a known general parsing technique, and a non-terminal symbol in the target language for each of the divided blocks. An ideal translation sentence and a block-divided parallel translation sentence made up of are accumulated. A block in the present invention refers to a unit composed of one or more consecutive words and a lower block. By defining in this way, the entire source language sentence is considered as the highest block, and the sentence can be expressed recursively and hierarchically in the form of a word and block sequence in the lower order. The block unit can be set to an arbitrary subtree on the parsing result (structural tree). However, in the present invention, it is desirable to use a section as a unit. Therefore, in the following, it is assumed that a section is a block unit. explain.

FIG. 3 is an example of the result of syntactic analysis for the English sentence “John bought a toy that was popular in Japan”. For syntax analysis, a well-known English parser Enju was used. S is a symbol indicating a clause other than a relative clause, and S-REL is a symbol indicating a relative clause. In the example of FIG. 3, it can be understood that the entire sentence can be considered as a clause, and a relative clause is embedded therein. When a substring in an English sentence expressed by S or S-REL is cut out in order from the top of the structure in FIG.
B0: John bought a toy [B1]
B1: that was popular in Japan
It is divided into two blocks. Here, [B1] included in the block B0 is a non-terminal symbol representing the position of the block B1 in the block B0. A block divided parallel translation is a combination of these blocks and an ideal translation including a non-terminal symbol in the target language. If Japanese is the target language, the ideal translation containing non-terminal symbols is
B0: John bought a toy [B1].

B1: It became popular in Japan, and each of these paired with each English sentence is a block-divided parallel translation. FIG. 4 shows a storage image of the block division parallel translation sentence in the block division parallel translation database 110. FIG. 4 is a translation between English and Japanese, where (a) is the English side and (b) is the Japanese side. In FIG. 4, __s0 and __s1 are non-terminal symbols, and when a plurality of blocks are inserted into one block, the non-terminal symbols are distinguished and expressed in this way, for example.

  The first embodiment is configured on the assumption that such a block division parallel translation is prepared in advance. A configuration in the case where it is not prepared will be described in a second embodiment.

  The translation training unit 120 learns the block translation model using the block division parallel translation including the non-terminal symbols read from the block division parallel translation database 110, and writes it into the block translation model database 130 (S1). For the learning of the block translation model, a known technique can be used. For example, a translation model learning program used together with the machine translation program Moses, a word translation probability estimation program GIZA ++, a word N-gram probability estimation program SRILM, or the like is used. Can do. Further, for optimization of the weight for each model, for example, a known technique called error rate minimization learning (Minimum Error Rate Training: MERT) may be used. In learning, in order to improve accuracy, a parallel sentence in a normal sentence unit that is not a block-divided parallel sentence may be used in combination. As the learning method in this case, existing methods (methods based on statistical models such as Patent Documents 1 and 2; methods based on dictionaries and rules such as Reference 1; and methods based on examples such as Reference 2) are used. it can.

[Reference document 1] Japanese Patent No. 3358096 [Reference document 2] Japanese Patent No. 4239505 By learning a block translation model using a block division parallel translation including a non-terminal symbol, only the translation relationship between the source language and the target language is obtained. Instead, the positional relationship between the position of the non-terminal symbol in the source language and the position of the non-terminal symbol in the target language is also learned. Therefore, by using this block translation model for the translation processing in the translation unit 150, a non-terminal symbol included in the input sentence of the source language block-divided by the block division unit 140 can be converted into an appropriate one in the target language in the block translation sentence. Can be placed in position.

  The block division unit 140 divides the input sentence of the source language to be translated into the target language into a plurality of blocks including non-terminal symbols based on a known general syntax analysis method (S2). In addition, since block division is performed on a word string, if the entire input sentence is a word string (for example, a sentence separated by using a blank character like English), it can be input as it is, but the input sentence is a character string. In the case (for example, Japanese) or a sentence including a part that is not divided into words, the source language word dividing unit 145 that divides the character string into word strings is provided in the preceding stage of the block dividing unit 140 as shown in FIG. It is necessary to provide. The division from the character string into the word string in the source language word dividing unit 145 can be performed using a known general morphological analysis method. An example of a Japanese morphological analysis program is Mecab.

  Using the block translation model read from the block translation model database 130, the translation unit 150 translates each block of the source language input sentence divided by the block division unit 140 into a translation sentence including a non-terminal symbol in the target language. (S3). Translation can be performed using a known machine translation technique (Patent Documents 1 to 3, Reference Documents 1 and 2, etc.).

  The combining unit 160 generates a translated sentence for the input sentence in the source language by combining the translated sentence of each block translated by the translation unit 150 based on the block insertion position represented by the non-terminal symbol (S4). .

  An example in which an input sentence in English shown below is translated into a translated sentence in Japanese using the machine translation apparatus 100 configured as described above will be described. Note that the block translation model is learned in advance.

Input sentence: we examined whether idiopathic pancreatitis is associated with CFTR mutations in persons who do not have lung disease of cystic fibrosis.
First, the block division unit 140 performs syntax analysis to obtain a syntax tree as shown in FIG. FIG. 5 shows an example of parsing by a known English parser Enju. Based on the analysis result, the block division unit 140 divides the input sentence into blocks including non-terminal symbols, using clauses as units of blocks. Since the nodes in the structure tree of FIG. 5 are S and S-REL, when a portion below the node of S or S-REL is divided as a block, the following is obtained.

1. We examined whether __s0.
2. idiopathic pacreatitis is associated with CFTR mutation in person __s0
3. who do not have lung disease of cystic fibrosis
In this division result, __s0 is a non-terminal symbol indicating the position where the next block in the list is inserted.

  Subsequently, the translation unit 150 translates the three blocks obtained as the output of the block division unit 140 into Japanese including non-terminal symbols using the block translation model stored in the block translation model database 130. The translation results when using the well-known machine translation program Moses are as follows.

1. We examined whether it was __s0.
2. In __s0 people, idiopathic pancreatitis is associated with CFTR mutations
3. No cystic fibrosis lung disease

  Then, the following translation is obtained by combining the three block translation results obtained as the output of the translation unit 150 based on the non-terminal symbols in the combining unit 160.

  Translation: We examined whether idiopathic pancreatitis was associated with CFTR mutations in people who did not develop cystic fibrosis lung disease.

  As can be seen from this translated sentence, the dependency relationship of the relative pronoun clause at the end of the input sentence is appropriately maintained, and the Japanese sentence has no sense of incongruity. On the other hand, when the above-mentioned input sentence is translated by a conventional machine translation device that does not perform block division, for example, the dependency relationship of the relative pronoun clause at the end of the input sentence may be lost as follows.

  Translated by the prior art: We were people with no cystic fibrosis lung disease who examined whether it was associated with idiopathic pancreatitis CFTR mutation.

  As described above, according to the machine translation apparatus 100 of the present invention, each block of an input sentence obtained by learning a block translation model from a block translation sentence including a non-terminal symbol and using the block translation model to include the non-terminal symbol. Translate. Therefore, in the translated text of each block, the non-terminal symbols are arranged at appropriate positions in the target language. Therefore, by simply combining the blocks based on the non-terminal symbols, the translated text in the appropriate word order without creating rules in advance. Can be obtained. Also, long modifier clauses can be rearranged in an appropriate order, improving translation quality.

  The first embodiment has a configuration on the premise that a block division parallel translation is prepared in advance, but it is necessary to create it when it is not prepared. The machine translation apparatus 200 according to the second embodiment is obtained by adding a configuration for creating a block-divided parallel translation to the configuration according to the first embodiment.

  FIG. 6 is a block diagram illustrating a configuration example of the machine translation apparatus 200, and FIG. 7 illustrates a process flow example of the machine translation apparatus 200. In addition to the components of the machine translation apparatus 100, the machine translation apparatus 200 includes a block division parallel translation unit creation unit 210, a word association degree model database 220, and a word connection degree model database 230.

  The block-divided parallel translation sentence is created from a parallel translation sentence for learning in which the translation is matched between the source language and the target language in sentence units. This bilingual sentence (a set of a source language learning sentence and an ideal translation sentence corresponding to the target language) is normally stored in the bilingual sentence database 240 in advance. FIG. 8 shows a storage image of the parallel translation sentence in the parallel translation database 240. FIG. 8 is a translation between English and Japanese, where (a) is the English side and (b) is the Japanese side.

  The block-divided bilingual sentence creation unit 210 performs processing by inputting the learning sentence of the source language divided into blocks and the word string of the ideal translation sentence in the target language of the learning sentence. Therefore, when the learning sentence of the source language is a character string or a sentence including a part that is not divided into words, the obtained word string is divided into word strings by the source language word dividing unit 145, and The block division unit 140 divides the data into blocks (S11), and inputs the block division parallel translation creation unit 210. In addition, when the ideal translation sentence in the target language is a character string or a sentence including a part that is not divided into words, the target language word dividing unit 245 divides the word into a word string, and generates a block division parallel translation creating unit. Input to 210. The target language word dividing unit 245 can divide the character string into word strings using a known general morphological analysis method, as with the source language word dividing unit 145.

The block split parallel translation creating unit 210 performs processing for associating each word in the input target language with each block in the source language, thereby blocking the word string in the target language and converting non-terminal symbols in the blocks in the source language. In the corresponding target language block, it is arranged at an appropriate position (S12). The process of associating each word in the target language with each block in the source language is as follows: the source language word string is F = f ₁ , f ₂ ,..., F _M , the number of blocks is K, and Block (k) (1 ≦ k ≦ K) represents a source language word included in the k-th block, = e ₁ word string target language E, e _2, · · ·, in e _N becomes translation, each e _1, e _2, .., E is equivalent to finding which block _N corresponds to.

  In the present invention, this problem is defined as a graph division problem as shown in FIG. 9 and solved by the same method as in Reference 3.

[Reference 3] X. Zhu, Z. Ghahramani, and J. Lafferty, "Semi-Supervised Learning Using Gaussian Fields and Harmonic Functions", Proceedings of the 20th International Conference on Machine Learning, 2003, p.912-919
Specifically, in FIG. 9, the source language of the block Block (k) (1 ≦ k ≦ K) each object language word _{e n (1 ≦ n ≦ N} ) constitutes a node respectively, and the node of the block There are branches (indicated by a thin solid line) connecting the nodes of the target language word and nodes (indicated by a thick solid line) of adjacent words in the target language. The graph of FIG. 9, if divided into K graphs each with nodes one block, or are determined each word e _n corresponds to which block. This is solved as an optimization problem of L in the following equation.

Here, w _ij is a weight of a branch connecting node i and node j, and v _i and v _j are k (1 ≦ k ≦ K) which are block IDs of node i and node j, respectively. If node i is a block node, v _i is the ID of that block, and the unknown is the ID of the block to which the N nodes of the target language word belong (ie, among the elements of the K + N-dimensional vector v, K is known and N is unknown). w _ij is the degree of relevance between the source language word in the block and the target language word that is the word node for the branch connecting the node of the block and the word node, and the target language for the branch connecting the word nodes Design to be the degree of word connection. In the machine translation apparatus 200, it is assumed that the word association degree is stored in the word association degree model database 220 as a word association degree model and the word connection degree is stored in the word connection degree model database 230 as a word connection degree model. . As the word relevance model and the word connection degree model, those separately learned from the parallel translation database 240 and other parallel translation databases are stored in advance. As the word relevance degree model, a known technique such as a word translation probability model used in a statistical machine translation system can be used. The word translation probability model can be configured, for example, as a list of word translation probabilities as shown in FIG. 10 obtained by a parallel translation database and a known word translation probability estimation program GIZA ++. In FIG. 10, each row is “English word”, “Japanese word”, “Conditional translation probability from Japanese word to English word”. Moreover, as a word connection degree model, what is based on well-known techniques, such as a word bigram model used with a statistical machine translation system, can be utilized. The word bigram model can be configured, for example, as a list of word bigram probabilities as shown in FIG. 11 obtained by the Japanese side of the parallel translation database and the known word N-gram probability estimation program SRILM. In FIG. 11, each row is “logarithm of conditional probability that the second word appears after the first word”, “first word”, “second word”, and “backoff probability”. By optimizing equation (1), the target language words will belong to the more strongly related blocks, and the target language words with a high degree of connection will belong to the same block. We can expect to obtain graph partitioning results suitable for the problem of block partitioning and mapping to the source language.

Based on the above definition, the block division parallel translation creating unit 230 obtains a vector v that optimizes the expression (1). Assuming that node numbers are assigned in the order of block nodes and word nodes, v is a vector obtained by concatenating a K-dimensional vector v _b related to block nodes and an N-dimensional vector v _w related to word nodes as shown in Equation (2). Can be expressed.

Similarly, a symmetric weight matrix W composed of branch weights w _ij can also be considered by dividing a part related to a block node and a word node (formula (3)).

In Equation (3), W _bb is the weight of the branch connecting the block nodes (in the present invention, the value is 0 because there is no branch between the block nodes), and W _bw and W _wb are the block node and the word node. The weight of the branch to be connected, W _ww is a matrix representing the weight of the branch connecting the word nodes.

Under the equations (2) and (3), the unknown part v _{w of} v that optimizes the equation (1) can be obtained by calculation of a matrix represented by the following equation according to the reference 3.

In equation (4), D _ww is a portion related to a branch connecting word nodes of a diagonal matrix D of (K + N) × (K + N) dimensions (similar to the relationship between W and W _{ww in} equation (3)), and Element d _i is

Is a matrix.

  Based on the above processing contents, an example of generating a block-divided parallel translation sentence by assigning a Japanese word of the parallel translation to each block of the English learning sentence in the block-divided parallel sentence creation section 230 will be shown.

  It is assumed that the bilingual sentence for learning read from the bilingual sentence database 240 is the following word string.

English: Although epidural corticosteroid injection are commonly used for sciatica, their efficacy has not been established.
Japanese: Epidural injections of corticosteroids are commonly used for sciatica, but their effectiveness has not been established.

  First, an English word string as a source language is divided into blocks in a form including non-terminal symbols in the block dividing unit 140 as follows.

1. Although __s0, __s1.
2. epidural corticosteroid injection are commonly used for sciatica
3. their efficacy has not been established
Then, in the block-divided bilingual sentence creation unit 230, by assigning each word of the bilingual Japanese sentence read from the bilingual sentence database 240 to the English sentence divided into blocks as described above, the block corresponding to the English sentence of each block Divided Japanese sentences are obtained as follows.

1. __s0 is __s1.
2. Epidural injection of corticosteroid is commonly used for sciatica
3. Its effectiveness has not been established

  As described above, according to the machine translation apparatus 200, even if a block-divided parallel translation is not prepared in advance, the translation is handled in units of sentences used for learning examples of existing techniques and statistical models. The block-divided parallel translation sentence can be generated from the parallel translation sentence, and the processing content by the machine translation apparatus 100 shown in the first embodiment can be executed using this.

  It should be noted that the function sharing of each component of the machine translation devices 100 and 200 of the present invention is not limited to the function sharing shown in the above embodiment, and can be appropriately changed without departing from the present invention. In addition, the processing of each step in the machine translation method of the present invention is not only executed in the time series described above, but also executed in parallel or individually as required by the processing capability of each constituent element that executes the processing. It is good to do.

Claims (7)

Based on the syntax analysis, the learning sentence in the source language divided into a plurality of blocks each consisting of one or more words and a non-terminal symbol representing the insertion position of the lower block based on the parsing, and the target language for each of the divided blocks An ideal translated sentence including a non-terminal symbol, and a block division parallel translation database in which block division parallel translations are stored;
A translation training unit that learns a block translation model using the block-divided parallel translation read from the block parallel translation database and writes the block translation model in the block translation model database;
A block translation model database in which the block translation model is stored;
Parse the source language input sentence to be translated into the target language, separate the tree structure at any selected node, and add a non-terminal symbol representing the selected node to the tree structure. A block divider for dividing the block into a number of blocks;
Using the block translation model read from the block translation model database, a translation unit that translates each block of the input sentence divided by the block division unit into a translation sentence including the non-terminal symbol in a target language,
A combining unit that generates a translated sentence for the input sentence by combining the translated sentence of each block based on a block insertion position represented by the non-terminal symbol;
A machine translation apparatus comprising: The machine translation device according to claim 1,
The block dividing unit further divides a learning sentence in a source language into a plurality of the blocks based on syntax analysis,
A word relevance model database in which a word relevance model indicating a relevance degree between a word included in a learning sentence in a source language and a word included in an ideal translation sentence in a target language of the learning sentence in the source language;
A word adjacency model database in which word adjacency models indicating the degree of articulation between words included in the ideal translation of the target language of the source language learning sentence are stored;
The learning sentence of the source language divided into the block units and the ideal translation sentence in the target language of the learning sentence are input, and the word association degree model read from the word association degree model database and the word connection degree model database Using the read word connectivity model, segment the source language learning sentence by identifying which block in the source language learning sentence corresponds to each word included in the target language ideal translation sentence. A block division parallel translation creation unit that generates an ideal translation including the non-terminal symbol in the target language for each block and writes the obtained block division parallel translation to the block division parallel translation database;
A machine translation device further comprising:   3. The machine translation apparatus according to claim 1, wherein the block is a node unit. For each divided block, the translation training section is divided into a plurality of blocks each of which is composed of one or more words and a non-terminal symbol representing the insertion position of the lower block based on the syntax analysis A translation training step of reading a block division parallel translation sentence including the ideal translation sentence including the non-terminal symbol in the target language from the block division parallel translation database, learning a block translation model using this, and writing it into the block translation model database; ,
The block division unit parses the input sentence of the source language to be translated into the target language, separates the tree structure at any selected node, and adds a non-terminal symbol representing the selected node to the tree structure. addition to, a block dividing step of dividing said blocks of multiple,
A translation unit translates each block of the input sentence divided in the block division step into a translation sentence including the non-terminal symbol in a target language using a block translation model read from the block translation model database Steps,
A combining step for generating a translated sentence for the input sentence by combining the translated sentence of each block based on a block insertion position represented by the non-terminal symbol;
Perform machine translation method. The machine translation method according to claim 4,
A learning sentence block dividing step of dividing a learning sentence of the source language into a plurality of the blocks based on syntax analysis;
The words and source language learning sentences included in the source language learning sentences read from the word relevance model database from the source language learning sentences divided into block units and the ideal translation sentences in the target language of the learning sentences. Between the words included in the ideal translation in the target language of the source language learning sentence read from the word connection model database and the word relevance model indicating the degree of association with the words included in the ideal translation in the target language By using a word-jointness model that indicates the degree of connection, it is possible to determine which block of the source language learning sentence each word contained in the ideal translation sentence in the target language corresponds to, respectively. An ideal translation including the non-terminal symbol in the target language for each of the divided blocks is generated, and the obtained block division parallel translation is stored in the block division parallel translation database. Machine translation method to further execute a block division bilingual sentence creating step burn them.   6. The machine translation method according to claim 4, wherein the block is a node unit.   A program for causing a computer to function as the machine translation device according to claim 1.

Images