JP2015001862A - Bilingual phrase learning device, statistical machine translation device, bilingual phrase learning method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve the problem in prior art that a translation model has to be updated every time a bilingual corpus is added.SOLUTION: When calculating a score for each phrase pair acquired from a j(2<=j<=N)th bilingual corpus (a supplemented bilingual corpus), one or more pieces of appearance frequency information corresponding to a (j-1)th bilingual corpus is used to calculate the score for each phrase pair corresponding to the j-th bilingual corpus, and the calculated score is used to create a translation model, and by utilizing the newly created translation model in linkage with the original translation model, a translation model can be easily upgraded stepwise.

Description

  The present invention relates to a parallel phrase learning device that learns a parallel phrase.

  In conventional statistical machine translation (see Non-Patent Document 1), bilingual knowledge such as a phrase table is extracted from bilingual data, and a translation model is learned. Based on the translation model, a translation system is realized. Further, in order to estimate an accurate translation model, it is necessary to learn a large amount of parallel translation data by a learning method called batch learning. Note that batch learning is a learning method in which all learning data is optimized.

  In particular, bilingual data is increasing year by year, and the prior art requires re-learning each time data is added. However, there was no guarantee that a better translation model would be estimated as a result of this relearning.

  In order to solve this problem, conventionally, there has been a method of dividing parallel translation data for each field, learning a local translation model in each field, and combining the translation models (see Non-Patent Document 2).

  Further, in the prior art, there is a method using a discriminator that appropriately assigns a field to an input sentence in a source language (see Non-Patent Document 3).

  In addition, in the prior art, a technique of adding a field-dependent feature and optimizing the parameter with respect to parallel translation data to which a field label is assigned is used (see Non-Patent Documents 3 to 6).

  Furthermore, in the conventional technique, there is a method called incremental learning. In this method, each time bilingual data is added, the model is updated according to the added amount (see Non-Patent Document 7).

Phillip Koehn, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proc. HLT. George Foster and Roland Kuhn. 2007. Mixture-model adaptation for smt. In Proc. Of the second workshop of SMT. Jia Xu, Yonggang Deng, Yuqing Gao, and Hermann Ney. 2007. Domain dependent statistical machine translation. MT Summit XI. Wei Wang, Klaus Macherey, Wolfgang Macherey, Franz Och, and Peng Xu. 2012. Improved domain adaptation for statistical machine translation.In Proc. Of AMTA. Yajuan Lu, Jin Huang, and Qun Liu. 2007. Improving statistical machine translation performance by training data selection and optimization.In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), pages 343-350. Jinsong Su, Hua Wu, Haifeng Wang, Yidong Chen, Xiaodong Shi, Huailin Dong, and Qun Liu. 2012.Translation model adaptation for statistical machine translation with monolingual topic information.In Proc. Of ACL. Abby Levenberg, Chris Callison-Burch, and Miles Osborne. 2010.Stream-based translation models for statistical machine translation.In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, HLT '10, pages 394- 402, Stroudsburg, PA, USA. Association for Computational Linguistics.

  However, the above-described technique based on the field application has a problem that the translation model cannot be estimated accurately because parallel translation data is reduced when learning locally in each field. In addition, the technique based on field application requires a discriminator in order to determine the weight for each field or to discriminate the field of the input sentence. In particular, the technologies according to Non-Patent Documents 2 and 3 have a problem that a discriminator that assigns an appropriate field to an input sentence is required, and the accuracy of translation depends on the performance of the discriminator.

  In addition, in the methods based on the features in Non-Patent Documents 4 to 6, there is a problem that accurate parallel translation data labeled in the field is required.

  Furthermore, the incremental learning in Non-Patent Document 7 is based on a technique that requires adjustment of complicated parameters such as an online EM algorithm, and becomes a complicated system such as further optimization of an added translation model. There was a problem.

  In summary, in the prior art, each time a bilingual corpus is added, the process for enhancing the translation model step by step is troublesome.

  In view of this point, an object of the present invention is to easily enrich a translation model step by step by using a translation model generated from an added bilingual corpus in connection with the original translation model.

  The bilingual phrase learning device according to the first aspect of the present invention is a bilingual information storage unit capable of storing N (N is a natural number of 2 or more) bilingual corpus having one or more bilingual information having a bilingual sentence and a tree structure of the bilingual sentence. And a score that is information about a phrase pair that is a pair of a first language phrase having one or more words in the first language and a second language phrase having one or more words in the second language and the appearance probability of the phrase pair One or more phrase appearance frequency information including a phrase table that can store one or more scored phrase pairs having each of the following, for each bilingual corpus, a phrase pair, and F appearance frequency information that is information relating to the appearance frequency of the phrase pair Are stored for each parallel corpus, a phrase appearance frequency information storage unit, a symbol for identifying a method for generating a new phrase pair, and information on the appearance frequency of the symbol. A symbol appearance frequency information storage unit capable of storing one or more symbol appearance frequency information having S appearance frequency information, and for each parallel corpus, using one or more phrase appearance frequency information, Generation phrase pair acquisition unit that acquires a phrase pair having a bilingual phrase, and update of the phrase appearance frequency information that increases the F appearance frequency information corresponding to the phrase pair by a predetermined value when the phrase pair can be acquired Part and phrase pair could not be acquired, using one or more symbol appearance frequency information, a symbol acquisition unit for acquiring one symbol, and S appearance frequency information corresponding to the symbol acquired by the symbol acquisition unit, A symbol appearance frequency information update unit that increases by a predetermined value, and two frames smaller than the phrase pair to be acquired if the phrase pair cannot be acquired. A partial phrase pair generation unit that generates a pair and a first process for generating a new phrase pair according to the symbol acquired by the symbol acquisition unit, or two smaller phrase pairs, and one or more phrase appearance frequency information A new second language phrase in which two first language phrases that form two phrase pairs are connected in order, and a second second language phrase that forms two phrase pairs in order. 2nd process which produces | generates one phrase pair which has a language phrase, or two smaller phrase pairs are produced | generated, and two produced | generated two phrase pairs are comprised using one or more phrase appearance frequency information A new first language phrase that connects two first language phrases in sequence and two second language phrases that form two phrase pairs in reverse order A new phrase pair generation unit that performs one of the third processes for generating a single phrase pair having a new second language phrase that is connected, and a phrase appearance frequency for the phrase pair generated by the new phrase pair generation unit Stored in the phrase update frequency information storage unit and the control unit that instructs to recursively perform the processing of the information update unit, symbol acquisition unit, symbol appearance frequency information update unit, partial phrase pair generation unit, and new phrase pair generation unit The score calculation unit that calculates a score for each phrase pair in the phrase table using one or more phrase appearance frequency information that has been performed, and the phrase table update that stores the score calculated by the score calculation unit in association with each phrase pair A score calculation unit for each phrase pair acquired from the j (2 <= j <= N) -th parallel corpus A bilingual phrase learning device that calculates a score for each phrase pair corresponding to the jth bilingual corpus using the one or more phrase appearance frequency information corresponding to the (j-1) th bilingual corpus when calculating the core. is there.

  With this configuration, the translation model can be easily enhanced in stages.

  The bilingual phrase learning device according to the second aspect of the present invention is different from the first aspect in that the bilingual information storage unit stores one or more bilingual corpora, the bilingual corpus receiving unit that receives the bilingual corpus, The bilingual corpus accumulating unit that accumulates the bilingual corpus accepted by the corpus accepting unit in the bilingual information storage unit, and the control unit stores the bilingual corpus accepted by the bilingual corpus accumulating unit in the bilingual information storage unit, Instructed to perform processing of the generated phrase pair acquisition unit, phrase appearance frequency information update unit, symbol acquisition unit, symbol appearance frequency information update unit, partial phrase pair generation unit, and new phrase pair generation unit for the bilingual corpus, and score calculation When calculating the score for each phrase pair acquired from the bilingual corpus received by the bilingual corpus receiving unit, The bilingual corpus accepting unit accepts one or more phrase appearance frequency information corresponding to one of the one or more bilingual corpora stored in the bilingual corpus before the product stores the bilingual corpus. This is a bilingual phrase learning device that calculates a score for each phrase pair corresponding to a bilingual corpus.

  With this configuration, the translation model can be easily enhanced in stages.

  The parallel phrase learning device according to the third aspect of the invention divides two or more parallel sentences into N groups and acquires the tree structure of the parallel sentences from the parallel sentences of each group. The bilingual corpus generating unit that accumulates the N bilingual corpora created in the bilingual information storage unit, and the score calculating unit calculates the score for each phrase pair acquired from one bilingual corpus. The bilingual phrase learning apparatus calculates a score for each phrase pair corresponding to one bilingual corpus using one or more phrase appearance frequency information corresponding to another bilingual corpus different from the bilingual corpus.

  With this configuration, the translation model can be easily enhanced in stages.

  In addition, the parallel phrase learning device according to the fourth aspect of the present invention provides the phrase calculation unit according to any one of the first to third aspects, wherein the score calculation unit uses a hierarchical Chinese restaurant process and each phrase corresponding to each parallel corpus. This is a bilingual phrase learning device that calculates a score for a pair.

  With this configuration, it is possible to easily enrich the translation model step by step using a hierarchical Chinese restaurant process.

  The statistical machine translation device according to the fifth aspect of the invention includes a phrase table learned by any one of the first to fourth parallel phrase learning devices, and a reception unit that receives a sentence in a first language having one or more words. The phrase acquisition unit that extracts one or more phrases from the sentence received by the reception unit and acquires one or more phrases in the second language from the phrase table using the score of the phrase table, and the phrase acquisition unit 1 acquired It is a statistical machine translation apparatus which comprises the sentence structure part which comprises the sentence of a 2nd language from the above phrase, and the output part which outputs the sentence which the sentence structure part comprised.

  With this configuration, it is possible to realize highly accurate machine translation using a translation model enriched in stages.

  According to the bilingual phrase learning device of the present invention, the translation model can be easily enhanced step by step.

Block diagram of bilingual phrase learning apparatus 1 in Embodiment 1 of the present invention The flowchart explaining operation | movement of the bilingual phrase learning apparatus 1 in Embodiment 1 of this invention. The flowchart explaining the phrase production | generation process in Embodiment 1 of this invention The figure which shows the example of the tree structure which comprises the parallel translation information in Embodiment 1 of this invention Block diagram of bilingual phrase learning device 2 in Embodiment 2 of the present invention The flowchart explaining operation | movement of the bilingual phrase learning apparatus 2 in Embodiment 2 of this invention. Block diagram of statistical machine translation apparatus 3 according to Embodiment 3 of the present invention The figure explaining the data set used in the experiment in the embodiment of the present invention The figure which shows the experimental result in embodiment of this invention The figure which shows the experimental result in embodiment of this invention The figure which shows the experimental result in embodiment of this invention Overview of a computer system according to an embodiment of the present invention The block diagram of the computer system in embodiment of this invention

  Hereinafter, embodiments of a parallel phrase learning device and the like will be described with reference to the drawings. In addition, since the component which attached | subjected the same code | symbol in embodiment performs the same operation | movement, description may be abbreviate | omitted again.

(Embodiment 1)
In this embodiment, a bilingual phrase learning apparatus that can easily enhance a translation model step by step by connecting a translation model generated from the added bilingual corpus to the original translation model will be described.

  FIG. 1 is a block diagram of a bilingual phrase learning apparatus 1 according to the present embodiment.

  The bilingual phrase learning device 1 includes a bilingual information storage unit 100, a phrase table 101, a phrase appearance frequency information storage unit 102, a symbol appearance frequency information storage unit 103, a bilingual corpus reception unit 104, a bilingual corpus storage unit 105, and a phrase table initialization unit. 106, generated phrase pair acquisition unit 107, phrase appearance frequency information update unit 108, symbol acquisition unit 109, symbol appearance frequency information update unit 110, partial phrase pair generation unit 111, new phrase pair generation unit 112, control unit 113, score calculation Unit 114, parsing unit 115, phrase table update unit 116, and tree update unit 117.

  The bilingual information storage unit 100 can store N (N is a natural number of 2 or 3) bilingual corpus. The bilingual corpus here has one or more bilingual information. The parallel translation information has a parallel translation sentence and a tree structure of the parallel translation sentence. The bilingual sentence is a pair of a first language sentence and a second language sentence. The first language sentence is a sentence in the first language. The second language sentence is a sentence in the second language. Here, the sentence means one or more words and includes a phrase. The bilingual sentence tree structure is information representing a correspondence of phrases (including words) obtained by dividing sentences in two languages in a tree structure.

  The bilingual information storage unit 100 may store one bilingual corpus before processing, and may accumulate one or more bilingual corpora after the second.

The phrase table 101 may store one or more scored phrase pairs for each of N parallel corpora. The phrase pair with a score has a phrase pair and a score. A phrase pair is a pair of a first language phrase and a second language phrase. The first language phrase is a phrase having one or more words in the first language. The second language phrase is a phrase having one or more words in the second language. Phrases are widely understood as including sentences. The score is information regarding the appearance probability of the phrase pair. In addition, the score, for example, is a phrase pair probability θ _t. Here, the phrase pair is a concept including a rule pair, and is widely understood. Also, one or more scored phrase pairs may be understood to have the same meaning as the translation model described above.

  The phrase appearance frequency information storage unit 102 can store one or more phrase appearance frequency information for each parallel corpus. The phrase appearance frequency information includes a phrase pair and F appearance frequency information. F appearance frequency information is information regarding the appearance frequency of phrase pairs. The F appearance frequency information is preferably the appearance frequency of the phrase pair, but may be the appearance probability of the phrase pair. Note that the initial values of the F appearance frequency information are all 0, for example.

  The symbol appearance frequency information storage unit 103 can store one or more pieces of symbol appearance frequency information. The symbol appearance frequency information includes a symbol and S appearance frequency information. The symbol is information for identifying a method for generating a new phrase pair. The symbol is, for example, any one of BASE, REG, and INV. Here, BASE is a symbol indicating that a phrase pair is generated from a base measure, REG is a normal non-terminal symbol, and INV is an inverted non-terminal symbol. The S appearance frequency information is information related to the appearance frequency of symbols. The S appearance frequency information is preferably a symbol appearance frequency, but may be a symbol appearance probability or the like. The initial value of the S appearance frequency information is, for example, 0 for all three symbols. Note that the base measure is a prior probability calculated by a word translation model such as IBM Model 1 and is a known technique, and thus detailed description thereof is omitted.

  The parallel corpus reception unit 104 receives a parallel corpus. Here, reception means reception of information input from an input device such as a keyboard, mouse, touch panel, reception of information transmitted via a wired or wireless communication line, recording on an optical disk, magnetic disk, semiconductor memory, or the like. It is a concept including reception of information read from a medium.

  The bilingual corpus input means may be anything such as a keyboard, mouse, or menu screen. The bilingual corpus accepting unit 104 can be realized by a device driver of input means such as a keyboard, control software for a menu screen, or the like.

  The bilingual corpus accumulating unit 105 accumulates the bilingual corpus received by the bilingual corpus receiving unit 104 in the bilingual information storage unit 100.

  The phrase table initialization unit 106 generates initial information of one or more phrase pairs with scores from one or more parallel translation information of the bilingual corpus and accumulates it in the phrase table 101. Note that the phrase table initialization unit 106 acquires, for example, a phrase pair that appears in the tree structure of the bilingual sentence included in one or more parallel translation information and the number of appearances as a phrase pair with a score, and accumulates them in the phrase table 101. In such a case, the score is the number of appearances. The phrase table initialization unit 106 normally generates initial information of one or more scored phrase pairs for each bilingual corpus and accumulates it in the phrase table 101. The phrase table initialization unit 106 generates initial information of one or more scored phrase pairs from one or more parallel information included in the bilingual corpus received by the bilingual corpus receiving unit 104 and stores the initial information in the phrase table 101. good.

  The generated phrase pair acquisition unit 107 acquires a phrase pair having a first language phrase and a second language phrase using one or more phrase appearance frequency information for each parallel corpus.

The generated phrase pair acquisition unit 107 acquires, for each bilingual corpus, one or more bilingual sentences stored in the bilingual corpus and appearances of one or more phrase pairs that constitute the tree structure of the bilingual sentences ( Usually, the appearance frequency “1”) is subtracted from the score of the phrase pair existing in the phrase table 101. Next, the generated phrase pair acquisition unit 107 acquires a phrase pair having a first language phrase and a second language phrase using one or more phrase appearance frequency information (accurately, it tries to acquire). Here, the use of 1 or more phrases appearance frequency information may be, for example, by using the probability distribution P _t phrase pairs. That is, it is preferable that the generated phrase pair acquisition unit 107 acquires a phrase pair having the first language phrase and the second language phrase using the phrase pair probability distribution P _t .

  When the generated phrase pair acquisition unit 107 or the new phrase pair generation unit 112 can acquire the phrase pair, the phrase appearance frequency information update unit 108 increases the F appearance frequency information corresponding to the phrase pair by a predetermined value. To do. The F appearance frequency information here is usually the appearance frequency of a phrase pair. The predetermined value is usually 1.

When the generated phrase pair acquisition unit 107 or the like cannot acquire a phrase pair, the symbol acquisition unit 109 acquires one symbol using one or more symbol appearance frequency information. Here, it is preferable to use the symbol probability distribution P _x (x; θ _x ) to use one or more symbol appearance frequency information. That is, it is preferable that the symbol acquisition unit 109 acquires one symbol using the probability distribution of symbols when the generation phrase pair acquisition unit 107 cannot acquire the generation phrase pair. One symbol is, for example, any one of BASE, REG, and INV. Note that “x” of P _x (x; θ _x ) is a symbol, and (θ _x ) is a probability that the symbol is adopted.

  The symbol appearance frequency information update unit 110 increases the S appearance frequency information corresponding to the symbol acquired by the symbol acquisition unit 109 by a predetermined value. The predetermined value is usually 1.

When the generated phrase pair acquisition unit 107 or the like cannot acquire a phrase pair, the partial phrase pair generation unit 111 generates two phrase pairs smaller than the phrase pair to be acquired. Moreover, the partial phrase pair production | generation part 111 normally produces | generates two phrase pairs smaller than the phrase pair which it was going to acquire using the prior probability of a phrase pair, when a phrase pair cannot be acquired. More specifically, for example, when a phrase pair in the j-th parallel corpus is to be generated, the partial phrase pair generation unit 111 sets the prior probability P ^j−1 of the phrase pair in the (j−1) -th parallel corpus. To generate two phrase pairs smaller than the phrase pair to be acquired. When j = 1, the partial phrase pair generation unit 111 generates two phrase pairs smaller than the phrase pair to be acquired using P _base (for example, IBM Model 1). For example, if the phrase pair you are trying to acquire is <red cookbook, red cookbook>, then “P _base (<red cookbook, red cookbook>) = P _x (REG) * P _t (<red, red>) * P_ {t} (<cookbook, cooking book>) + P _x (REG) * P _t (<red, red cooking>) * P _t (<cookbook, book>) + P _x (INV) * P _t (< red, book>) * P _t (<cookbook, red food>) + P _x (INV) * P _t (<red, cook book>) * P _t (<cookbook, red>) + P _x (BASE) * P _base (<red cookbook>) ”. Note that P _base is a prior probability calculated by a word translation model such as IBM Model 1, for example.

  The new phrase pair generation unit 112 performs either the first process, the second process, or the third process according to the symbol acquired by the symbol acquisition unit 109. The new phrase pair generation unit 112 performs the first process when the symbol acquired by the symbol acquisition unit 109 is BASE, performs the second process when the symbol is REG, and when the symbol is INV A third process is performed.

  Here, the first process is a process of generating a new phrase pair. The first process is a process for generating a new phrase pair using the prior probability of the phrase pair. Here, when processing is performed on the j (2 <= j <= N) th parallel corpus, the prior probability of the phrase pair used in the first processing is (j-1) th parallel translation. This is the prior probability of the phrase pair for the corpus.

  In the second process, two smaller phrase pairs are generated, and two or more first language phrases constituting the two generated phrase pairs are sequentially connected using one or more phrase appearance frequency information. This is a process of generating one phrase pair having a single language phrase and a new second language phrase in which two second language phrases constituting two phrase pairs are sequentially connected.

Further, the third process generates two smaller phrase pairs, and uses the one or more phrase appearance frequency information to newly connect the two first language phrases constituting the generated two phrase pairs in order. This is a process of generating one phrase pair having a single language phrase and a new second language phrase in which two second language phrases constituting two phrase pairs are connected in reverse order. Here, using one or more phrase appearance frequency information may mean using a phrase pair generation probability (P _hier ).

The control unit 113 performs the phrase appearance frequency information update unit 108, the symbol acquisition unit 109, the symbol appearance frequency information update unit 110, the partial phrase pair generation unit 111, and the new phrase pair generated by the new phrase pair generation unit 112. The phrase pair generation unit 112 is instructed to perform the process recursively. Note that “recursively” usually means that the recursive process ends when the processing target becomes a word pair. Note that recursive processing is terminated when the processing target is generated the phrase (without using the base measure) directly from P _t. Also, recursive processing may generate a BASE from P _x, and ends when that generated phrase pair from P _base.

  Further, the control unit 113 accumulates the bilingual corpus accepted by the bilingual corpus accumulating unit 105 in the bilingual information storage unit 100, and then generates a generated phrase pair acquisition unit 107, a phrase appearance frequency information update unit 108, and a symbol acquisition unit for the bilingual corpus 109, the symbol appearance frequency information update unit 110, the partial phrase pair generation unit 111, and the new phrase pair generation unit 112 may be instructed to perform processing.

  The score calculation unit 114 calculates a score for each phrase pair in the phrase table 101 using one or more phrase appearance frequency information stored in the phrase appearance frequency information storage unit 102.

  When the score calculation unit 114 calculates a score for each phrase pair acquired from the j (2 <= j <= N) -th parallel corpus, one or more phrases corresponding to the (j-1) -th parallel corpus Using the appearance frequency information, a score for each phrase pair corresponding to the j-th parallel corpus is calculated.

  When the score calculation unit 114 calculates a score for each phrase pair acquired from the bilingual corpus received by the bilingual corpus receiving unit 104, the bilingual corpus accumulating unit 105 stores the bilingual corpus before storing the bilingual corpus. The score for each phrase pair corresponding to the bilingual corpus received by the bilingual corpus accepting unit 104 is calculated using one or more phrase appearance frequency information corresponding to any one of the one or more bilingual corpora stored in You may do it.

Further, the score calculation unit 114 may calculate a score for each phrase pair corresponding to each bilingual corpus using the hierarchical Chinese restaurant process according to Equation 1.

From the above, it can be said that the bilingual phrase learning device 1 does not estimate model parameters for all the bilingual data <F, E>, but learns a part of the bilingual data specified in a certain field. Furthermore, it can be said that the bilingual phrase learning apparatus 1 uses a model learned in another field as a prior probability, instead of using a model such as IBM Model 1. Specifically, the bilingual data <F, E> is divided into ^J fields <F ¹ , E ¹ >… <F ^J , E ^J >, and the parameters of the translation model of the jth field θ _t ^j is learned from the parallel translation data <F ^j , E ^j > of the j-th field using the model P ^j-1 obtained in the previous j-1 field as the prior probability (Formula 2 reference). The translation model of Equation 2 is called a hierarchical Pitman-Yor model, and is used in, for example, an ngram language model or field application. Further, when the hierarchical Pitman-Yor model is expressed as a hierarchical Chinese restaurant process, it is expressed as Equation 1. In the parallel translation data <F, E>, “F” is a source language sentence, and “E” is a target language sentence (second language sentence).

  The parsing unit 115 obtains a tree structure of bilingual sentences (including phrases) that maximizes the score calculated by the score calculation unit 114. In more detail, the purging unit 115 acquires a tree structure by an ITG chart parser. The ITG chart parser is described in "M. Saers, J. Nivre, and D. Wu. Learning stochastic bracketing inversion transduction grammars with a cubic time biparsing algorithm. In Proc. IWPT, 2009."

  The phrase table update unit 116 stores the score calculated by the score calculation unit 114 in association with each phrase pair. Moreover, the phrase table update part 116, when the phrase pair corresponding to the score which the score calculation part 114 calculated does not exist in the phrase table 101, the phrase pair with a score which has the score and frame pair which the score calculation part 114 calculated The phrase table 101 may be accumulated.

  The tree update unit 117 accumulates the tree structure acquired by the parsing unit 115 in the bilingual corpus. Here, the tree update unit 117 normally overwrites the tree structure. That is, the old tree structure in the bilingual corpus is updated to a new tree structure.

  The bilingual information storage unit 100, the phrase table 101, the phrase appearance frequency information storage unit 102, and the symbol appearance frequency information storage unit 103 are preferably non-volatile recording media, but can also be realized by volatile recording media.

  The process in which the bilingual corpus is stored in the bilingual information storage unit 100 or the like is not limited. For example, a bilingual corpus or the like may be stored in the bilingual information storage unit 100 or the like via a recording medium, and a bilingual corpus or the like transmitted via a communication line or the like is stored in the bilingual information storage unit 100 or the like. Alternatively, the bilingual corpus or the like input via the input device may be stored in the bilingual information storage unit 100 or the like.

  Bilingual corpus accumulation unit 105, phrase table initialization unit 106, generated phrase pair acquisition unit 107, phrase appearance frequency information update unit 108, symbol acquisition unit 109, symbol appearance frequency information update unit 110, partial phrase pair generation unit 111, new phrase The pair generation unit 112, the control unit 113, the score calculation unit 114, the parsing unit 115, the phrase table update unit 116, and the tree update unit 117 can be normally realized by an MPU, a memory, or the like. The processing procedure of the bilingual corpus storage unit 105 or the like is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  Next, operation | movement of the parallel translation phrase learning apparatus 1 is demonstrated using the flowchart of FIG. In this flowchart, when the bilingual phrase learning device 1 sequentially receives N (N is a natural number of 2 or 3) bilingual corpora and constructs a phrase table from the j (j <N) -th bilingual corpus, (j -1) The case where the 1st phrase table is used is demonstrated.

  (Step S201) The parallel corpus accepting unit 104 substitutes 1 for a counter i.

  (Step S202) The parallel corpus reception unit 104 determines whether or not an i-th parallel corpus has been received. If the i-th parallel corpus is accepted, the process proceeds to step S203. If the i-th parallel corpus is not accepted, the process returns to step S202.

  (Step S203) The phrase table initialization unit 106 generates initial information of one or more phrase pairs with scores from one or more pieces of parallel translation information included in the i-th parallel corpus, and associates it with i to the phrase table 101. To accumulate.

(Step S204) The generated phrase pair acquisition unit 107 acquires one or more parallel translations included in the parallel translation corpus received in Step S201, and the appearance of one or more phrase pairs constituting the tree structure of each parallel translation Minute (usually “1” of appearance frequency) is a phrase pair existing in the phrase table 101, and is subtracted from the score of the phrase pair corresponding to i. Next, the generated phrase pair acquisition unit 107 attempts to generate one phrase pair using the phrase pair probability distribution P ^i-1 corresponding to (i-1). If “i = 1”, the generated phrase pair acquisition unit 107 attempts to generate one phrase pair using the probability distribution P _base . The probability distribution P _base is, for example, IBM Model 1. Further, the phrase pair probability distribution can be calculated by, for example, a Pitman-Yor process using the phrase pair frequency (F appearance frequency information) corresponding to (i-1). The phrase pair frequency (F appearance frequency information) is stored in the phrase appearance frequency information storage unit 102. Further, since the calculation of the probability based on the Pitman-Yor process is a known technique, the description thereof is omitted.

  (Step S205) The partial phrase pair generation unit 111 and the like perform phrase generation processing. The phrase generation process is, for example, a phrase generation process using two or more hierarchies using hierarchical ITG. Details of the phrase generation processing will be described with reference to the flowchart of FIG.

  (Step S206) The parallel corpus reception unit 104 increments the counter i by one.

  (Step S207) The parallel corpus reception unit 104 determines whether or not “i <= N” is satisfied. If “i <= N” is satisfied, the process returns to step S202, and if “i <= N” is not satisfied, the process ends.

  Next, details of the phrase generation processing in step S205 will be described using the flowchart of FIG.

  (Step S301) The partial phrase pair generation unit 111 determines whether a phrase pair has been generated in the process of generating the previous phrase pair. If the phrase pair can be generated, the process goes to step S302; otherwise, the process goes to step S305.

  (Step S302) The phrase appearance frequency information update unit 108 increases the F appearance frequency information corresponding to the phrase pair generated in the previous phrase pair generation process by a predetermined value (usually “1”). . When the phrase pair does not exist in the phrase appearance frequency information storage unit 102, the phrase appearance frequency information update unit 108 associates the generated phrase pair with the F appearance frequency information, and the phrase appearance frequency information storage unit 102 To accumulate.

  (Step S303) The score calculation unit 114 calculates the score of the phrase pair corresponding to the updated phrase appearance frequency information. Here, the score calculation unit 114 uses the phrase appearance frequency information corresponding to (i-1) when calculating the score of the phrase pair (see Expressions 1 and 2).

  (Step S304) The phrase table update unit 116 configures a scored phrase pair having the score calculated in step S303, and writes it into the phrase table 101. In addition, when the said phrase pair does not exist in the phrase table 101, the phrase table update part 116 comprises a phrase pair with a score, and adds it to the phrase table 101 newly. If the phrase pair exists in the phrase table 101, the phrase table update unit 116 updates the score corresponding to the phrase pair to the score calculated in step S303, and returns to the upper process (S206). .

(Step S305) The partial phrase pair generation unit 111 uses, for example, a base measure P _dac or a probability distribution P ^i-1 corresponding to the (i−1) -th parallel corpus to be smaller than the phrase pair to be generated 2 Generate one phrase pair.

  (Step S306) The symbol acquisition unit 109 acquires one symbol x using one or more symbol appearance frequency information.

  (Step S307) The symbol appearance frequency information update unit 110 increases the S appearance frequency information corresponding to the symbol x acquired by the symbol acquisition unit 109 by a predetermined value (usually “1”).

  (Step S308) The new phrase pair generation unit 112 determines whether or not the symbol x acquired in step S306 is “BASE”. If the symbol x is “BASE”, go to step S309, and if not “BASE”, go to step S310.

  (Step S309) The new phrase pair generation unit 112 generates a new phrase pair using the prior probability of the phrase pair, and jumps to step S302.

  (Step S310) The new phrase pair generation unit 112 determines whether or not the symbol x acquired in Step S306 is “REG”. If the symbol x is “REG”, the process goes to step S311, and if it is not “REG”, the process goes to step S315. If the symbol x is not “REG”, the symbol x is “INV”.

  (Step S311) The new phrase pair generation unit 112 generates two smaller phrase pairs. In addition, let two phrase pairs here be the 1st phrase pair and the 2nd phrase pair.

  (Step S312) The phrase generation process of FIG. 3 is performed on the first phrase pair generated in Step S311.

  (Step S313) The phrase generation process of FIG. 3 is performed on the second phrase pair generated in Step S311.

  (Step S314) The new phrase pair generation unit 112 sequentially connects the two phrase pairs generated in Step S312 and Step S313, generates one phrase pair, and jumps to Step S302.

  (Step S315) The new phrase pair generation unit 112 generates two smaller phrase pairs. The two phrase pairs here are referred to as a third phrase pair and a fourth phrase pair.

  (Step S316) The phrase generation process of FIG. 3 is performed on the third phrase pair generated in Step S315.

  (Step S317) The phrase generation process of FIG. 3 is performed on the fourth phrase pair generated in Step S315.

  (Step S318) The new phrase pair generation unit 112 concatenates the two phrase pairs generated in steps S316 and S317 in reverse order, generates one phrase pair, and jumps to step S302.

  2 and 3, it is preferable that the generation of the tree structure by the purging unit 115 and the update process of the tree structure by the tree update unit 117 are performed after step S304 and before the return. The tree structure to be updated is the tree structure of the i-th parallel corpus in the parallel corpus.

  Hereinafter, a specific operation of the parallel phrase learning device 1 according to the present embodiment will be described.

  Now, it is assumed that the symbol appearance frequency information storage unit 103 stores the symbols “BASE”, “REG”, “INV”, and three symbol appearance frequency information that are combinations of the appearance frequencies of the respective symbols.

  In this situation, it is assumed that the bilingual corpus accepting unit 104 first accepts the first bilingual corpus and stores the first bilingual corpus in association with 1 in the bilingual information storage unit 100.

  Next, the phrase table initialization unit 106 generates initial information of one or more scored phrase pairs from one or more parallel translation information of the first bilingual corpus, associates it with 1, and stores it in the phrase table 101. accumulate.

Next, the generated phrase pair acquisition unit 107 acquires one parallel translation from the first parallel corpus. Next, the generated phrase pair acquisition unit 107 uses the occurrence of one or more phrase pairs (usually “1” of appearance frequency) constituting the acquired bilingual tree structure as a phrase pair existing in the phrase table 101. Subtract from the score. Next, the generated phrase pair acquisition unit 107 tries to generate the phrase pair <f, e> that is the parallel translation using the probability distribution P _base ¹ . It is assumed that the probability distribution P _base ¹ is estimated in advance by, for example, IBM MODEL 1 and the bilingual phrase learning device 1 holds the probability distribution P _base ¹ in advance.

  And the partial phrase pair production | generation part 111 performs a process as follows, when it is judged that the phrase pair was not able to be produced | generated in the process of the production | generation of a previous phrase pair.

That is, the partial phrase pair generation unit 111 recursively generates two phrase pairs smaller than the phrase pair to be generated using P _base ¹ . And a new phrase pair is produced | generated by combining two smaller phrase pairs produced | generated. It is assumed that the probability of the phrase pair correspondence <f, e> is given by Equation 3.

Also, θ _t ¹ is a parameter of the translation model, and is a table representing all <f, e> probability values. In this case, θ _t ¹ is estimated in the pitman-You process of Equation 4.

Next, the symbol acquisition unit 109 uses the three symbol appearance frequency information to generate a symbol according to the symbol probability distribution P _x (x; θ _x ). Then, the symbol appearance frequency information update unit 110 increases the S appearance frequency information corresponding to the symbol “x = reg” by one.

Next, when the generated symbol x is “x = base”, the new phrase pair generation unit 112 generates a new phrase pair directly from P _base ¹ . When the generated symbol x is “x = reg”, the new phrase pair generation unit 112 generates <f ₁ , e ₁ > and <f ₂ , e ₂ > from the phrase pair generation probability (P _hier ). One phrase pair <f ₁ f ₂ , e ₁ e ₂ > is created. When the generated symbol x is “x = inv”, the new phrase pair generation unit 112 generates <f ₁ , e ₁ > and <f ₂ , e ₂ > from P _hier , and f ₁ and f ₂ Are arranged in reverse order to create one phrase pair <f ₂ f ₁ , e ₁ e ₂ >.

  Then, the phrase appearance frequency information update unit 108 updates the phrase appearance frequency information of the newly created phrase pair.

Further, the score calculation unit 114 calculates the score of the phrase pair corresponding to the updated phrase appearance frequency information using P _base ¹ .

  Next, the phrase table update unit 116 updates the phrase table.

  Next, the purging unit 115 uses the score calculated by the score calculation unit 114 to acquire a new tree structure that maximizes the score of the tree structure. Then, the tree updating unit 117 accumulates the acquired tree structure in the bilingual corpus and updates the old tree structure to the new tree structure.

  With the above processing, for example, a phrase pair with a multi-level granularity can be learned for the phrase pair “Mrs. Smith's red cookbook / Smith ’s red cookbook” as shown in FIG. FIG. 4 is an example of a tree structure constituting the parallel translation information.

  Moreover, the construction method of the phrase table 101 in this specific example is, for example, as follows.

As features of the phrase table, conditional probabilities P _t (f | e) and P _t (e | f), a lexical weighting probability, a phrase penalty, and the like are used. Here, the conditional probability is calculated using the model probability P _t . That is, the conditional probability is calculated using Equation 5 and Equation 6. Then, for example, the score calculation unit 114 calculates a score by multiplying each feature of the phrase table by a predetermined weight and taking the sum of these values. Further, the lexical weighting probability can be calculated using words constituting the phrase. This calculation is a known technique (see P. Koehn, FJOch, and D. Marchu. Statistical phrase-based translation. In Proc. NAACL, pp. 48-54, 2003.). The phrase penalty is “1” for all phrases, for example.

In addition, the term of Σ in Equation 5 represents P _t (e), and among all <e, f>, enumerate phrase pairs having a frequency of 1 or more and matching e. The probability values are totaled. Here, f ^~ ( ^~ is located immediately above f) is f constituting a phrase pair in which e matches among all <e, f>. Further, the term of Σ in Equation 6 represents P _t (f), and among all <e, f>, enumerate phrase pairs having a frequency of 1 or more and matching f. The probability values are totaled. Here, e ^~ ( ^~ exists immediately above f) is e that constitutes a phrase pair in which f matches among all <e, f>. ^{Further, c (<e, f ~} >) indicates the frequency of <e, ^{f ~>. Further, c (<e ~, f} >) indicates the frequency of ^<e ~, f>.

  Note that the phrase table update unit 116 enters only the phrase pair p that appears once or more in the sample into the phrase table 101. In addition, two features are added. The first is the phrase pair coincidence probability Pt (<f, e>) by the model. The second feature is the average posterior probability of a span containing a phrase pair <f, e> based on the posterior probability of the span calculated by the inside-outside algorithm. The span probability is high for a frequently occurring phrase pair or a phrase pair constructed based on a frequently occurring phrase pair, which is useful for determining how reliable the phrase pair is. Phrase extraction based on this model probability is called MOD. The span probability can be calculated by an ITG chart parser.

  Then, the above processing is performed for all the parallel translation sentences included in the first parallel corpus.

  Next, it is assumed that the bilingual corpus accepting unit 104 accepts the second bilingual corpus and stores the second bilingual corpus in association with 2 in the bilingual information storage unit 100.

  Next, the phrase table initialization unit 106 generates initial information of one or more phrase pairs with scores from one or more pieces of parallel translation information included in the second parallel corpus, associates it with 2, and stores it in the phrase table 101. accumulate.

Next, the generated phrase pair acquisition unit 107 acquires one parallel translation from the second parallel corpus. Next, the generated phrase pair acquisition unit 107 uses the occurrence of one or more phrase pairs (usually “1” of appearance frequency) constituting the acquired bilingual tree structure as a phrase pair existing in the phrase table 101. Subtract from the score. Then, generated phrase pair acquisition unit 107, a phrase pair is the translation sentence <f, e> and attempts to generate using the probability distribution P ^1. Probability distribution P ¹ is a probability distribution obtained in the above process for the first bilingual corpus.

  And the partial phrase pair production | generation part 111 performs a process as follows, when it is judged that the phrase pair was not able to be produced | generated in the process of the production | generation of a previous phrase pair.

In other words, partial phrase pair generation unit 111 uses the P ^1, recursively generates a tries to create a phrase pair is less than two phrases pairs. And a new phrase pair is produced | generated by combining two smaller phrase pairs produced | generated. It is assumed that the second translation model (θ _t ² ) is estimated in the pitman-You process of Equation 7.

Next, the symbol acquisition unit 109 uses the three symbol appearance frequency information to generate a symbol according to the symbol probability distribution P _x (x; θ _x ). Then, the symbol appearance frequency information update unit 110 increases the S appearance frequency information corresponding to the symbol “x = reg” by one.

Then, if the generated symbol x is "x = base", the new phrase pair generation unit 112, to generate direct a new phrase pair from P ^1. When the generated symbol x is “x = reg”, the new phrase pair generation unit 112 generates <f ₁ , e ₁ > and <f ₂ , e ₂ > from the phrase pair generation probability (P _hier ). One phrase pair <f ₁ f ₂ , e ₁ e ₂ > is created. When the generated symbol x is “x = inv”, the new phrase pair generation unit 112 generates <f ₁ , e ₁ > and <f ₂ , e ₂ > from P _hier , and f ₁ and f ₂ Are arranged in reverse order to create one phrase pair <f ₂ f ₁ , e ₁ e ₂ >.

  Then, the phrase appearance frequency information update unit 108 updates the phrase appearance frequency information of the newly created phrase pair. The phrase appearance frequency information is phrase appearance frequency information corresponding to the second parallel corpus.

Furthermore, the score calculation unit 114, a score of phrase pair corresponding to the updated phrase appearance frequency information is calculated using the P ^1.

  Next, the phrase table update unit 116 updates the phrase table corresponding to the second parallel corpus.

  Next, the purging unit 115 uses the score calculated by the score calculation unit 114 to acquire a new tree structure that maximizes the score of the tree structure. Then, the tree updating unit 117 accumulates the acquired tree structure in the bilingual corpus and updates the old tree structure to the new tree structure. This tree structure is a tree structure corresponding to the second parallel corpus.

  Then, the above processing is performed for all the translations included in the second parallel corpus. In the phrase table 101, one or more phrase pairs with scores corresponding to 2 are accumulated.

It is assumed that the above processing is also performed for the third and subsequent parallel corpora. Then, it is assumed that the phrase table 101 stores a large number of scored phrase pairs corresponding to the first to (j−1) th parallel corpora. For example, it is assumed that the probability distribution of the (j−1) th many phrase pairs is P ^j−1 . Here, j is a natural number of 3 or more.

  In this situation, it is assumed that the bilingual corpus accepting unit 104 accepts the j-th bilingual corpus and stores the j-th bilingual corpus in association with j in the bilingual information storage unit 100.

  Next, the phrase table initialization unit 106 generates initial information of one or more scored phrase pairs from one or more pieces of parallel translation information included in the j-th parallel corpus, and associates it with j in the phrase table 101. accumulate.

Next, the generated phrase pair acquisition unit 107 acquires one parallel translation from the parallel corpus. Next, the generated phrase pair acquisition unit 107 uses the occurrence of one or more phrase pairs (usually “1” of appearance frequency) constituting the acquired bilingual tree structure as a phrase pair existing in the phrase table 101. Subtract from the score. Next, the generated phrase pair acquisition unit 107 tries to generate the phrase pair <f, e>, which is the parallel translation, from the probability distribution P ^J-1 of the (j−1) th phrase pair set. The probability distribution P ^J-1 is a probability distribution acquired by the process for the (j−1) -th parallel corpus.

  And the partial phrase pair production | generation part 111 performs a process as follows, when it is judged that the phrase pair was not able to be produced | generated in the process of the production | generation of a previous phrase pair.

That is, the partial phrase pair production | generation part 111 produces | generates two phrase pairs smaller than the phrase pair which it was going to produce | generate recursively using probability distribution ^Pj-1 . And a new phrase pair is produced | generated by combining two smaller phrase pairs produced | generated. Note that θ _t ^j is estimated in the pitman-You process of Equation 8.

Next, the symbol acquisition unit 109 uses the three symbol appearance frequency information to generate a symbol according to the symbol probability distribution P _x (x; θ _x ). Then, the symbol appearance frequency information update unit 110 increases the S appearance frequency information corresponding to the symbol “x = reg” by one.

Then, if the generated symbol x is "x = base", the new phrase pair generation unit 112, directly generate a new phrase pair from P ^1. When the generated symbol x is “x = reg”, the new phrase pair generation unit 112 generates <f ₁ , e ₁ > and <f ₂ , e ₂ > from the phrase pair generation probability (P _hier ). One phrase pair <f ₁ f ₂ , e ₁ e ₂ > is created. When the generated symbol x is “x = inv”, the new phrase pair generation unit 112 generates <f ₁ , e ₁ > and <f ₂ , e ₂ > from P _hier , and f ₁ and f ₂ Are arranged in reverse order to create one phrase pair <f ₂ f ₁ , e ₁ e ₂ >.

  Then, the phrase appearance frequency information update unit 108 updates the phrase appearance frequency information of the newly created phrase pair. The phrase appearance frequency information is phrase appearance frequency information corresponding to the j-th parallel corpus.

Moreover, the score calculation part 114 calculates the score of the phrase pair corresponding to the updated phrase appearance frequency information using ^Pj-1 .

  Next, the phrase table update unit 116 updates the phrase table corresponding to the j-th parallel corpus.

  Next, the purging unit 115 uses the score calculated by the score calculation unit 114 to acquire a new tree structure that maximizes the score of the tree structure. Then, the tree updating unit 117 accumulates the acquired tree structure in the bilingual corpus and updates the old tree structure to the new tree structure. This tree structure is a tree structure corresponding to the second parallel corpus.

  Then, the above processing is performed for all the translation sentences included in the j-th parallel corpus. In the phrase table 101, one or more phrase pairs with scores associated with j are accumulated.

  As described above, according to the present embodiment, the translation model generated from the added bilingual corpus can be connected to the original translation model, and as a result, the translation model can be easily enriched step by step.

  Moreover, according to this Embodiment, the size of a phrase table can be reduced significantly, maintaining the precision of the machine translation using the phrase table created with the parallel translation phrase learning apparatus 1. FIG. That is, according to the present embodiment, many appropriate phrase pairs can be learned.

Note that the processing in the present embodiment may be realized by software. Then, this software may be distributed by software download or the like. Further, this software may be recorded and distributed on a recording medium such as a CD-ROM. This also applies to other embodiments in this specification. Note that the software that implements the information processing apparatus according to the present embodiment is the following program. That is, in this program, the computer-accessible recording medium can store bilingual corpus of N (N is a natural number of 2 or more) having one or more bilingual information having a bilingual sentence and a bilingual tree structure. A phrase pair that is a pair of an information storage unit, a first language phrase having one or more words in the first language, and a second language phrase having one or more words in the second language, and the appearance probability of the phrase pair One or more phrase appearance frequency information having a phrase table that can store one or more scored phrase pairs having a score that is information, a phrase pair, and F appearance frequency information that is information related to the appearance frequency of the phrase pair A phrase appearance frequency information storage unit that can store each bilingual corpus, a symbol that identifies a method for generating a new phrase pair, and the symbol A symbol appearance frequency information storage unit that can store one or more symbol appearance frequency information having S appearance frequency information that is information relating to the appearance frequency, and the computer displays the one or more phrase appearances for each bilingual corpus Using the frequency information, a generated phrase pair acquisition unit that acquires a phrase pair having a first language phrase and a second language phrase;
When a phrase pair can be acquired, the F appearance frequency information corresponding to the phrase pair is increased by a predetermined value, and when the phrase pair cannot be acquired, one or more symbols appear. A symbol acquisition unit that acquires one symbol using the frequency information; a symbol appearance frequency information update unit that increases S appearance frequency information corresponding to the symbol acquired by the symbol acquisition unit by a predetermined value; If a phrase pair cannot be acquired, a partial phrase pair generation unit that generates two phrase pairs smaller than the phrase pair that was to be acquired, and a first that generates a new phrase pair according to the symbols acquired by the symbol acquisition unit Or generating two smaller phrase pairs and using the one or more phrase appearance frequency information, One phrase pair having a new first language phrase in which the two first language phrases constituting the pair are sequentially connected and a new second language phrase in which the two second language phrases constituting the two phrase pairs are sequentially connected The second process of generating the two or two smaller phrase pairs, and using the one or more phrase appearance frequency information, the two first language phrases constituting the generated two phrase pairs are connected in order A third process of generating a single phrase pair having a new first language phrase and a new second language phrase in which the two second language phrases constituting the two phrase pairs are connected in reverse order. For the phrase pair generated by the new phrase pair generation unit and the new phrase pair generation unit, the phrase appearance frequency information update unit, the symbol capture Stored in the phrase appearance frequency information storage unit, a control unit that instructs to recursively perform the processing of the symbol appearance frequency information update unit, the partial phrase pair generation unit, and the new phrase pair generation unit. A score calculator that calculates a score for each phrase pair in the phrase table using one or more phrase appearance frequency information, and a phrase table update that stores the score calculated by the score calculator in association with each phrase pair The score calculation unit corresponds to the (j-1) th parallel corpus when calculating the score for each phrase pair acquired from the j (2 <= j <= N) th parallel corpus. Calculating the score for each phrase pair corresponding to the j-th parallel corpus using one or more phrase appearance frequency information And a program that causes the computer to function.

  The bilingual information storage unit stores one or more bilingual corpora, and in the upper program, the bilingual corpus receiving unit that receives the bilingual corpus and the bilingual corpus received by the bilingual corpus receiving unit are the bilingual information storing unit. A bilingual corpus accumulating unit for accumulating the generated bilingual corpus in the bilingual corpus after the bilingual corpus accumulating unit accumulates the accepted bilingual corpus in the bilingual information storage unit. Instructing the processing of the unit, the phrase appearance frequency information update unit, the symbol acquisition unit, the symbol appearance frequency information update unit, the partial phrase pair generation unit, and the new phrase pair generation unit, the score calculation unit Is a score for each phrase pair acquired from the parallel corpus received by the parallel corpus reception unit. When calculating, one or more phrase appearance frequency information corresponding to any one of the one or more parallel corpora stored in the parallel translation information storage before the parallel corpus is stored in the parallel corpus It is preferable to make the computer function as calculating a score for each phrase pair corresponding to the bilingual corpus received by the bilingual corpus receiving unit.

(Embodiment 2)
In the present embodiment, a bilingual phrase learning apparatus that learns a plurality of fields independently, replaces prior probabilities of each field with models obtained in other fields, and connects a plurality of models hierarchically will be described.

  FIG. 5 is a block diagram of the bilingual phrase learning device 2 in the present embodiment. As shown in FIG. 5, the bilingual phrase learning device 2 is different from the bilingual phrase learning device 1 in the bilingual corpus generation unit 201, and does not include the bilingual corpus reception unit 104 and the bilingual corpus accumulation unit 105.

  The bilingual corpus generation unit 201 divides two or more bilingual sentences into N groups, and acquires N bilingual corpora created by acquiring a bilingual tree structure from the bilingual sentences of each group in the bilingual information storage unit 100. accumulate. N is a natural number of 2 or 3 or more. The dividing method is not limited. In the case where a class identifier for identifying a class is assigned to the parallel translation, the parallel corpus generation unit 201 may divide two or more parallel translations into N groups using the class identifier. Further, the bilingual corpus generation unit 201 may divide two or more bilingual sentences into N groups so as to include the same number of bilingual sentences.

  The bilingual corpus generation unit 201 can usually be realized by an MPU, a memory, or the like. The processing procedure of the bilingual corpus generation unit 201 is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  Next, operation | movement of the parallel translation phrase learning apparatus 2 is demonstrated using the flowchart of FIG. In the flowchart of FIG. 6, the description of the same steps as those in the flowchart of FIG. 2 is omitted.

  (Step S601) The bilingual corpus generation unit 201 divides two or more bilingual sentences stored in the bilingual information storage unit 100 into N groups. Each group has a bilingual corpus having one or more bilingual sentences.

  (Step S602) The bilingual corpus generation unit 201 configures a tree structure of one or more bilingual sentences in each group and accumulates them in the bilingual information storage unit 100. With this processing, the final translation corpus of N groups is stored in the translation information storage unit 100. The bilingual corpus here has one or more bilingual information having a bilingual sentence and a tree structure of the bilingual sentence.

  (Step S603) The phrase table initialization unit 106 acquires the i-th parallel corpus from the parallel translation information storage unit 100. Go to step S203.

  In the flowchart of FIG. 6, when a phrase table corresponding to the bilingual corpus of the jth group is configured, the phrase table corresponding to the bilingual corpus of the (j−1) th group is usually used. However, when the phrase table corresponding to the bilingual corpus of the jth group is configured, the phrase table has already been acquired, and the phrase table of another group (for example, the phrase table corresponding to the bilingual corpus of the third group) ) May be used.

  As described above, according to the present embodiment, the translation model generated from the added bilingual corpus can be connected to the original translation model, and as a result, the translation model can be easily enriched step by step.

  As described above, according to the present embodiment, parallelization is facilitated by, for example, dividing bilingual data for each field and learning a local model in each field. Further, according to the present embodiment, when combining the statistical models obtained as a result of learning, it is not necessary to recalculate the weights, and the translation model can be easily enriched.

  In addition, if the bilingual phrase learning apparatus 1 demonstrated in the said embodiment is used, it will become possible as follows. That is, according to the bilingual phrase learning device 1 or the bilingual phrase learning device 2, the bilingual data is newly updated every day, and when bilingual data is newly added to existing large-scale bilingual data, the cost of re-learning is greatly increased. It becomes possible to make it low. In particular, when dealing with new fields such as patent data, the bilingual phrase learning device 1 or the bilingual phrase learning device 2 can easily learn an existing statistical model as a prior probability and newly estimate the parameters of the model. Become. In addition, the parallel phrase learning device 1 or the parallel phrase learning device 2 adds new bilingual data each time a new expression appears in daily conversation, etc., and has a general model as a prior probability. It becomes possible to create.

  Note that the software that implements the information processing apparatus according to the present embodiment is the following program. That is, the computer-accessible recording medium includes a parallel translation information storage unit that can store N (N is a natural number of 2 or more) parallel corpus having one or more parallel translation information having a parallel translation sentence and a parallel translation tree structure. A score that is information about a phrase pair that is a pair of a first language phrase having one or more words in the first language and a second language phrase having one or more words in the second language and the appearance probability of the phrase pair One or more phrase appearance frequency information having a phrase table that can store one or more scored phrase pairs having the following, a phrase pair, and F appearance frequency information that is information regarding the appearance frequency of the phrase pair, A phrase appearance frequency information storage unit that can be stored every time, a symbol that identifies a method of generating a new phrase pair, and an appearance frequency of the symbol A symbol appearance frequency information storage unit capable of storing one or more symbol appearance frequency information having S appearance frequency information as information, and the program causes the computer to execute the one or more phrases for each parallel corpus. Using the appearance frequency information, if a phrase pair acquisition unit that acquires a phrase pair having a first language phrase and a second language phrase and a phrase pair can be acquired, F appearance frequency information corresponding to the phrase pair is obtained. A phrase appearance frequency information update unit that increases by a predetermined value, a symbol acquisition unit that acquires one symbol using one or more symbol appearance frequency information when the phrase pair cannot be acquired, and the symbol The symbol appearance frequency information update unit that increases the S appearance frequency information corresponding to the symbol acquired by the acquisition unit by a predetermined value and the phrase pair cannot be acquired. A partial phrase pair generation unit that generates two phrase pairs smaller than the phrase pair to be acquired, and a first process for generating a new phrase pair according to the symbol acquired by the symbol acquisition unit, or 2 Two smaller phrase pairs are generated, and using the one or more phrase appearance frequency information, a new first language phrase in which two first language phrases constituting the generated two phrase pairs are sequentially connected, and two A second process for generating one phrase pair having a new second language phrase in which two second language phrases constituting the phrase pair are sequentially connected, or two smaller phrase pairs are generated, and the one or more Using the phrase appearance frequency information of, the two first language phrases that make up the two generated phrase pairs are connected in order A new one that performs one of the third processes of generating one phrase pair having a first language phrase and a new second language phrase in which two second language phrases constituting two phrase pairs are connected in reverse order For the phrase pair generated by the phrase pair generation unit and the new phrase pair generation unit, the phrase appearance frequency information update unit, the symbol acquisition unit, the symbol appearance frequency information update unit, the partial phrase pair generation unit, and Each phrase pair in the phrase table is determined using a control unit that instructs to recursively perform the processing of the new phrase pair generation unit, and one or more phrase appearance frequency information stored in the phrase appearance frequency information storage unit. A score calculation unit that calculates a score for the phrase, and a phrase that stores the score calculated by the score calculation unit in association with each phrase pair When the score calculation unit calculates the score for each phrase pair acquired from the j (2 <= j <= N) th parallel corpus, the (j-1) th parallel corpus Is a program that causes a computer to function as a score for each phrase pair corresponding to the j-th parallel corpus using one or more phrase appearance frequency information corresponding to.

  Further, in the upper program, two or more parallel translations are divided into N groups, and N parallel corpora created by acquiring the parallel translation tree from the parallel translations of each group are stored in the parallel translation information storage unit. The bilingual corpus generation unit further functions, and the score calculation unit corresponds to another bilingual corpus different from the one bilingual corpus when calculating the score for each phrase pair acquired from the one bilingual corpus. It is preferable to cause the computer to function as a score for each phrase pair corresponding to the one bilingual corpus using one or more phrase appearance frequency information.

(Embodiment 3)
In the present embodiment, a statistical machine translation apparatus 3 using a phrase table 101 learned using the parallel phrase learning apparatus 1 or the parallel phrase learning apparatus 2 will be described.

  FIG. 7 is a block diagram of the statistical machine translation apparatus 3 in the present embodiment. The statistical machine translation apparatus 3 includes a phrase table 101, a reception unit 301, a phrase acquisition unit 302, a sentence composition unit 303, and an output unit 304.

  The phrase table 101 is a phrase table learned by the bilingual phrase learning device 1 or the bilingual phrase learning device 2.

  The accepting unit 301 accepts a sentence in a first language having one or more words. Here, reception means reception of information input from an input device such as a keyboard, mouse, touch panel, reception of information transmitted via a wired or wireless communication line, recording on an optical disk, magnetic disk, semiconductor memory, or the like. It is a concept including reception of information read from a medium. The first language sentence input means may be anything such as a keyboard, mouse, or menu screen. The accepting unit 301 can be realized by a device driver for input means such as a keyboard, control software for a menu screen, or the like.

  The phrase acquisition unit 302 extracts one or more phrases from the sentence received by the reception unit 301, and acquires one or more phrases in the second language from the phrase table 101 using the score of the phrase table 101. Note that the processing of the phrase acquisition unit 302 is a known technique.

  The sentence composition unit 303 composes a sentence in the second language from one or more phrases acquired by the phrase acquisition unit 302. The processing of the sentence composition unit 303 is a known technique.

  The output unit 304 outputs the sentence constructed by the sentence composition unit 303. Here, output refers to display on a display, projection using a projector, printing on a printer, sound output, transmission to an external device, storage in a recording medium, output to another processing device or other program, etc. It is a concept that includes delivery of processing results.

  The phrase acquisition unit 302 and the sentence composition unit 303 can be usually realized by an MPU, a memory, or the like. The processing procedure of the phrase acquisition unit 302 and the like is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The output unit 304 may or may not include an output device such as a display or a speaker. The output unit 304 can be implemented by output device driver software, or output device driver software and an output device.

  Further, the operation of the statistical machine translation apparatus 3 suffices to perform a well-known phrase-based statistical machine translation process, and a detailed description thereof will be omitted.

  As described above, according to the present embodiment, it is possible to perform machine translation with high accuracy by using hierarchically connected phrase tables.

Hereinafter, the experimental result of the parallel phrase learning device 1 or the parallel phrase learning device 2 will be described.
(Experiment)

  The information of the data set used in this experiment is shown in FIG. In this experiment, the task of Chinese-English translation (translation from Chinese to English) was used. In this experiment, three data sets having different sizes shown in FIG. 8 were used. In FIG. 8, “Data set” is the name of the data set, “Corpus” is the name of the corpus in the data set, and “# sent.pairs” is the number of parallel translations.

  The data set “IWSLT” in FIG. 8 is a data set used in IWSLT2012 OLYMPICS, and consists of two training sets (HIT corpus and BTEC corpus). The HIT Corpus is closely related to the 2008 Beijing Olympics. The BTEC corpus is a multilingual speech corpus that includes tourism-related sentences.

  The data set “FBIS” in FIG. 8 is a set of news articles and does not have domain (field) information. Therefore, the entire corpus was divided into five domains using a latent directory distribution (LDA) tool called PLDA (see http://code.google.com/p/plda/). In each of the five domains, the source side (first language side) and the target side (second language side) are connected like a single sentence (Zhiyuan Liu, Yuzhou Zhang, Edward Y Chang, and Maosong Sun. 2011. Plda +: Parallel latent dirichlet allocation with data placement and pipeline processing. See ACM Transactions on Intelligent Systems and Technology (TIST), 2 (3): 1-18.

  The data set “LDC” in FIG. 8 includes various domains such as news, magazines, and finance, and is composed of five corpora acquired from the LDC.

In this experiment, the following five phrase pair extraction methods were used to evaluate the parallel phrase learning device 1 or the parallel phrase learning device 2. The method of the parallel phrase learning device 1 or the parallel phrase learning device 2 is referred to as “Hier-combin”.
(1) GIZA-linear

In this method, phrase pairs consist of GIZA ++ (see Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational linguistics, 29 (1): 19-51.) And a maximum length of 7 "grow- Extracted within each domain using the "diag-final-and" method. In this method, phrase tables composed of various domains are linearly combined by averaging feature values.
(2) Pialign-linear

This method is similar to GIZA-linear, but pialign toolkit (Graham Neubig, Taro Watanabe, Eiichiro Sumita, Shinsuke Mori, and Tatsuya Kawahara. 2011. An unsupervised model for joint phrase alignment and extraction. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 632-641, Portland, Oregon, USA, June. Association for Computational Linguistics.) And different. Also in this method, the extracted phrase pairs are linearly combined by averaging feature quantities.
(3) GIZA-batch

In this method, the data set is not divided into domains, but treated as a single corpus. In this method, a heuristic GIZA-based phrase extraction method similar to GIZA-linear is executed.
(4) Pialign-batch

In this method, similar to GIZA-batch, a single model is estimated as a single merged corpus. Since Pialign cannot handle large-scale data, it was not possible to experiment with the largest LDC data set.
(5) Pialign-adaptive

  In this method, alignment and phrase pairs are extracted by the same method as Pialign-batch. In this method, the translation probability is estimated by an adaptive method using monolingual topic information.

  Further, in the method “Hier-combin” of the parallel phrase learning device 1 or the parallel phrase learning device 2, phrase pairs were extracted by the same method as “Pialign-linear”. In the phrase table connection process, the translation probability of each phrase pair is estimated by “Hier-combin”. The other features are linearly combined by averaging feature amounts. “Pialign” uses default parameters. The parameter “samps” is set to 5. “5” indicates that five samples are generated for one parallel sentence.

Furthermore, in this experiment, batch-MIRA (Colin Cherry and George Foster. 2012. Batch tuning strategies for statistical machine translation. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 427-436, Montr´eal, Canada, June. Association for Computational Linguistics.). In order to evaluate translation quality, case-insensitive BLEU-4 metrics (Kishore Papineni, Salim Roukos, ToddWard, and Wei Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA, July. Association for Computational Linguistics.).
FIG. 9 shows the results of experiments conducted in the above environment. In FIG. 9, “BLEU” indicates an evaluation value of translation quality, and “Size” indicates the number of phrase pairs. In FIG. 9, it can be seen that “Hier-combin” is superior to “Pialign-linear”. Note that “Hier-combin” and “Pialign-linear” only differ in translation probability, and the number of phrase pairs and phrase pairs is the same.

  The performance of “Pialign-adaptive” is higher than that of “Pialign-linear” and lower than that of “Hier-combin”. This proves that the adaptive method using monolingual topic information is useful for tasks. However, “Hier-combin” using a hierarchical Pitman-yor process can estimate a more accurate translation probability based on all data from various domains. That is, in FIG. 9, it can be seen that “Hier-combin” has a high translation quality evaluation for various data sets with a relatively small number of phrase pairs as compared to other methods.

  In more detail, in FIG. 9, it can be seen that “Hier-combin” achieves competitive performance using a much smaller phrase table compared to “GIZA-batch”. . The number of phrase pairs generated by the “Hier-combin” method is 73.9%, 52.7%, and 45.4% for each data set compared to that of “GIZA-batch”. It is getting smaller.

  In the IWLST2012 data set, there is a large difference between the HIT corpus and the BTEC corpus, and the “Hier-combin” method improved 0.814 BLEU value compared to the “Pialign-batch” method. On the other hand, in the FBIS data set, the “Hier-combin” method is 0.09 compared to the “GIZA-batch” method because it is artificially divided into subdomains and the clear allocation difference is not clear. The BLEU value was low.

  In “Hier-combin”, phrase pairs can be acquired independently from each of a plurality of domains. Therefore, in "Hier-combin", processing can be performed on different machines for each domain, and parallel processing is also possible.

  FIG. 10 shows the time taken for alignment and phrase pair extraction when the “FBIS” data set is used. In FIG. 10, “Batch” is a batch-based ITGs sampling method (“pialign-batch”). FIG. 10 shows experimental results using a 2.7 GHz E5-2680 CPU and 128 GB memory. In FIG. 10, “Parallel Extraction” is the time when parallel processing is performed, “Integrating” is the time of concatenation processing, and “Total” is the total time of parallel processing and concatenation processing.

  Comparing “Hier-combin” and “pialign-batch”, according to FIG. 10, the time spent for training in “Hier-combin” is much more than almost a quarter of “pialign-batch”. Very few. On the other hand, according to FIG. 9, “Hier-combin” has a slightly higher BLEU value than “pialign-batch”.

  The “Hier-combin” method of hierarchical connection uses the property of the hierarchical Pitman-Yor process. And the “Hier-combin” method has the advantage of smoothing effect. Also, by using the “Hier-combin” method, it is possible to generate a simple phrase table based on all data from various domains with a more accurate probability step by step. The phrase pair extraction in traditional SMT is batch-based, but the “Hier-combin” method can extract phrase pairs very efficiently, while the translation accuracy is not inferior to the traditional SMT method. It can be said.

  FIG. 11 shows BLEU values when different combining methods are used for each of the three data sets in the “Hier-combin” method. In FIG. 11, the similarity is arranged as a key, “Descending” is descending and “Ascending” is ascending. The similarity between data was calculated using a perplexity index using a 5-gram language model.

  FIG. 12 shows the external appearance of a computer that executes the program described in this specification to realize the bilingual phrase learning apparatus and the like of the various embodiments described above. The above-described embodiments can be realized by computer hardware and a computer program executed thereon. FIG. 12 is an overview diagram of the computer system 300, and FIG. 13 is a block diagram of the system 300.

  In FIG. 12, a computer system 300 includes a computer 301 including a CD-ROM drive, a keyboard 302, a mouse 303, and a monitor 304.

  In FIG. 13, in addition to the CD-ROM drive 3012, the computer 301 includes an MPU 3013, a bus 3014 connected to the MPU 3013 and the CD-ROM drive 3012, a ROM 3015 for storing a program such as a bootup program, and an MPU 3013. And a RAM 3016 for temporarily storing instructions of the application program and providing a temporary storage space, and a hard disk 3017 for storing the application program, the system program, and data. Although not shown here, the computer 301 may further include a network card that provides connection to a LAN.

  A program that causes the computer system 300 to execute the functions of the bilingual phrase learning apparatus and the like of the above-described embodiment may be stored in the CD-ROM 3101, inserted into the CD-ROM drive 3012, and further transferred to the hard disk 3017. . Alternatively, the program may be transmitted to the computer 301 via a network (not shown) and stored in the hard disk 3017. The program is loaded into the RAM 3016 at the time of execution. The program may be loaded directly from the CD-ROM 3101 or the network.

  The program does not necessarily include an operating system (OS), a third-party program, or the like that causes the computer 301 to execute the functions of the bilingual phrase learning device of the above-described embodiment. The program only needs to include an instruction portion that calls an appropriate function (module) in a controlled manner and obtains a desired result. How the computer system 300 operates is well known and will not be described in detail.

  Further, the computer that executes the program may be singular or plural. That is, centralized processing may be performed, or distributed processing may be performed.

  In each of the above embodiments, it is needless to say that two or more communication means (terminal information transmission unit, terminal information reception unit, etc.) existing in one device may be physically realized by one medium. .

  In each of the above embodiments, each process (each function) may be realized by centralized processing by a single device (system), or by distributed processing by a plurality of devices. May be.

  The present invention is not limited to the above-described embodiments, and various modifications are possible, and it goes without saying that these are also included in the scope of the present invention.

  As described above, the translation phrase learning device according to the present invention can easily enhance the translation model step by step by connecting the translation model generated from the added translation corpus to the original translation model. It is useful as an apparatus for machine translation.

1, 2 Bilingual phrase learning device 3 Statistical machine translation device 100 Bilingual information storage unit 101 Phrase table 102 Phrase appearance frequency information storage unit 103 Symbol appearance frequency information storage unit 104 Bilingual corpus reception unit 105 Bilingual corpus storage unit 106 Phrase table initialization Unit 107 generated phrase pair acquisition unit 108 phrase appearance frequency information update unit 109 symbol acquisition unit 110 symbol appearance frequency information update unit 111 partial phrase pair generation unit 112 new phrase pair generation unit 113 control unit 114 score calculation unit 115 parsing unit 116 phrase table Update unit 117 Tree update unit 201 Bilingual corpus generation unit

Claims (7)

A bilingual information storage unit capable of storing N (N is a natural number of 2 or more) bilingual corpus having one or more bilingual information having a bilingual sentence and a bilingual tree structure;
A phrase pair that is a pair of a first language phrase having one or more words in the first language and a second language phrase having one or more words in the second language, and a score that is information relating to the appearance probability of the phrase pair; A phrase table that can store one or more scored phrase pairs with each bilingual corpus;
A phrase appearance frequency information storage unit capable of storing, for each bilingual corpus, one or more phrase appearance frequency information including a phrase pair and F appearance frequency information that is information relating to the appearance frequency of the phrase pair;
A symbol appearance frequency information storage unit that can store one or more symbol appearance frequency information including a symbol for identifying a method for generating a new phrase pair and S appearance frequency information that is information about the appearance frequency of the symbol;
For each bilingual corpus, using the one or more phrase appearance frequency information, a generated phrase pair acquisition unit that acquires a phrase pair having a first language phrase and a second language phrase;
When the phrase pair can be acquired, the phrase appearance frequency information update unit that increases the F appearance frequency information corresponding to the phrase pair by a predetermined value;
If the phrase pair could not be acquired, a symbol acquisition unit that acquires one symbol using one or more symbol appearance frequency information,
A symbol appearance frequency information update unit that increases S appearance frequency information corresponding to the symbol acquired by the symbol acquisition unit by a predetermined value;
If the phrase pair could not be acquired, a partial phrase pair generation unit that generates two phrase pairs smaller than the phrase pair to be acquired;
In accordance with the symbol acquired by the symbol acquisition unit, a first process for generating a new phrase pair, or two smaller phrase pairs are generated, and the generated two phrases are generated using the one or more phrase appearance frequency information. One phrase having a new first language phrase in which two first language phrases constituting a phrase pair are sequentially connected and a new second language phrase in which two second language phrases constituting two phrase pairs are sequentially connected 2nd process which produces | generates a pair, or two smaller phrase pairs are produced | generated, The two 1st language phrases which comprise the produced | generated two phrase pairs are sequentially used using said one or more phrase appearance frequency information A new second language phrase that connects the new first language phrase connected together and the two second language phrases that make up two phrase pairs in reverse order. And new phrase pair generation unit that performs one of a third process of generating a single phrase pair having a's,
For the phrase pair generated by the new phrase pair generation unit, the phrase appearance frequency information update unit, the symbol acquisition unit, the symbol appearance frequency information update unit, the partial phrase pair generation unit, and the new phrase pair generation unit A control unit instructing to perform the process of recursively,
A score calculation unit that calculates a score for each phrase pair in the phrase table, using one or more phrase appearance frequency information stored in the phrase appearance frequency information storage unit;
A phrase table update unit that stores the score calculated by the score calculation unit in association with each phrase pair;
The score calculation unit
When calculating a score for each phrase pair acquired from the j (2 <= j <= N) th parallel corpus, using one or more phrase appearance frequency information corresponding to the (j-1) th parallel corpus , A parallel phrase learning device for calculating a score for each phrase pair corresponding to the j-th parallel corpus. The bilingual information storage unit
Contains one or more parallel corpora,
A bilingual corpus accepting unit that accepts a bilingual corpus;
A bilingual corpus storage unit that stores the bilingual corpus received by the bilingual corpus receiving unit in the bilingual information storage unit;
The controller is
After the bilingual corpus accumulation unit accumulates the accepted bilingual corpus in the bilingual information storage unit, the generated phrase pair acquisition unit for the bilingual corpus, the phrase appearance frequency information update unit, the symbol acquisition unit, the symbol appearance Instructing the processing of the frequency information update unit, the partial phrase pair generation unit, and the new phrase pair generation unit,
The score calculation unit
When calculating the score for each phrase pair acquired from the bilingual corpus received by the bilingual corpus receiving unit, one or more stored in the bilingual information storage unit before the bilingual corpus accumulating unit accumulates the bilingual corpus The bilingual translation of Claim 1 which calculates the score with respect to each phrase pair corresponding to the bilingual corpus which the said bilingual corpus received using the one or more phrase appearance frequency information corresponding to one of the bilingual corpus of Phrase learning device. A bilingual corpus generation unit that accumulates N bilingual corpora created by dividing two or more bilingual sentences into N groups and acquiring a bilingual sentence tree structure from the bilingual sentences of each group in the bilingual information storage unit; In addition,
The score calculation unit
When calculating the score for each phrase pair acquired from one bilingual corpus, using one or more phrase appearance frequency information corresponding to another bilingual corpus different from the one bilingual corpus, The bilingual phrase learning device according to claim 1, wherein a score for each corresponding phrase pair is calculated. The score calculation unit
The bilingual phrase learning apparatus according to any one of claims 1 to 3, wherein a score for each phrase pair corresponding to each bilingual corpus is calculated using a hierarchical Chinese restaurant process according to Formula 9.
In Equation 9, f is a source language phrase, e is a target language phrase, F is a source language sentence, E is a target language sentence, C ^j is all customers in the j-th parallel translation data <F, E>, s ^j is a strength corresponding to the j-th parallel translation data, d ^j is a discount corresponding to the j-th parallel translation data, T ^j is all tables in the parallel translation data <F, E>, and c ^j _{<f, e>} is j The number of customers corresponding to each <f, e> in the j-th translation data, t ^j is the number of tables corresponding to each <f, e> in the j-th translation data, P _base (<f _i , e _i >) is the prior probability of the model estimated in advance. A phrase table learned by the bilingual phrase learning device according to any one of claims 1 to 4,
An accepting unit that accepts a sentence in a first language having one or more words;
A phrase acquisition unit that extracts one or more phrases from the sentence received by the reception unit, and acquires one or more phrases of a second language from the phrase table using a score of the phrase table;
A sentence constructing unit that constructs a sentence in a second language from one or more phrases acquired by the phrase acquiring unit;
A statistical machine translation apparatus comprising: an output unit that outputs a sentence formed by the sentence composing unit. The recording medium is
A bilingual information storage unit capable of storing N (N is a natural number of 2 or more) bilingual corpus having one or more bilingual information having a bilingual sentence and a bilingual tree structure;
A phrase pair that is a pair of a first language phrase having one or more words in the first language and a second language phrase having one or more words in the second language, and a score that is information relating to the appearance probability of the phrase pair; A phrase table that can store one or more scored phrase pairs having
A phrase appearance frequency information storage unit capable of storing, for each bilingual corpus, one or more phrase appearance frequency information including a phrase pair and F appearance frequency information that is information relating to the appearance frequency of the phrase pair;
A symbol appearance frequency information storage unit capable of storing at least one symbol appearance frequency information including a symbol for identifying a method for generating a new phrase pair and S appearance frequency information that is information relating to the appearance frequency of the symbol. ,
Phrase appearance frequency information update unit, symbol acquisition unit, symbol appearance frequency information update unit, partial phrase pair generation unit, new phrase pair generation unit, control unit, score calculation unit, parallel translation phrase learning method that can be realized by the phrase table update unit There,
The generated phrase pair acquisition unit acquires a phrase pair having a first language phrase and a second language phrase using the one or more phrase appearance frequency information for each parallel corpus;
When the phrase appearance frequency information update unit can acquire a phrase pair, the phrase appearance frequency information update step for increasing the F appearance frequency information corresponding to the phrase pair by a predetermined value;
When the symbol acquisition unit fails to acquire a phrase pair, a symbol acquisition step of acquiring one symbol using one or more symbol appearance frequency information;
The symbol appearance frequency information update unit increases the S appearance frequency information corresponding to the symbol acquired in the symbol acquisition step by a predetermined value;
When the partial phrase pair generation unit cannot acquire a phrase pair, a partial phrase pair generation step of generating two phrase pairs smaller than the phrase pair to be acquired;
The new phrase pair generation unit generates a new phrase pair according to the symbol acquired in the symbol acquisition step, or generates two smaller phrase pairs, and the one or more phrase appearance frequency information A new first language phrase in which two first language phrases composing the generated two phrase pairs are connected in order and a new first language phrase in which two second language phrases composing two phrase pairs are connected in order. Second process for generating one phrase pair having bilingual phrases, or two smaller phrase pairs are generated, and the generated two phrase pairs are configured using the one or more phrase appearance frequency information A new first language phrase that connects two first language phrases in sequence and two second language phrases that form two phrase pairs And new phrase pair generation step of performing one of a third process of generating a single phrase pair having a new second language phrases linked in reverse order's,
The control unit, for the phrase pair generated in the new phrase pair generation step, the phrase appearance frequency information update step, the symbol acquisition step, the symbol appearance frequency information update step, the partial phrase pair generation step, and the A control step that instructs to recursively perform the processing of the new phrase pair generation step;
A score calculating step of calculating a score for each phrase pair in the phrase table, using one or more phrase appearance frequency information stored in the phrase appearance frequency information storage unit;
A phrase table update step for storing the score calculated in the score calculation step in association with each phrase pair, and
The score calculating step includes:
When calculating a score for each phrase pair acquired from the j (2 <= j <= N) th parallel corpus, using one or more phrase appearance frequency information corresponding to the (j-1) th parallel corpus A parallel phrase learning method for calculating a score for each phrase pair corresponding to the j-th parallel corpus. Computer-accessible recording media
A bilingual information storage unit capable of storing N (N is a natural number of 2 or more) bilingual corpus having one or more bilingual information having a bilingual sentence and a bilingual tree structure;
A phrase pair that is a pair of a first language phrase having one or more words in the first language and a second language phrase having one or more words in the second language, and a score that is information relating to the appearance probability of the phrase pair; A phrase table that can store one or more scored phrase pairs having
A phrase appearance frequency information storage unit capable of storing, for each bilingual corpus, one or more phrase appearance frequency information including a phrase pair and F appearance frequency information that is information relating to the appearance frequency of the phrase pair;
A symbol appearance frequency information storage unit capable of storing at least one symbol appearance frequency information including a symbol for identifying a method for generating a new phrase pair and S appearance frequency information that is information relating to the appearance frequency of the symbol. ,
Computer
For each bilingual corpus, using the one or more phrase appearance frequency information, a generated phrase pair acquisition unit that acquires a phrase pair having a first language phrase and a second language phrase;
When the phrase pair can be acquired, the phrase appearance frequency information update unit that increases the F appearance frequency information corresponding to the phrase pair by a predetermined value;
If the phrase pair could not be acquired, a symbol acquisition unit that acquires one symbol using one or more symbol appearance frequency information,
A symbol appearance frequency information update unit that increases S appearance frequency information corresponding to the symbol acquired by the symbol acquisition unit by a predetermined value;
If the phrase pair could not be acquired, a partial phrase pair generation unit that generates two phrase pairs smaller than the phrase pair to be acquired;
In accordance with the symbol acquired by the symbol acquisition unit, a first process for generating a new phrase pair, or two smaller phrase pairs are generated, and the generated two phrases are generated using the one or more phrase appearance frequency information. One phrase having a new first language phrase in which two first language phrases constituting a phrase pair are sequentially connected and a new second language phrase in which two second language phrases constituting two phrase pairs are sequentially connected 2nd process which produces | generates a pair, or two smaller phrase pairs are produced | generated, The two 1st language phrases which comprise the produced | generated two phrase pairs are sequentially used using said one or more phrase appearance frequency information A new second language phrase that connects the new first language phrase connected together and the two second language phrases that make up two phrase pairs in reverse order. And new phrase pair generation unit that performs one of a third process of generating a single phrase pair having a's,
For the phrase pair generated by the new phrase pair generation unit, the phrase appearance frequency information update unit, the symbol acquisition unit, the symbol appearance frequency information update unit, the partial phrase pair generation unit, and the new phrase pair generation unit A control unit instructing to perform the process of recursively,
A score calculation unit that calculates a score for each phrase pair in the phrase table, using one or more phrase appearance frequency information stored in the phrase appearance frequency information storage unit;
Function as a phrase table update unit that stores the score calculated by the score calculation unit in association with each phrase pair;
The score calculation unit
When calculating a score for each phrase pair acquired from the j (2 <= j <= N) th parallel corpus, using one or more phrase appearance frequency information corresponding to the (j-1) th parallel corpus A program that causes a computer to function as a score for each phrase pair corresponding to the j-th parallel corpus.