JP6613666B2 - Word rearrangement learning device, word rearrangement device, method, and program - Google Patents

Description

  The present invention relates to a word rearrangement learning device, a word rearrangement device, a method, and a program, and more particularly to a word rearrangement learning device, a word rearrangement device, a method, and a program for rearranging words in a source language sentence. .

  The machine translation process from language A to language B is performed by translating a phrase of language A (hereinafter referred to as the source language) into a phrase of language B (hereinafter referred to as the target language) and a translated target language phrase in the target language. Broadly divided into two with appropriate sorting. Statistical translation technology widely used in this field is a statistical model of word translation and word sorting based on the correspondence between words in the target language and words in the target language estimated from a large number of parallel translations. In other words, a method of searching for a target language translation sentence constituted by a probable phrase translation and phrase rearrangement based on a statistical model of the source language input sentence is employed. In general, it is extremely difficult to comprehensively search for all translation candidates, so the number of translation candidates for each word is limited, and the distance of word sorting is limited within a certain range. By doing so, machine translation processing with a practical calculation amount is realized. However, depending on the combination of the source and target languages to be translated, the corresponding phrases may appear in very different orders, and the sort is large enough to accurately translate between such languages. Since translation processing in consideration of distance is required, there is a problem that an increase in calculation amount is unavoidable.

  As a technique for dealing with this problem, there is a technique called “pre-ordering” that rearranges words in the source language so as to approach the order of the corresponding words in the target language before performing the translation process.

  Non-Patent Document 1 is intended for translation from German to English, and Patent Document 1 is intended for translation from English to Japanese.The language of the input sentence (source language) is changed to the order of the corresponding words in the translated language (target language). It uses a rule that rearranges them closer to each other. While these technologies can be sorted fairly accurately by using source language parsing and appropriate rules, the rules that are required differ depending on the source and target languages. Must be defined.

  Non-patent document 2 performed using a statistical model is an example of a prior rearrangement method that can be realized regardless of language.

  In Non-Patent Document 2, in addition to the label of the parent node and child node, the main words and the words at both ends of the partial syntax tree with the child node as the vertex, and the word adjacent to the partial syntax tree are the features (features). By using the entropy identification model, it is determined whether or not to change the order of the child nodes.

  In the method described in Non-Patent Document 2 that implements language-independent pre-ordering, a pattern or learning data on whether or not to change the order of child nodes on the syntax tree of the source language is manually assigned to the bilingual data Alternatively, it is acquired based on the word association information automatically estimated by the methods shown in Non-Patent Document 3 and Non-Patent Document 4. There is no overlap between the target language side partial word string A corresponding to the left child node and the target language side partial word string B corresponding to the right child node, and the target language side partial word string A is more than the partial word string B. If it is on the left side, it can be determined that rearrangement of the child nodes is unnecessary, and if it is on the right side, rearrangement of the child nodes is necessary.

JP 2011-175500 A

Michael Collins et al., “Clause Restructuring for Statistical Machine Translation”, In Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, pp. 531-540, 2005. Chi-Ho Li et al., “A Probabilistic Approach to Syntax-based Reordering for Statistical Machine Translation,” Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics, pp. 720-727, 2007. Peter F. Brown et al., “The Mathematics of Statistical Machine Translation: Parameter Estimation,” Computational Linguistics, pp. 268-311, 1993. Jason Riesa et al., “Feature-Rich Language-Independent Syntax-Based Alignment for Statistical Machine Translation,” Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pp. 497-507, 2011. Nang Yang et al., “A Ranking--based Approach to Word Reordering for Statistical Machine Translation”, Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics, pp. 912-920, 2012.

  The word correspondence in the parallel translation may have a one-to-many or many-to-many relationship, and an erroneous correspondence is likely to be given in the automatically estimated word correspondence, so that there are many overlaps in the partial word strings on the target language side. Can occur.

  Using the word association information of Non-Patent Document 5, the number of intersections in word order is calculated, and a rearrangement model that realizes rearrangement so that the number of intersections becomes smaller is used. However, since the amount of calculation for finding the element rearrangement that minimizes the number of intersections increases exponentially with the increase in the number of elements, the amount of calculation is reduced by adding constraints in the possible rearrangement order. .

  Non-Patent Document 2 eliminates overlap by removing low-probability one-to-many and many-to-many word correspondences, but considering the ambiguity of the language, even word correspondences with relatively low probabilities are not an error In addition, there is a possibility that the overlap is not eliminated even by removing the word correspondence and the number of learning examples is reduced.

  The present invention has been made in view of the above circumstances, and learns a word rearrangement model that can appropriately determine prior word rearrangement even when there is an overlap in word association. An object of the present invention is to provide a word rearrangement learning device, method, and program.

  Moreover, it aims at providing the word rearrangement apparatus, method, and program which can determine the prior rearrangement of a word appropriately.

  In order to achieve the above object, the word rearrangement learning device according to the present invention performs syntactic analysis of the source language for each of the source language sentences included in a plurality of pairs of source language sentences and target language sentences to be translated. A syntactic analysis unit that generates a syntax tree of the source language sentence, and for each of the plurality of pairs, based on the syntax tree of the source language sentence generated by the syntactic analysis unit, the source language sentence of the pair And an automatic word associating unit for associating words between the target language sentences, and a plurality of words of the target language sentence as the source language based on the result of word association by the automatic word associating part If the word is associated with at least one word of the sentence, the result of the word association is a central word of the plurality of words of the target language sentence, and at least one word of the source language sentence; Vs And, for each of the plurality of pairs, for each node having two child nodes in the syntax tree of the source language sentence of the pair, two child nodes obtained using the result of the word association A word string of the target language sentence corresponding to the word string of the source language sentence corresponding to the word string of the source language sentence represented by the two child nodes whose order has been reversed. And a word rearrangement correct determination unit for determining whether to reverse the order of the two child nodes, and the word rearrangement correct determination unit for each of the plurality of pairs. A feature amount extracting unit that extracts a feature amount for each node for which the determination has been made, a correct answer for each node determined for each of the plurality of pairs by the word rearrangement correct answer determining unit, and the feature Based on the feature quantity for each node extracted for each of the plurality of pairs by the extraction unit, the order of the two child nodes is determined for a node having two child nodes in the syntax tree of the source language sentence. A word rearrangement model learning unit that learns a word rearrangement model for determining whether to invert or not.

  A word rearrangement learning method according to the present invention includes a word rearrangement learning apparatus including a syntactic analysis unit, an automatic word association unit, a word rearrangement correct answer determination unit, a feature amount extraction unit, and a word rearrangement model learning unit. In the replacement learning method, the syntactic analysis unit performs a syntactic analysis of the source language for each of the source language sentences included in a plurality of pairs of source language sentences and target language sentences to be translated, and the source language sentences And the automatic word association unit generates, for each of the plurality of pairs, a source language sentence and an object of the pair based on the syntax tree of the source language sentence generated by the syntactic analysis unit Words are associated with each other between language sentences, and the word rearrangement corrective determination unit determines that the plurality of words of the target language sentence are the original words based on the result of word association by the automatic word association unit. language If the word is associated with at least one word, a central word of the plurality of words of the target language sentence and at least one word of the source language sentence are obtained as a result of the word association. Two child nodes obtained by using the result of the word association for each of the plurality of pairs, each node having two child nodes in the syntax tree of the source language sentence of the pair. A word string of the target language sentence corresponding to the word string of the source language sentence corresponding to the word string of the source language sentence represented by the two child nodes whose order has been reversed. And determining whether or not to reverse the order of the two child nodes, and the feature amount extraction unit determines whether the correct word determination unit determines the correct answer for each of the plurality of pairs. For each node determined, and the word rearrangement model learning unit determines the correct answer for each node determined for each of the plurality of pairs by the word rearrangement correct answer determination unit; The two child nodes for a node having two child nodes in a syntax tree of a source language sentence based on the feature amount for each node extracted for each of the plurality of pairs by the feature amount extraction unit A word rearrangement model for determining whether or not to reverse the order of is learned.

  The word rearrangement device according to the present invention includes a syntactic analysis unit that performs a syntactic analysis of a source language based on an input source language sentence and generates a syntax tree of the source language sentence, and a syntax of the source language sentence A feature amount extraction unit that extracts a feature amount for each node having two child nodes in the tree, and a node that has two child nodes in the syntax tree of the source language sentence extracted by the feature amount extraction unit. Based on the feature quantity and the word rearrangement model learned by the word rearrangement learning device, the order of the two child nodes is reversed for a node having two child nodes in the syntax tree of the source language sentence. And a word rearrangement determining unit that rearranges words of the source language sentence based on the determination result.

  A word rearrangement method according to the present invention is a word rearrangement method in a word rearrangement device including a syntactic analysis unit, a feature amount extraction unit, and a word rearrangement determination unit, wherein the syntactic analysis unit receives the input original Based on the language sentence, syntactic analysis of the source language sentence is performed to generate a syntax tree of the source language sentence, and the feature amount extraction unit is configured to perform processing for each node having two child nodes in the syntax tree of the source language sentence. A feature amount is extracted, and the word rearrangement determination unit extracts the feature amount for each node having two child nodes in the syntax tree of the source language sentence extracted by the feature amount extraction unit, and the word arrangement described above Based on the word rearrangement model learned by the replacement learning method, it is determined whether or not to reverse the order of the two child nodes for a node having two child nodes in the syntax tree of the source language sentence. And, on the basis of the result of the decision, carry out the sort of word of the source language sentence.

  Moreover, the program of this invention is a program for functioning a computer as each part which comprises said word rearrangement learning apparatus.

  Moreover, the program of this invention is a program for functioning a computer as each part which comprises said word rearrangement apparatus.

  As described above, according to the word rearrangement learning device, method, and program of the present invention, based on the result of word association, a plurality of words in the target language sentence are at least one word in the source language sentence. Is associated with at least one word of the source language sentence and has two child nodes in the syntax tree of the source language sentence. For each node, use the result of word association to determine the correct answer of whether to reverse the order of the two child nodes, and to the two children for the node having two child nodes in the syntax tree of the source language sentence By learning the word rearrangement model to determine whether to reverse the order of the nodes, even if there is an overlap in word mapping, determine the preordering of words appropriately It is possible.

  Further, according to the word rearrangement device, method, and program of the present invention, based on the feature amount for each node having two child nodes in the syntax tree of the source language sentence and the learned word rearrangement model By determining whether to reverse the order of the two child nodes for a node having two child nodes in the syntax tree of the source language sentence and rearranging the words in the source language sentence, Sorting can be determined.

It is a block diagram which shows the functional structure of the word rearrangement learning apparatus which concerns on embodiment of this invention. It is a figure which shows the example of a syntax tree. It is a block diagram which shows the functional structure of the word rearrangement apparatus which concerns on embodiment of this invention. It is a block diagram which shows the functional structure of the machine translation learning apparatus which concerns on embodiment of this invention. It is a block diagram which shows the functional structure of the machine translation apparatus which concerns on embodiment of this invention. It is a flowchart figure which shows the word rearrangement learning process routine in the word rearrangement learning apparatus which concerns on embodiment of this invention. It is a flowchart figure which shows the word rearrangement process routine in the word rearrangement apparatus which concerns on embodiment of this invention. It is a flowchart figure which shows the machine translation learning process routine in the machine translation learning apparatus which concerns on embodiment of this invention. It is a flowchart figure which shows the machine translation processing routine in the machine translation apparatus concerning embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Outline of Embodiment of the Present Invention>
In the embodiment of the present invention, it is defined that the word correspondence overlap in learning of pre-ordering using the syntax tree of the source language is associated with the central word of the target language side word group corresponding to a word in the source language. To solve it. In other words, the problem that it is difficult to determine whether or not to rearrange child nodes according to one-to-many and many-to-many correspondence in word association is avoided by using the central word.

  In addition, a second automatic word association is used that uses the result of automatic word association obtained by previously removing words in the source language and the target language in which there is no word correspondence between the source language and the target language as a feature quantity. By doing this, the accuracy of automatic word matching, which is a precondition for learning of pre-ordering, is improved, and more accurate pre-sorting is possible. Japanese is an example of the source language, and English is an example of the target language.

<Configuration of word rearrangement learning device according to embodiment of the present invention>
Next, the configuration of the word rearrangement learning device according to the embodiment of the present invention will be described. As shown in FIG. 1, a word rearrangement learning device 100 according to an embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program and various data for executing a word rearrangement learning processing routine described later. And a computer including Functionally, the word rearrangement learning device 100 includes an input unit 10, a calculation unit 20, and an output unit 90 as shown in FIG.

  The input unit 10 accepts input of a set of pairs of a source language sentence and a target language sentence, which are parallel translation sentences, as learning data for word rearrangement. In addition, the input unit 10 receives input of word association information in a pair of a source language sentence and a target language sentence that are parallel translation sentences.

  The calculation unit 20 includes a source language sentence database 22, a target language sentence database 24, a syntactic analysis unit 26, an automatic word association unit 32, a word association information database 42, a second automatic word association model learning unit 44, and a second automatic word. An association model output unit 46, a second automatic word association model 48, a word rearrangement correct answer determination unit 50, a feature amount extraction unit 52, a word rearrangement model learning unit 54, and a word rearrangement model 56. .

  The source language sentence database 22 stores a set of parallel language sentences received by the input unit 10.

  The target language sentence database 24 stores a set of target language sentences of parallel translations received by the input unit 10.

  The syntactic analysis unit 26 includes a source language syntactic analysis unit 28 and a target language syntactic analysis unit 30.

  The source language syntactic analysis unit 28 performs a syntactic analysis of the source language for each source language sentence stored in the source language sentence database 22 and generates a syntax tree of the source language sentence. Processing in the source language syntactic analysis unit 28 may include word division and part-of-speech assignment. The syntactic analysis method may be a known technique, for example, software such as Berkeley Parser or Enju for English, and software such as Haruniwa or Ckylark for Japanese, but the configuration in the embodiment of the present invention is a specific syntactic analysis technique. Does not depend on. In the embodiment of the present invention, it is assumed that the syntax tree is a binary tree. Therefore, when the result of syntactic analysis is a polyary tree, the tree is converted to a binary tree by a known technique. .

  The target language syntactic analysis unit 30 performs syntactic analysis of the target language for each target language sentence stored in the target language sentence database 24, and generates a syntax tree of the target language sentence. The method of syntactic analysis is the same as that of the source language syntactic analysis unit 28.

  The automatic word association unit 32 determines, based on the syntax tree of the source language sentence and the syntax tree of the target language sentence generated by the syntactic analysis unit 26, for each pair of the source language sentence and the target language sentence of the parallel translation sentence. A word is associated between a pair of source language sentence and target language sentence. The automatic word association unit 32 includes a first automatic word association unit 34 and a second automatic word association unit 40.

  The first automatic word association unit 34 includes a language information editing unit 36 and a first word association unit 38.

  The language information editing unit 36 uses the source language sentence and the target language sentence syntax tree generated by the syntactic analysis unit 26 for each of the paired source language sentence and target language sentence pair, based on the source language sentence and the target language sentence syntax tree. In order to reduce the difference in word or word order between the target language and the target language, the word that does not correspond between the source language and the target language can be changed according to the syntactic structure, part of speech, and surface layer information. Delete from the syntax tree of the language sentence, supplement words in the other language, or rearrange at least one word order of the source language sentence and the target language sentence by an existing word rearrangement method. In the configuration using the word association result of the first automatic word association unit 34 as a feature amount, the word association accuracy in the second automatic word association unit 40 is increased by reducing the difference in words and word order. Done for. For example, when matching English and Japanese words, the English articles “a”, “an”, “the” and the Japanese case particles “ha”, “ga”, “wa”, etc. In many cases, the corresponding word does not exist, and an error in word correspondence is likely to occur. Therefore, it is expected that the accuracy of matching other words can be improved by removing in advance. Further, by utilizing the rearrangement for bringing English closer to Japanese and the method for adding words to English as described in Patent Document 1, the difference in word order in the bilingual sentence is reduced, and the improvement of word matching accuracy is expected. it can.

  The first word associating unit 38 includes a source language sentence and a target whose language information editing unit 36 has removed or added words or rearranged words for each pair of the source language sentence and the target language sentence of the parallel translation sentence. Based on the language sentence, word association is performed for the parallel translation sentence whose language information has been edited. As a word association method, a known technique, for example, software such as GIZA ++ that performs association using a model described in Non-Patent Document 3 can be used. Further, since the word association result obtained by the first word association unit 38 is for the source language sentence and the target language sentence edited by the language information editing unit 36, the source language sentence and the target language sentence before being edited. It is necessary to replace with the correspondence in. In the present embodiment, the correspondence information can be easily read by storing information on word removal / addition and word rearrangement in the language information editing unit 36.

  The second automatic word association unit 40 uses a second automatic word association model, which will be described later, and the result of the first automatic word association unit 34, and the source language sentence of the bilingual sentence for which no word association information is given and For each target language sentence pair, word association is performed for the pair.

  It is preferable to use the result of the first automatic word association unit 34 as the feature amount. For example, an association score is added to a pair of words associated as a result of the first automatic word association unit 34. Further, as a simple configuration, a configuration using only the result of unsupervised learning of the first automatic word association unit 34, or the result of unsupervised learning of the first automatic word association unit 34 is used as the second automatic word association unit. The structure which is not used in 40 may be used. When the configuration using only the result of unsupervised learning of the first automatic word association unit 34 is used, the result of the unsupervised learning method of non-patent document 3 is used as the association result, and supervised learning of non-patent document 4 is performed. Should not be used. In the configuration in which the result of unsupervised learning by the first automatic word association unit 34 is not used in the processing of the second automatic word association unit 40, the method of Non-Patent Document 4 is used, but the unsupervised learning result is characterized. Do not use it as a quantity. Further, in the case of the above simple configuration, the first and second two-stage automatic word association is not necessary. In this case, the automatic word association unit 32 may include only one of the first automatic word association unit 34 and the second automatic word association unit 40.

  The word correspondence information database 42 stores word correspondence information given to each of at least one pair of the source language sentence and the target language sentence of the parallel translation sentence.

  The second automatic word association model learning unit 44 uses the supervised learning based on the word association result by the first automatic word association unit 32 and the word association information stored in the word association information database 42. Learn the association model. For learning the word association model, a known technique, for example, a method described in Non-Patent Document 4 can be used. For learning, a part to which word correspondence information is given is selected and used from a plurality of parallel translation sentences.

  As the feature quantity for each word pair that is a candidate for mapping, the surface layer of the words in the source language or the target language, parts of speech, syntactic labels, syntactic structure, other mapping restrictions specialized for numbers and symbols, GIZA ++, etc. were used. A word association result by unsupervised learning of the first automatic word association unit 32 is used. In particular, in the present embodiment, the word association result that has been processed by the language information editing unit 36 is used as the word association result by unsupervised learning. It should be noted that a plurality of word association results by unsupervised learning are used, for example, the word association results when non-corresponding word removal is performed and the word association results when not performed are used as different feature amounts. Is possible. Alternatively, the model may be learned by using a parameter update algorithm with constraints described in Non-Patent Document 6 using separately provided word correspondence information.

[Non-patent document 6]: David Talbot, “Constrained EM for Parallel Text Alignment,” Natural Language Engineering, vol.11 pp. 263-277, 2005.

  The second automatic word association model output unit 46 outputs the word association model learned by the second automatic word association model learning unit 44 and stores it in the second automatic word association model 48.

  The word rearrangement correct answer determination unit 50, for each syntax tree of the source language sentence generated by the syntactic analysis unit 26, based on the result of the word association by the automatic word association unit 32, Among the nodes having two child nodes above, the node that is rearranged to identify the node that approaches the word order of the target language is identified, and the correct answer in learning by the word rearrangement model learning unit 54 described later is determined. The word correspondence information used in the second automatic word association model learning unit 44 may also be used.

  Hereinafter, the principle of identifying nodes that are rearranged to approach the target language in the word order will be described.

First, as a reference for approaching the word order of the target language, the order correlation between the word order of the source language sentence and the word order of the target language sentence corresponding to each word of the source language sentence is used. The rank correlation calculation will be described based on the example of FIG. The English (source language) syntax tree shown in Example 1 (the underline corresponding to the node of the tree is the part-of-speech or phrase label, the word corresponding to the leaf is a word, the number is an ID starting from 0), Japanese (Target language) word string (numbers are IDs starting from 0), and integer vectors in which the IDs of the corresponding target language sentences are listed in order from the beginning of the words of the source language sentence in the correspondence of words represented by broken lines think of. Here, when a plurality of words of the target language sentence are associated with the words of the source language sentence, the ID stored in the vector is the median value of the IDs of the words of the corresponding target language sentence (element The method of determining the median when the number is even is not specified, but the average of the elements immediately before and after the center is the median, so the average may be used, or the value immediately before or after the center may be substituted. Also good). In Example 1, the median value of the word corresponding to the first “is” is 2 (when the median value for even elements is immediately before the center), and so on, followed by 0 and 1, so v _m = [2, 0, 1] vector. At the same time, consider the case where the order of the two child nodes is reversed. Example 2 shows an example in which the child nodes (NP and VBZ) of the root node (VP) are inverted. The integer vector in this case is v _w = [0, 1, 2]. The difference in rank correlation coefficient between the case where the order of child nodes is not reversed and the case where the order of child nodes is not reversed is calculated.

D = R (v _m ) − R (v _w ) (1)

  If the difference is positive, it is closer to the target language's word order if the order of the child nodes is not reversed. Therefore, between these child nodes, "Do not invert" is determined as the correct answer, and when the difference is negative, the order of the child nodes is inverted. Since it is closer to the word order of the target language, “invert” between the child nodes is determined as the correct answer. If the difference becomes 0, the order is indefinite, and is excluded from the learning data in the next word rearrangement model learning unit 54.

  The rank correlation scale R is preferably a rank correlation coefficient called Kendall's tau (τ) shown in the following equation.

(2)

  In the above formula (2), v is an integer vector, | v | is the number of elements of v, and P is the number of elements in an ascending order among a set of arbitrary two elements of v.

  For each syntax tree of the source language sentence generated by the syntactic analysis unit 26, the feature quantity extraction unit 52 determines the correct answer for the rearrangement of the child nodes by the word rearrangement correct answer determination unit 50. For each node on the syntax tree, a feature quantity for determining whether or not to rearrange is extracted. The details of the feature amount are not specified in the present embodiment, but the leftmost and rightmost words of the subtree having the vertices at the left and right child nodes, the labels of the parent node and the child node described in Non-Patent Document 2. In addition to the main word, the combination of labels of child nodes, the concatenation of leaf words of child nodes, the combination of labels of parent nodes, child nodes, grandchild nodes, etc. This is suitable for rearranging the accuracy.

  The word rearrangement model learning unit 54 is obtained by the rearrangement correct answer for each node and the feature quantity extraction unit 52 for each syntax tree of the source language sentence whose correct answer is determined by the word rearrangement correctness determination unit 50. The word rearrangement model is learned by using the feature amount for each node. A learning method is not particularly defined, but a known binary or multi-level classification technique such as a support vector machine or a maximum entropy identification model can be used.

  The word rearrangement model 56 stores the word rearrangement model learned by the word rearrangement model learning unit 54. The word rearrangement model stored in the word rearrangement model 56 is output by the output unit 90.

<Configuration of word rearrangement device>
Next, the configuration of the word rearrangement device according to the embodiment of the present invention will be described. As shown in FIG. 3, the word rearrangement device 200 according to the embodiment of the present invention includes a CPU, a RAM, a ROM for storing a program and various data for executing a word rearrangement processing routine described later, It can comprise with the computer which includes. Functionally, the word rearrangement device 200 includes an input unit 210, a calculation unit 220, and an output unit 290 as shown in FIG.

  The input unit 210 receives input of a source language sentence that is a word rearrangement target from an input device such as a keyboard. Note that the input unit 210 may accept input from the outside via a network or the like.

  The calculation unit 220 includes a syntactic analysis unit 226, a feature amount extraction unit 252, a word rearrangement determination unit 254, and a word rearrangement model 256.

  The syntactic analysis unit 226 includes a source language syntactic analysis unit 228. Similar to the source language syntactic analysis unit 28 in the word rearrangement learning device 100, the source language syntactic analysis unit 228 syntactically analyzes the input source language sentence and generates a syntax tree of the source language sentence.

  Based on the syntax tree of the source language sentence generated by the syntactic analysis unit 226, the feature quantity extraction unit 252 performs a child operation on the syntax tree of the source language sentence, similarly to the feature quantity extraction unit 52 of the word rearrangement learning device 100. For each node having two nodes, a feature quantity for word rearrangement is extracted.

  The word rearrangement model 256 stores the same word rearrangement model as the word rearrangement model 56 of the word rearrangement learning device 100.

  The word rearrangement determining unit 254 obtains the word rearrangement model stored in the word rearrangement model 256 and the feature amount extraction unit 252 for each node having two child nodes on the syntax tree of the source language sentence. Based on the determined feature amount, it is determined whether or not the child nodes are rearranged for the node. Although the determination method is not particularly defined, a known technique according to the method used for learning the word rearrangement model can be used.

  The word rearrangement determination unit 254 reflects the rearrangement determination result in the syntax tree, and outputs the result to the storage medium or terminal by the output unit 290 in the form of a word string or a syntax tree.

<Configuration of machine translation learning device>
Next, the configuration of the machine translation learning device according to the embodiment of the present invention will be described. As shown in FIG. 4, the machine translation learning device 300 according to the embodiment of the present invention includes a CPU, a RAM, a ROM that stores a program and various data for executing a machine translation learning processing routine described later, It can comprise with the computer which includes. Functionally, the machine translation learning apparatus 300 includes an input unit 310, a calculation unit 320, and an output unit 390 as shown in FIG.

  The input unit 310 accepts input of a set of pairs of source language sentences and target language sentences, which are parallel translation sentences, as learning data for machine translation. However, each of the source language sentences is a word rearranged by the word rearrangement device 200 so as to be closer to the target language.

  The input unit 310 receives an input of a set of target language sentences.

  The input unit 310 also accepts input of a set of pairs of source language sentences and target language sentences, which are parallel translation sentences, as learning data for model weight adjustment. However, each of the source language sentences is a word rearranged by the word rearrangement device 200 so as to be closer to the target language.

  The calculation unit 320 includes a parallel translation original language sentence database 322, a parallel translation target language sentence database 324, a word association part 326, a translation model learning part 328, a target language sentence database 330, a language model learning part 332, a translation model 334, and a language model 336. , A weight adjustment parallel translation database 338, a weight adjustment unit 340, and a model weight storage unit 342.

  The bilingual source language sentence database 322 stores a set of source language sentences of the bilingual sentences accepted by the input unit 310.

  The bilingual target language sentence database 324 stores a set of target language sentences of the bilingual sentences received by the input unit 310.

  The word association unit 326 performs word association between the source language sentence and the target language sentence of each pair of the source language sentence and the target language sentence that are parallel translations. As a method of word association, the same method as the first automatic word association unit 34 or the second automatic word association unit 40 in the word rearrangement learning device 100 may be used, or another method may be used. Similarly, a known technique can be used for the association.

  The translation model learning unit 328 translates the source language phrase into the target language phrase based on the result of the word association by the word association unit 326 for each pair of the source language sentence and the target language sentence that are parallel translation sentences. Learn the translation model that calculates the probability. A known technique, for example, the method of Non-Patent Document 7, can be used for model learning. Also, a method of directly learning a translation model from bilingual data without passing through word association, for example, the method of Non-Patent Document 8 may be used.

[Non-Patent Document 7]: Phillip Koehn et al., “Statistical Phrase-based Translation,” Proc. HLT-NAACL, pp. 263-270, 2003.
[Non-Patent Document 8]: Graham Neubig et al., “An Unsupervised Model for Joint Phrase Alignment and Extraction,” Proc. ACL, pp. 632-641, 2011.

  The target language sentence database 330 stores a set of target language sentences received by the input unit 310.

  The language model learning unit 332 learns the language model of the target language based on the set of target language sentences stored in the target language sentence database 330. Although the type of language model and its learning method are not particularly defined, a known word N-gram language model and various learning methods thereof can be used.

  The translation model 334 stores a translation model learned by the translation model learning unit 328.

  The language model 336 stores a language model learned by the language model learning unit 332.

  The weight adjustment parallel translation database 338 stores a set of pairs of source language sentences and target language sentences, which are parallel translations, received by the input unit 310.

  The weight adjustment unit 340 generates a translation model based on a set of target language sentences stored in the target language sentence database 330, a translation model stored in the translation model 334, and a translation model stored in the language model 336. And adjust the weight for each of the language models.

  When machine translation is performed using a plurality of statistical models, improvement in translation accuracy can be expected by setting an appropriate weight for each model. For weight adjustment, a translation result obtained when a source language sentence for weight adjustment is translated using a known technique, for example, a parallel translation sentence for weight adjustment described in Non-Patent Document 9, is used for weight adjustment. It is possible to use a method of repeatedly performing the process of updating the weight so as to approach the target language sentence.

[Non-Patent Document 9] Franz Josef Och, “Minimum Error Rate Training in Statistical Machine Translation,” Proc. ACL, pp. 160-167, 2003.

  The model weight storage unit 342 stores the weight for each of the translation model and the language model adjusted by the weight adjustment unit 340.

  The output unit 390 outputs the translation model stored in the translation model 334, the translation model stored in the language model 336, and the weight stored in the model weight storage unit 342.

<Configuration of machine translation device>
Next, the configuration of the machine translation apparatus according to the embodiment of the present invention will be described. As shown in FIG. 5, a machine translation apparatus 400 according to an embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program for executing a machine translation processing routine to be described later and various data. It can consist of computers. Functionally, the machine translation apparatus 400 includes an input unit 410, a calculation unit 420, and an output unit 490 as shown in FIG.

  The input unit 410 receives an input of a source language sentence to be translated. However, the source language sentence is a word rearranged by the word rearrangement device 200 so as to be closer to the target language.

  The calculation unit 420 includes a translation model 422, a language model 424, a model weight storage unit 426, and a translation execution unit 428.

  The translation model 422 stores the same translation model as the translation model 334 of the machine translation learning device 300.

  The language model 424 stores the same language model as the language model 336 of the machine translation learning device 300.

  The model weight storage unit 426 stores a weight for each of the translation model and the language model, which is the same as the model weight storage unit 342 of the machine translation learning device 300.

  Based on the translation model stored in the translation model 422, the language model stored in the language model 424, and the weight stored in the model weight storage unit 426, the translation execution unit 428 received the input by the input unit 410. A translation process for translating the source language sentence into the target language sentence is executed. As a translation method, a known technique, for example, the technique of Non-Patent Document 7, can be used.

  The translation result is output to a terminal or a storage medium via the output unit 490.

<Operation of word rearrangement learning device>
Next, the operation of the word rearrangement learning device 100 according to the embodiment of the present invention will be described. First, the input unit 10 accepts input of a set of pairs of source language sentences and target language sentences that are parallel translations, and the source language sentence set is stored in the source language sentence database 22, and the target language sentence set is Are stored in the target language sentence database 24. The input unit 10 accepts word correspondence information and stores it in the word correspondence information database 42. Then, the CPU executes the program stored in the ROM of the word rearrangement learning device 100, thereby executing the word rearrangement learning processing routine shown in FIG.

  First, in step S100, a set of pairs of source language sentences and target language sentences, which are parallel translation sentences stored in the source language sentence database 22 and the target language sentence database 24, is read.

  Next, in step S102, syntactic analysis is performed on each source language sentence included in the set of parallel translation pairs read in step S100, and a syntax tree of the source language sentence is generated. In step S104, syntactic analysis is performed on each target language sentence included in the set of parallel translation pairs read in step S100, and a syntax tree of the target language sentence is generated.

  In step S106, for the source language sentence and the target language sentence included in the set of parallel translation pairs read in step S100, the word or word order difference between the source language and the target language is reduced. Remove / add and rearrange words.

  In step S108, based on the syntax tree of the source language sentence and the syntax tree of the target language sentence, the linguistic information is edited in step S106 for each of the pair of translation sentences included in the pair of parallel sentence pairs. To associate words.

  In step S110, the word correspondence information stored in the word correspondence information database 42 is read.

  In step S112, for each pair of translated sentences included in the set of parallel sentence pairs, a feature amount for word association is extracted based on the result of word association in step S108.

  In step S114, the word association model is learned based on the word association information read in step S110 and the feature amount extracted in step S112.

  In step S116, a feature quantity for word association is extracted based on the word association result in step S108 for each of the parallel sentence pairs included in the bilingual sentence pair set. . In step S118, for each of the pairs of translated sentences included in the pair of translated sentences, the word association is performed based on the word association model learned in step S114 and the feature amount extracted in step S116. I do.

  In step S120, for each source language sentence of the parallel translation sentence included in the set of parallel translation sentence pairs, based on the result of word association in step S118, a plurality of words of the target language sentence of the parallel translation sentence are In the case of being associated with at least one word of the source language sentence, in the result of the word association, the central word of the plurality of words of the target language sentence of the parallel translation sentence and at least one of the source language sentence Try to associate two words. Further, for each source language sentence of the parallel translation sentence included in the pair of parallel translation sentences, for each node having two child nodes on the syntax tree of the source language sentence based on the result of word association The correct answer is determined whether to invert between the child nodes.

  In step S122, for each source language sentence in the bilingual sentence, whether the correct answer is determined in step S120 is rearranged for each node having two child nodes on the syntax tree of the source language sentence. A feature amount for determining whether or not is extracted.

  In step S124, the word rearrangement model is learned based on the correct answer of each node determined in step S120 and the feature amount of each node extracted in step S122. At the same time, it is output by the output unit 90, and the word rearrangement learning processing routine is terminated.

<Operation of word rearrangement device>
Next, the operation of the word rearrangement device 200 according to the embodiment of the present invention will be described. First, when an input unit 210 accepts a source language sentence to be rearranged for input to the machine translation learning device 300 or the machine translation device 400, the program stored in the ROM of the word rearrangement device 200 is converted to a CPU. Is executed, the word rearrangement processing routine shown in FIG. 7 is executed.

  First, in step S200, the source language sentence received by the input unit 210 is read.

  Next, in step S202, as in step S102, syntactic analysis is performed on the source language sentence received by the input unit 210 to generate a syntax tree.

  In step S204, in order to determine whether or not to rearrange each node having two child nodes based on the syntax tree of the source language sentence generated in step S202, as in step S122. Extract feature values.

  In step S208, the child nodes are rearranged or rearranged for each node based on the feature amount of each node extracted in step S204 and the word rearrangement model stored in the word rearrangement model 256. Decide if there is no. Then, the rearrangement determination result is reflected in the syntax tree, and output in the form of a word string or syntax tree by the output unit 290, and the word rearrangement processing routine is terminated.

  The word string that has been rearranged and output by the output unit 290 is used as an input of the machine translation learning device 300 or the machine translation device 400.

<Operation of machine translation learning device>
Next, the operation of the machine translation learning apparatus 300 according to the embodiment of the present invention will be described. First, the input unit 310 accepts input of a set of pairs of source language sentences and target language sentences, which are parallel translation sentences, and the set of source language sentences is stored in the parallel source language sentence database 322, and a set of target language sentences. Is stored in the bilingual target language sentence database 324. However, each of the input source language sentences is a word rearranged by the word rearrangement device 200 so as to approach the word order of the target language.

  The input unit 310 accepts a set of target language sentences and stores them in the target language sentence database 330. Further, the input unit 310 accepts input of a set of pairs of source language sentences and target language sentences, which are parallel translation sentences, as learning data for weight adjustment of the model, and is stored in the parallel sentence database 338 for weight adjustment. . However, each of the input source language sentences is a word rearranged by the word rearrangement device 200 so as to approach the word order of the target language.

  Then, when the CPU executes the program stored in the ROM of the machine translation learning device 300, the machine translation learning process routine shown in FIG. 8 is executed.

  First, in step S300, a set of pairs of source language sentences and target language sentences, which are parallel translation sentences, stored in the parallel source language sentence database 322 and the target language sentence database 324 is read.

  Next, in step S302, a word is associated with each of the parallel sentence pairs included in the parallel sentence pair set.

  In step S304, a translation model is learned based on the word association result in step S302, stored in the translation model 334, and output by the output unit 390.

  In step S306, a set of target language sentences stored in the target language sentence database 330 is read.

  In step S308, a language model is learned based on the set of target language sentences read in step S306, stored in the language model 336, and output by the output unit 390.

  In step S310, a set of parallel translation pairs stored in the weight adjustment parallel translation database 338 is read.

  In step S312, the weight of each model is calculated based on the set of parallel translation pairs read in step S310, the translation model stored in the translation model 334, and the language model stored in the language model 336. Adjust, store in the model weight storage unit 342, and output by the output unit 390, and terminate the machine translation learning processing routine.

<Operation of machine translation device>
Next, the operation of the machine translation apparatus 400 according to the embodiment of the present invention will be described. First, when a source language sentence to be machine-translated is received by the input unit 410, the CPU executes the program stored in the ROM of the machine translation apparatus 400, thereby executing the machine translation processing routine shown in FIG. . However, the source language sentence to be machine-translated is a word rearranged by the word rearrangement device 200 so as to be closer to the target language.

  First, in step S400, the source language sentence received by the input unit 410 is read.

  Next, in step S402, based on the translation model stored in the translation model 422, the language model stored in the language model 424, and the weight of each model stored in the model weight storage unit 426, the above step is performed. A process of translating the source language sentence obtained in S400 into the target language sentence is executed, the translation result is output by the output unit 490, and the machine translation process routine is terminated.

<Example>
Next, examples in which Japanese-English translation experiments were conducted will be described below.

  As described in the above embodiment, in the configuration using the first automatic word association and the second automatic word association in the word rearrangement learning device 100, in the machine translation result by the machine translation device 400, the machine translation The value of RIBES, a kind of evaluation value, was 76.98%, far exceeding the 69.30% without pre-sorting. Further, in the configuration in which the word rearrangement learning device learns only by the first automatic word association, the RIBES value in the machine translation result by the machine translation device is 75.47%, and the first automatic word association and the second automatic word It has been found that a larger effect can be obtained when both of the correspondences are used.

  As described above, according to the word rearrangement learning device according to the embodiment of the present invention, based on the result of word association, the plurality of words in the target language sentence are at least one word in the source language sentence. Is associated with at least one word of the source language sentence and has two child nodes in the syntax tree of the source language sentence. For each node, the word string of the target language sentence corresponding to the word string of the source language sentence represented by the two child nodes and the two child nodes with the order reversed are obtained by using the result of the word association. Based on the word string of the target language sentence corresponding to the word string of the source language sentence, a correct answer is determined as to whether or not to reverse the order of the two child nodes, and two child nodes are included in the syntax tree of the source language sentence. 2 for nodes By learning the word rearrangement model for determining whether to reverse the order of child nodes, even if there is an overlap in word mapping, determine the preordering of words appropriately Can do.

  In addition, it becomes possible to learn the word rearrangement model used in the pre-arrangement for correctly translating the word order in machine translation based on highly accurate word association and using more learning examples. As a result, it is possible to obtain high translation accuracy.

  Moreover, according to the word rearrangement device according to the embodiment of the present invention, the feature amount for each node having two child nodes in the syntax tree of the source language sentence, and the word rearrangement model learned as described above Based on the above, it is determined whether or not to reverse the order of the two child nodes for the node having two child nodes in the syntax tree of the source language sentence, and by reordering the words of the source language sentence, The pre-ordering of words can be determined.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, the word rearrangement learning device may not generate the syntax tree of the target language sentence. In this case, the target language syntactic analysis unit 30 becomes unnecessary, and the automatic word association unit may perform word association based on the syntax tree of the source language sentence and the target language sentence.

  In the present embodiment, the case where the word rearrangement device and the machine translation learning device are configured as separate devices has been described as an example. However, the present invention is not limited to this, and the word rearrangement device and the machine The translation learning device may be configured as one device. Moreover, although the case where the word rearrangement device and the machine translation device are configured as separate devices has been described as an example, the present invention is not limited to this, and the word rearrangement device and the machine translation device are one device. It may be configured.

  Further, in the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium or provided via a network. It is also possible to do.

10, 210, 310, 410 Input unit 20, 220, 320, 420 Operation unit 26 Syntax analysis unit 28 Source language syntax analysis unit 30 Target language syntax analysis unit 32 Automatic word association unit 34 First automatic word association unit 36 Language Information editing unit 38 First word association unit 40 Second automatic word association unit 42 Word association information database 44 Second automatic word association model learning unit 50 Word rearrangement correct decision unit 52 Feature amount extraction unit 54 Model learning unit 56 Word rearrangement model 90, 290, 390, 490 Output unit 100 Word rearrangement learning device 200 Word rearrangement device 226 Syntax analysis unit 228 Source language syntax analysis unit 252 Feature quantity extraction unit 254 Word rearrangement determination unit 300 Machine translation learning device 326 Word association unit 328 Translation model learning unit 332 Language model learning unit 340 Weight adjustment unit 4 00 Machine translation device 428 Translation execution unit

Claims (7)

A syntactic analysis unit that performs syntactic analysis of the source language for each of the source language sentences included in a plurality of pairs of source language sentences and target language sentences to be translated, and generates a syntax tree of the source language sentences;
Automatic word association for associating words between the source language sentence and the target language sentence of the pair based on the syntax tree of the source language sentence generated by the syntactic analysis unit for each of the plurality of pairs And
For each of the plurality of pairs, for each node having two child nodes in the syntax tree of the source language sentence of the pair, a word string of the source language sentence represented by the two child nodes is a first source language word. Column,
A word string of the source language sentence obtained by reversing the order of the two child nodes is a second source language word string;
Each word of the word string representing the target language sentence is attached with position information that is an integer that specifies the position of each word relative to the first word,
A vector in which the position information of the words of the target language sentence corresponding to each word in order from the first word of the first source language word string is arranged as a first series,
A vector in which the position information of the words of the target language sentence corresponding to each word in order from the first word of the second source language word string is arranged as a second series,
When the second sequence is closer to the ascending order than the first sequence, the one obtained by inverting the order of the two child nodes is determined as a correct answer, and the first sequence than the second sequence is determined. If the sequence is closer to ascending order, a word rearrangement correct answer determination unit that determines a correct answer that does not reverse the order of the two child nodes;
For each of the plurality of pairs, a feature amount extraction unit that extracts a feature amount for each node for which the correct answer is determined by the word rearrangement correct determination unit;
The correct answer for each node determined for each of the plurality of pairs by the word rearrangement correct answer determination unit, and each node extracted for each of the plurality of pairs by the feature amount extraction unit A word rearrangement model for determining whether or not to reverse the order of the two child nodes for a node having two child nodes in the syntax tree of the source language sentence A replacement model learning unit;
Including
When there are a plurality of words of the target language sentence corresponding to one word of the source language sentence, the median value of the position information of the plurality of words is determined as the original value in the first series and the second series. A word rearrangement learning device characterized in that it is position information of a word of a target language sentence corresponding to one word of a language sentence.   The automatic word association unit, when there is a word that does not correspond between the source language sentence and the target language sentence, which is obtained based on the result of the syntactic analysis and the surface layer of the word, After deleting a language sentence or the target language sentence or adding a predetermined word corresponding to the non-corresponding word to the source language sentence or the target language sentence so as to correspond to the non-corresponding word, the pair The word rearrangement learning apparatus according to claim 1, wherein word association is performed between the source language sentence and the target language sentence. The automatic word association unit
For each of the plurality of pairs, a first automatic word that associates words between the source language sentence and the target language sentence of the pair based on the syntax tree of the source language sentence generated by the syntactic analysis unit An association unit;
Using the result of word association between the source language sentence and the target language sentence for each of the plurality of pairs by the first automatic word association unit as a feature amount, for each of the plurality of pairs, The word rearrangement learning apparatus according to claim 2, further comprising: a second automatic word association unit that associates words between a pair of source language sentences and target language sentences. A syntactic analysis unit that performs syntactic analysis of the source language based on the input source language sentence and generates a syntax tree of the source language sentence;
A feature quantity extraction unit that extracts a feature quantity for each node having two child nodes in the syntax tree of the source language sentence;
The word rearrangement learning according to any one of claims 1 to 3, and the feature amount for each node having two child nodes in the syntax tree of the source language sentence extracted by the feature amount extraction unit. Determining whether to reverse the order of the two child nodes for a node having two child nodes in the syntax tree of the source language sentence based on the word rearrangement model learned by the apparatus; A word rearrangement determination unit for rearranging words of the source language sentence based on the result of
Word sorter including A word rearrangement learning method in a word rearrangement learning device including a syntactic analysis unit, an automatic word association unit, a word rearrangement correct determination unit, a feature amount extraction unit, and a word rearrangement model learning unit,
The syntactic analysis unit performs a syntactic analysis of the source language for each of the source language sentences included in a plurality of pairs of source language sentences and target language sentences to be translated to generate a syntax tree of the source language sentences,
The automatic word association unit, for each of the plurality of pairs, based on the syntax tree of the source language sentence generated by the syntactic analysis unit, the word between the source language sentence and the target language sentence of the pair Make a match,
The word rearrangement correct answer determination unit, for each of the plurality of pairs, for each node having two child nodes in the syntax tree of the source language sentence of the pair, of the source language sentence represented by the two child nodes Let the word string be the first source language word string,
A word string of the source language sentence obtained by reversing the order of the two child nodes is a second source language word string;
Each word of the word string representing the target language sentence is attached with position information that is an integer that specifies the position of each word with respect to the first word,
A vector in which the position information of the words of the target language sentence corresponding to each word in order from the first word of the first source language word string is arranged as a first series,
A vector in which the position information of the words of the target language sentence corresponding to each word in order from the first word of the second source language word string is arranged as a second series,
When the second sequence is closer to the ascending order than the first sequence, the one obtained by inverting the order of the two child nodes is determined as a correct answer, and the first sequence than the second sequence is determined. When the sequence is closer to the ascending order, the correct answer is determined not to reverse the order of the two child nodes,
The feature amount extraction unit extracts a feature amount for each node for which the correct answer is determined by the word rearrangement correct answer determination unit for each of the plurality of pairs.
The word rearrangement model learning unit determines the correct answer for each node determined for each of the plurality of pairs by the word rearrangement correct answer determination unit, and each of the plurality of pairs by the feature amount extraction unit. And determining whether to reverse the order of the two child nodes for a node having two child nodes in the syntax tree of the source language sentence based on the feature amount extracted for each node. Learning the word rearrangement model,
When there are a plurality of words of the target language sentence corresponding to one word of the source language sentence, the median value of the position information of the plurality of words is determined as the original value in the first series and the second series. A word rearrangement learning method, comprising: position information of a word of a target language sentence corresponding to one word of a language sentence. Computer program to function as each section of words sorted learning equipment according to any one of claims 1 to 3.   The program for functioning a computer as each part of the word rearrangement apparatus of Claim 4.