JP5791097B2 - Parallel translation phrase learning apparatus, phrase-based statistical machine translation apparatus, parallel phrase learning method, and parallel phrase production method - Google Patents

Description

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

  In the first conventional method, after word alignment, heuristic phrase unit alignment is comprehensively performed to create a phrase table for phrase-based statistical machine translation (see, for example, Non-Patent Document 1). In this first method, lexical ambiguity is resolved with a long phrase, and sparse data is handled with a short phrase at the same time.

  In contrast to this heuristic method (first method), a method for directly calculating the alignment of phrase units from parallel translation data has been proposed (see, for example, Non-Patent Documents 2 to 5). In the methods described in Non-Patent Documents 2 and 3, phrases are comprehensively calculated, whereas in the methods described in Non-Patent Documents 4 and 5, the restrictions of Inversion Transduction Grammar (ITG) are used.

  In the methods described in Non-Patent Documents 3 to 5, a short phrase can be extracted by assigning a high probability to a simple phrase using a stochastic process based on a nonparametric Bayes method.

More specifically, ITG is a kind of synchronous context free grammar, and is characterized by rearranging words when generating nonterminal symbols. By using ITG restriction, the amount of calculation can be reduced, and the maximum likelihood solution and marginal probability of alignment can be calculated in polynomial time. In ITG, the generation probability of a certain phrase pair is P _flat (<e, f>; θ _x , θ _t ), and parameterized by the phrase pair probability θ _t and the symbol probability θ _x . The conventional ITG model uses the following generation process. <E, f> indicates a pair of a phrase e in the first language (for example, English) and a phrase f in the second language (for example, Japanese).

That is, in the conventional ITG model, first, a symbol (symbol) x is generated according to a polynomial distribution Px (x; θ _x ). Possible values for x are term, reg, or inv. Here, Term is a terminal symbol, reg is a normal non-terminal symbol, and inv is an inverted non-terminal symbol.

Second, according to the value of x, when “x = term (terminal symbol)”, a phrase pair is generated according to the phrase pair probability Pt (<e, f>; θ _t ). In addition, when “x = reg (ordinary non-terminal symbol)”, the phrase pair <e ₁ , f ₁ > and <e ₂ , f ₂ > are generated according to P _flat , and <e ₁ e ₂ , f ₁ f ₂ Use> to merge into one phrase pair. Furthermore, when “x = inv (inverted non-terminal symbol)”, two phrase pairs <e ₁ , f ₁ > and <e ₂ , f ₂ > are generated according to P _flat , but f ₁ and f ₂ are in reverse order To get <e ₁ e ₂ , f ₂ f ₁ >.

Then, the product of P _flat for each sentence is taken, and the corpus likelihood can be calculated as shown in Equation 1.

A conventional ITG model (referred to as FLAT) can be learned as it is by maximum likelihood estimation, but a very long phrase pair (one phrase per phrase) is obtained with the maximum likelihood solution. Therefore, the problem of a long phrase is solved by using a prior probability P (θ) = P (θ _x , θ _t ) that gives a high probability to a simple phrase dictionary (see Non-Patent Document 5).

Here, using Dirichlet distribution prior probability theta _x, the theta _t utilizing Pitman-Yor process based on non-parametric Bayesian method (see Non-Patent Document 6). The Pitman-Yor process is shown in Equation 2.

In Equation 2, d is a discount parameter of the Pitman-Yor process, and s is a strength parameter. Further, in Equation 2, the discount parameter d and the strength parameter s are estimated using the technique described in Non-Patent Document 6. In Equation 2, P _base is a base measure described later. Furthermore, since the Dirichlet distribution is a known technique in Equation 2, description thereof is omitted.

The advantage of using the prior distribution by the Pitman-Yor process lies in the nature of the stochastic process of storing the generated phrase pairs. The probability of phrase pairs that are frequently generated from the distribution is increased, and they are more likely to be generated (this effect is called “rich-gets-richer effect”). By using the Pitman-Yor process, a phrase table composed of fewer and more useful phrases can be constructed. Also, only phrases generated from P _t (phrase pair probability distribution) are stored. Also, the flat (ITG model), since only the minimum phrase pair terminal symbol is generated from P _t, only minimal phrase pairs also being stored.

Also, P _base in Equation 2 is the prior probability of the phrase pair in the model, and prior knowledge about the ease of phrase alignment can be incorporated into the model by appropriately determining it. Here, P _base is of no corresponding phrase (| e | = 0 or | f | = 0) to choose whether or not to generate a with a certain probability P _u, to produce a phrase without the support from P _bu, corresponding Yes Generate a phrase pair from Pba.
P _ba can be calculated by the following Equation 3 as described in Non-Patent Document 7.

In Equation 3, P _pois is a Poisson distribution having an average length parameter λ. To avoid long phrases, use a small value for λ. P _m1 is an IBM model 1 probability based on the word probability (see Non-Patent Document 8). By using this, if the translation probability of the words constituting the phrase is high, the probability of the phrase is also high. By using the geometric mean of conditional probabilities in both directions, phrases that match both models are preferentially aligned. Further, in Equation 3, when words constituting the phrase e is _{e 1 ··· e_ n, Puni (} e) is the product of the unigram probability of each word.

In _Pbu , a non-empty word string from e and f is defined as g, and the probability is defined as in the following Equation 4.

In Equation 4, _Pbu is divided by 2 in order to consider both e and f. Further, in Equation 4, when words constituting the phrase g is _{g 1 ··· g_ n, Puni (} g) is the product of the unigram probability of each word.

P. Koehn, F. J. Och, and D. Marcu. Statistical phrase-based translation. In Proc. NAACL, pp. 48-54, 2003. Daniel Marcu and William Wong. 2002. A phrase-based, joint probability model for statistical machine transla- tion. Pages 133-139. J. DeNero and D. Klein. The complexity of phrase alignment problems.In Proc.ACL, pp. 25-28, 2008. P. Blunsom and T. Cohn. Inducing synchronous gram- mars with slice sampling. In Proc. NAACL, 2010. H. Zhang, C. Quirk, R. C. Moore, and D. Gildea. Bayesian learning of non-compositional phrases with synchronous parsing. Proc. ACL, pp. 97-105, 2008. Y. W. Teh.A hierarchical Bayesian language model based on Pitman-Yor processes.In Proc.ACL, 2006. J. DeNero, A. Bouchard-C ^ ot_e, and D. Klein. Sam-pling alignment structure under a Bayesian translation model. In Proc. EMNLP, pp. 314-323, 2008. J. DeNero and D. Klein. Discriminative modeling of extraction sets for machine translation. In Proc. ACL, pp. 1453-1463, 2010.

  However, in the first conventional method, the optimum phrase for translation is not always obtained, and it is necessary to extract a large number of phrases in order to improve translation accuracy.

  In the maximum likelihood estimation in the method described in Non-Patent Document 2, only very long phrase pairs can be extracted.

  Further, in the methods described in Non-Patent Documents 3 to 5, as described above, only the extraction of the phrase of the minimum unit is stored, and the phrases having a plurality of granularities cannot be directly modeled. For this reason, after calculating the phrase alignment of the minimum unit, the phrases are comprehensively extracted by heuristics, and eventually, a two-stage method has to be used. Moreover, in order to exhaustively extract phrases, inappropriate phrase pairs have been learned.

  Specifically, when the conventional ITG model is used, for example, for the phrase pair “Mrs. Smith's red cookbook”, as shown in FIG. 11, the phrase “Mrs. / San "," Smith "," 's / no "," red / red "," cookbook / cookbook "can only be acquired.

  The bilingual phrase learning device of the first invention includes 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. One or more having a phrase table that can store one or more scored phrase pairs having a score that is information relating to the appearance probability of the phrase pair, a phrase pair, and F appearance frequency information that is information relating to the appearance frequency of the phrase pair One or more symbol appearances having a phrase appearance frequency information storage unit that can store the phrase appearance frequency information, a symbol that identifies a method for generating a new phrase pair, and S appearance frequency information that is information about the appearance frequency of the symbol A first language phrase and a second language frame using a symbol appearance frequency information storage unit capable of storing frequency information and one or more phrase appearance frequency information A generated phrase pair acquisition unit that acquires a phrase pair including: a phrase appearance frequency information update unit that increases F appearance frequency information corresponding to the phrase pair by a predetermined value if the phrase pair can be acquired; If the phrase pair cannot be acquired, the symbol acquisition unit for acquiring one symbol and the S appearance frequency information corresponding to the symbol acquired by the symbol acquisition unit using one or more symbol appearance frequency information are determined in advance. The symbol appearance frequency information update unit that increases by the value obtained, the partial phrase pair generation unit that generates two phrase pairs smaller than the phrase pair to be acquired, and the symbol acquisition unit when the phrase pair could not be acquired The first process to generate a new phrase pair according to the symbol, or two smaller phrase pairs to generate one or more phrase occurrences Using information, a new first language phrase that connects two first language phrases that make up the two generated phrase pairs in order, and a new second language phrase that connects the two second language phrases that make up the two phrase pairs in order. The second process of generating one phrase pair having a bilingual phrase, or two smaller phrase pairs are generated, and the generated two phrase pairs are configured using one or more phrase appearance frequency information Generates a single phrase pair having a new first language phrase in which two first language phrases are connected in order and a new second language phrase in which two second language phrases constituting two phrase pairs are connected in reverse order Phrase appearance frequency for the phrase pair generated by the new phrase pair generation unit that performs any of the third processes and the phrase pair generation unit The information update unit, the symbol acquisition unit, the symbol appearance frequency information update unit, the partial phrase pair generation unit, and the control unit that recursively processes the new phrase pair generation unit, and the phrase appearance frequency information storage unit 1 Using the above phrase appearance frequency information, a score calculation unit that calculates a score for each phrase pair in the phrase table, and a phrase table update unit that stores the score calculated by the score calculation unit in association with each phrase pair This is a bilingual phrase learning device.

  With this configuration, a large number of appropriate phrase pairs can be learned.

  Moreover, the bilingual phrase learning device of the second invention has a first language phrase and a second language phrase using the probability distribution of the phrase pair, with respect to the first invention. When the generated phrase pair is acquired and the symbol acquisition unit cannot acquire the phrase pair, the symbol probability distribution is used to acquire one symbol, and the partial phrase pair generation unit cannot acquire the phrase pair. If the base measure is used, two phrase pairs smaller than the phrase pair to be generated are generated, and the first process is a process for generating a new phrase pair using the phrase pair base measure, and the score The calculation unit uses the one or more phrase appearance frequency information stored in the phrase appearance frequency information storage unit, and based on the nonparametric Bayes method, the phrase table A bilingual phrase learning device for calculating a score for each phrase pairs.

  With this configuration, a large number of appropriate phrase pairs can be learned.

  In addition, the phrase-based statistical machine translation device according to the third aspect of the invention provides a phrase table learned by the parallel phrase learning device and a sentence in a first language having one or more words with respect to the first or second invention. A phrase acquisition unit that extracts one or more phrases from a sentence received by the reception unit, and acquires one or more phrases in the second language from the phrase table using a score of the phrase table, and a phrase acquisition It is a phrase-based statistical machine translation device comprising a sentence constructing unit that constructs a sentence in the second language from one or more phrases acquired by the unit, and an output unit that outputs a sentence constructed by the sentence constructing unit.

  With this configuration, accurate machine translation can be performed using a large number of appropriate phrase pairs.

  According to the bilingual phrase learning apparatus according to the present invention, a large number of appropriate phrase pairs can be learned.

Block diagram of bilingual phrase learning apparatus according to Embodiment 1 The flowchart explaining operation | movement of the parallel translation phrase learning apparatus. A flowchart for explaining the operation of the phrase generation process Diagram explaining phrase pairs that can be learned Diagram showing the specifications of the corpus Figure showing the results of the experiment Diagram showing comparison between phrase extraction based on model probability and conventional method Block diagram of phrase-based statistical machine translation apparatus in Embodiment 2 Overview of the computer system in the above embodiment Block diagram of a computer system in the above embodiment The figure explaining the phrase pair which can be learned in the prior art

  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 the present embodiment, a parallel phrase learning apparatus that accumulates a plurality of levels of parallel phrases will be described.

  In the present embodiment, a phrase having a plurality of granularities is directly expressed by a probability model by using a hierarchical ITG model. For this reason, in the parallel phrase learning apparatus in the present embodiment, high translation accuracy can be realized without performing phrase extraction based on heuristics.

  FIG. 1 is a block diagram of a bilingual phrase learning apparatus 1 according to the present embodiment. The parallel phrase learning device 1 includes a parallel corpus 100, a phrase table 101, a phrase appearance frequency information storage unit 102, a symbol appearance frequency information storage unit 103, a phrase table initialization unit 113, a generated phrase pair acquisition unit 104, and a phrase appearance frequency information update. Unit 105, symbol acquisition unit 106, symbol appearance frequency information update unit 107, partial phrase pair generation unit 108, new phrase pair generation unit 109, control unit 110, score calculation unit 111, parsing unit 114, phrase table update unit 112, and A tree update unit 115 is provided.

  The bilingual corpus 100 can store one or more bilingual information having a bilingual sentence and a tree structure of the bilingual sentence. A 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 tree structure of the bilingual sentence is, for example, information as shown in FIG.

The phrase table 101 can store one or more scored phrase pairs. 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.

  The phrase appearance frequency information storage unit 102 can store one or more phrase appearance frequency information. 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.

  The phrase table initialization unit 113 generates initial information of one or more phrase pairs with scores from one or more pieces of parallel translation information of the bilingual corpus 100 and stores the initial information in the phrase table 101. Note that the phrase table initialization unit 113 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 the phrase pair in the phrase table 101. In such a case, the score is the number of appearances.

The generated phrase pair acquisition unit 104 acquires one or more bilingual sentences stored in the bilingual corpus 100, and appearances of one or more phrase pairs that constitute the tree structure of the bilingual sentences (usually, appearance frequency). (1)) is subtracted from the phrase pair score existing in the phrase table 101. Next, the generated phrase pair acquisition unit 104 acquires a phrase pair that includes the first language phrase and the second language phrase by using one or more phrase appearance frequency information (accurately, 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 suitable for the generated phrase pair acquisition unit 104 to acquire a phrase pair having a first language phrase and a second language phrase using the phrase pair probability distribution P _t .

  When the generated phrase pair acquisition unit 104 or the like can acquire a phrase pair, the phrase appearance frequency information update unit 105 increases the F appearance frequency information corresponding to the phrase pair by a predetermined value. The F appearance frequency information here is the appearance frequency of the passage and phrase pair. The predetermined value is usually 1. The generated phrase pair acquisition unit 104 and the like are the generated phrase pair acquisition unit 104 and the new phrase pair generation unit 109.

When the generated phrase pair acquisition unit 104 or the like cannot acquire a phrase pair, the symbol acquisition unit 106 acquires one symbol using one or more symbol appearance frequency information. Here, it is preferable to use the probability distribution P _x (x; θ _x ) of the symbols to use one or more symbol appearance frequency information. That is, it is preferable that the symbol acquisition unit 106 acquires one symbol using the probability distribution of symbols when the generation phrase pair acquisition unit 104 cannot acquire the generation phrase pair. One symbol is, for example, any one of BASE, REG, and INV.

  The symbol appearance frequency information update unit 107 increases the S appearance frequency information corresponding to the symbol acquired by the symbol acquisition unit 106 by a predetermined value. The predetermined value is usually 1.

When the generated phrase pair acquisition unit 104 or the like cannot acquire a phrase pair, the partial phrase pair generation unit 108 generates two phrase pairs smaller than the phrase pair to be acquired. In addition, when the phrase pair cannot be acquired, the partial phrase pair generation unit 108 normally generates two phrase pairs smaller than the phrase pair to be acquired using the prior probability of the phrase pair. More specifically, for example, the partial phrase pair generation unit 108 generates two phrase pairs smaller than the phrase pair to be acquired, using the base measure P _dac acquired from the prior probability of the phrase pair. The _dac of the base measure P _dac is an abbreviation of (“divide-and-conquer”), and is a probability based on a mechanism for dividing a long phrase pair into short phrase pairs. For example, if the phrase pair you are trying to acquire is <red cookbook, red cookbook>, then “P _dac (<red cookbook, red cookbook>) = P _x (REG) * P _t (<red, red>) * P_ {t} (<cookbook, book>) + P _x (REG) * P _t (<red, red dish>) * P _t (<cookbook, book>) + P _x (INV) * P _t (< red, book>) * P _t (<cookbook, red food>) + P _x (INV) * P _t (<red, food book>) * P _t (<cookbook, red>) + P _x (BASE) * P _base (<red cookbook>) ”.

  The new phrase pair generation unit 109 performs either the first process, the second process, or the third process according to the symbol acquired by the symbol acquisition unit 106. The new phrase pair generation unit 109 performs the first process when the symbol acquired by the symbol acquisition unit 106 is BASE, performs the second process when the symbol is REG, and performs the second process when the symbol is INV. Perform the third process.

  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. Moreover, since the first process is a known technique, the description is omitted.

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 ).

For the phrase pair generated by the new phrase pair generation unit 109, the control unit 110 includes a phrase appearance frequency information update unit 105, a symbol acquisition unit 106, a symbol appearance frequency information update unit 107, a partial phrase pair generation unit 108, and a new phrase. The processing of the pair generation unit 109 is performed 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 the phrase pairs from P_ _base.

The score calculation unit 111 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. Here, using one or more phrase appearance frequency information means, for example, using a Pitman-Yor process based on the nonparametric Bayes method as shown in Equation 5. That is, the score calculation unit 111 calculates a score for each phrase pair in the phrase table 101 based on the nonparametric Bayes method using one or more phrase appearance frequency information stored in the phrase appearance frequency information storage unit 102. It is preferable to do.

Equation 5 means that the probability distribution of the phrase pair is generated from a Pitman-Yor process using d, s, and P _dac as parameters.

  The parsing unit 114 acquires a tree structure of bilingual sentences (including phrases) that maximizes the score calculated by the score calculation unit 111. In more detail, the purging unit 114 acquires a tree structure using 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 tree update unit 115 accumulates the tree structure acquired by the parsing unit 114 in the bilingual corpus 100. Here, the tree update unit 115 normally overwrites the tree structure. That is, the old tree structure in the bilingual corpus 100 is updated to a new tree structure.

  The phrase table update unit 112 stores the score calculated by the score calculation unit 111 in association with each phrase pair. Moreover, the phrase table update part 112, when the phrase pair corresponding to the score which the score calculation part 111 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 111 calculated The phrase table 101 may be accumulated.

  The phrase table 101, the phrase appearance frequency information storage unit 102, the symbol appearance frequency information storage unit 103, or the bilingual corpus 100 is preferably a non-volatile recording medium, but can also be realized by a volatile recording medium.

  The process of storing a phrase pair with a score in the phrase table 101 or the like is not limited. For example, a phrase pair with a score may be stored in the phrase table 101 or the like via a recording medium, and a phrase pair with a score transmitted via a communication line or the like is stored in the phrase table 101 or the like. Alternatively, a phrase pair with a score or the like input via an input device may be stored in the phrase table 101 or the like.

  Generated phrase pair acquisition unit 104, phrase appearance frequency information update unit 105, symbol acquisition unit 106, symbol appearance frequency information update unit 107, partial phrase pair generation unit 108, new phrase pair generation unit 109, control unit 110, score calculation unit 111 The phrase table update unit 112, the phrase table initialization unit 113, the parsing unit 114, and the tree update unit 115 can be usually realized by an MPU, a memory, or the like. The processing procedure of the generated phrase pair acquisition unit 104 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. It is assumed that the phrase table initialization unit 113 generates the initial stage phrase table 101 before the operation of the flowchart of FIG. In addition, the flowcharts of FIGS. 2 and 3 describe a process of acquiring a scored phrase pair using one translation information in the translation corpus 100. That is, normally, a phrase pair with a score is repeatedly obtained for a large number of pieces of parallel translation information in the parallel translation corpus 100. In addition, it is preferable to repeatedly obtain a scored phrase pair even for one parallel translation information.

(Step S201) The generated phrase pair acquisition unit 104 acquires one or more bilingual sentences stored in the bilingual corpus 100, and appearances of one or more phrase pairs constituting 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 104 attempts to generate one phrase pair using the phrase pair probability distribution P _t . Further, the phrase pair probability distribution P _t can be calculated by, for example, a Pitman-Yor process using the phrase pair frequency (F appearance frequency information) in the phrase appearance frequency information storage unit 102. Note that the calculation of the probability based on the Pitman-Yor process is a known technique, and thus the description thereof is omitted.

  (Step S202) The partial phrase pair generation unit 108 or the like performs a phrase generation process and ends the process. The phrase generation process is a phrase generation process using two or more hierarchies using hierarchical ITG. The phrase generation process will be described with reference to the flowchart of FIG.

  Next, the phrase generation processing in step S102 will be described using the flowchart in FIG.

  (Step S301) The partial phrase pair generation unit 108 determines whether or not a phrase pair has been generated in the previous phrase pair generation process. 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 105 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 105 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 111 calculates the score of the phrase pair corresponding to the updated phrase appearance frequency information.

  (Step S304) The phrase table update unit 112 configures a scored phrase pair having the score calculated in step S303, and writes it into the phrase table 101. When the phrase pair does not exist in the phrase table 101, the phrase table update unit 112 configures a score-added phrase pair and newly adds the phrase pair to the phrase table 101. When the phrase pair exists in the phrase table 101, the phrase table update unit 112 updates the score corresponding to the phrase pair to the score calculated in step S303. Then, the process returns to the upper process (step S202 and the like).

(Step S305) The partial phrase pair generation unit 108 generates two phrase pairs smaller than the phrase pair to be generated using the base measure P _dac .

  (Step S306) The symbol acquisition unit 106 acquires one symbol x using one or more symbol appearance frequency information.

  (Step S307) The symbol appearance frequency information update unit 107 increases the S appearance frequency information corresponding to the symbol x acquired by the symbol acquisition unit 106 by a predetermined value (usually “1”).

  (Step S308) The new phrase pair generation unit 109 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 109 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 109 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 109 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 109 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 109 generates two smaller phrase pairs. In addition, let two phrase pairs here be the 1st phrase pair and the 2nd phrase pair.

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

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

  (Step S318) The new phrase pair generation unit 109 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, after step S304 and before returning, generation of a tree structure by the parsing unit 114 and update processing of the tree structure (the tree structure in the bilingual corpus 100) by the tree update unit 115 Is preferably performed.

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

Now, the phrase table 101, and a number of the scored phrase pair stored, the probability distribution of the phrase pair is P _t, and to.

  The phrase appearance frequency information storage unit 102 stores one or more phrase appearance frequency information that is a pair of a phrase pair and an appearance frequency.

  Further, the symbol appearance frequency information storage unit 103 stores three symbol appearance frequency information, which are combinations of the symbols “BASE”, “REG”, “INV”, and the appearance frequency of each symbol.

In such a situation, first, the generated phrase pair acquisition unit 104 of the parallel phrase learning device 1 acquires one parallel sentence from the parallel corpus 100. Next, the generated phrase pair acquisition unit 104 uses the occurrence of one or more phrase pairs constituting the acquired bilingual tree structure (usually “1” of appearance frequency) as a phrase pair existing in the phrase table 101. Subtract from the score. Next, the generated phrase pair acquisition unit 104 tries to generate the phrase pair <e, f>, which is the parallel translation, from the phrase pair probability distribution P _t .

  If the partial phrase pair generation unit 108 determines that the phrase pair could not be generated in the previous phrase pair generation process, the partial phrase pair generation unit 108 performs the following process.

That is, the partial phrase pair generation unit 108 recursively generates two phrase pairs smaller than the phrase pair to be generated, using the base measure P _dac . And a new phrase pair is produced | generated by combining two smaller phrase pairs produced | generated. The base measure P _dac is defined, and the equation of θ _t is as shown in Equation 5.

  Moreover, the generation process of Pdac is a generation process based on ITG as follows.

That is, the symbol acquisition unit 106 generates a symbol according to the probability distribution P _x (x; θ _x ) of the symbol using the three symbol appearance frequency information. Then, the symbol appearance frequency information update unit 107 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 109 generates a new phrase pair directly from P _base . When the generated symbol x is “x = reg”, the new phrase pair generation unit 109 generates <e ₁ , f ₁ > and <e ₂ , f ₂ > from P _hier and generates one phrase pair < Create e ₁ e ₂ , f ₁ f ₂ >. When the generated symbol x is “x = inv”, the new phrase pair generation unit 109 generates <e ₁ , f ₁ > and <e ₂ , f ₂ > from P _hier , and f ₁ and f ₂ Are arranged in reverse order to create one phrase pair <e ₁ e ₂ , f ₂ f ₁ >.

  Then, the phrase appearance frequency information update unit 105 updates the phrase appearance frequency information of the newly created phrase pair.

  Moreover, the score calculation part 111 calculates the score of the phrase pair corresponding to the updated phrase appearance frequency information.

  Then, the phrase table update unit 112 updates the phrase table.

  In addition, the purging unit 114 acquires a new tree structure that maximizes the score of the tree structure, using the score calculated by the score calculation unit 111. Then, the tree updating unit 115 accumulates the acquired tree structure in the bilingual corpus 100 and updates the old tree structure to the new tree structure.

  With the above processing, as shown in FIG. 4, phrase pairs having a multi-level granularity can be learned for the phrase pair “Mrs. Smith's red cookbook”.

  In addition, 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 Pt (f | e) and Pt (e | f), lexical weight-ing probabilities, phrase penalties, etc. are used. Here, the conditional probability is calculated using the model probability P _t . That is, the conditional probability is calculated using Equation 6 and Equation 7. Then, for example, the score calculation unit 111 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. Such calculation is known technology (see P. Koehn, FJ Och, and D. Marcu. Statistical phrase-based translation. In Proc. NAACL, pp. 48-54, 2003.)
It is. The phrase penalty is “1” for all phrases, for example.

Note that the phrase table update unit 112 puts only phrase pairs that appear at least once in the sample into the phrase table 101. In addition, two features are added. The first is the phrase pair coincidence probability Pt (<e, f>) by the model. The second feature is the average posterior probability of a span containing a phrase pair <e, f> 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.
(Experiment)

  Hereinafter, the experimental result of the parallel phrase learning device 1 will be described. In this experiment, the method of the bilingual phrase learning device 1 was evaluated by the tasks of French-English translation and Japanese-English translation.

  Workshop on Statistical Machine Translation (WMT) (C. Callison-Burch, et al. Findings of the 2010 joint workshop on statistical machine translation and metrics for machine translation. In Proc. WMT / MetricsMATR, pp. 17 {53 , 2010.), we used a news commentary corpus for translation model learning, and a news commentary and Europarl corpus for language model learning. Japanese to English translation is NTCIR's patent translation task (A.Fujii, M.Utiyama, M.Yamamoto, and T.Utsuro.Overview of the patent translation task at the NTCIR-7 workshop. In Proc. NTCIR-7, pp. 389 -400, 2008.), the first 100,000 sentences of the parallel corpus were used as the translation model, and the entire parallel corpus was used as the language model. The corpus specifications are shown in FIG. As preprocessing of data, word division (tokenization) and lower case are performed, and only sentences of 40 words or less are used for learning a translation model. Moses (see P. Koehn, et al. Moses: Open source toolkit for statistical machine translation. In Proc. ACL, 2007.) is used as a decoder. The maximum phrase length is 7, and the language model is a 5-gram model using Kneser-Ney smoothing. The evaluation standard is BLEU score up to 4-gram. In the first experiment, the phrase extraction (mod) using the flat and hier model probabilities is compared with the accuracy of phrase extraction based on alignment (giza) and heuristics obtained from GIZA ++. In the case of giza, the standard learning setting up to Model4 is used and combined with the alignment results in both directions with grow-diag-final-and. In the method of the parallel phrase learning device 1, 100 iterations are learned and the last sample is used. Actually, the likelihood increased monotonically up to the 100th iteration, but the translation accuracy became almost the same after the 5th to 10th iteration. Since one iteration took about 1.3 hours for one core, good translation accuracy could be realized in 6.5 to 13 hours.

  The experimental results are shown in FIG. FIG. 6 shows the BLEU score and the phrase table size. In FIG. 6, the phrase table using the probability of the hierarchical model in both French and English and Japanese and English slightly exceeded the accuracy of phrase extraction based on GIZA ++ and heuristics. In other words, the complete probabilistic model exceeded the phrase extraction based on heuristics. Furthermore, the size of the phrase table obtained by the method of the bilingual phrase learning apparatus 1 is also less than 20% of the conventional method. Also, when using model probabilities, hier (this method) greatly exceeded flat (conventional method based on ITG model). This is because high accuracy cannot be obtained if only the minimum phrase is used.

  Furthermore, FIG. 7 shows a comparison between phrase extraction based on model probabilities and the conventional method. FIG. 7 shows translation accuracy and phrase table size by various phrase extraction methods. In FIG. 7, in addition to the proposed method mod using model probabilities using hier and flat alignment, heuristic extraction using the phrase heur-p, block heur-b, and word heur-w as the minimum unit was compared. It can be seen from FIG. 7 that the combination of hier and mod significantly reduces the size of the phrase table while showing almost the same or higher accuracy than the heuristic extraction.

  As mentioned above, 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 bilingual 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.

(Embodiment 2)
In the present embodiment, a phrase-based statistical machine translation device using the phrase table 101 learned by the parallel phrase learning device 1 in the first embodiment will be described.

  FIG. 8 is a block diagram of the phrase-based statistical machine translation apparatus 2 in the present embodiment.

  The phrase-based statistical machine translation device 2 includes a phrase table 101, a reception unit 201, a phrase acquisition unit 202, a sentence composition unit 203, and an output unit 204.

  The phrase table 101 is a phrase table learned by the parallel phrase learning device 1.

  The accepting unit 201 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 201 can be realized by a device driver of an input unit such as a keyboard, menu screen control software, or the like.

  The phrase acquisition unit 202 extracts one or more phrases from the sentence received by the reception unit 201, 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 202 is a known technique.

  The sentence composing unit 203 composes a sentence in the second language from one or more phrases acquired by the phrase acquiring unit 202. The processing of the sentence composition unit 203 is a known technique.

  The output unit 204 outputs the sentence constructed by the sentence constructing unit 203. 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 202 and the sentence composition unit 203 can be usually realized by an MPU, a memory, or the like. The processing procedure of the phrase acquisition unit 202 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).

  The output unit 204 may be considered as including or not including an output device such as a display or a speaker. The output unit 204 can be implemented by output device driver software, or output device driver software and an output device.

  Further, the operation of the phrase-based statistical machine translation apparatus 2 suffices to perform a well-known phrase-based statistical machine translation process, and thus detailed description thereof is omitted.

  As described above, according to this embodiment, it is possible to perform highly accurate machine translation using a phrase table that can be realized with a small storage area.

  FIG. 9 shows the external appearance of a computer that implements the bilingual phrase learning apparatus and the like of the above-described embodiment. The above-described embodiments can be realized by computer hardware and a computer program executed thereon. FIG. 9 is an overview diagram of the computer system 340, and FIG. 10 is a block diagram of the computer system 340.

  In FIG. 9, the computer system 340 includes a computer 341 including an FD drive and a CD-ROM drive, a keyboard 342, a mouse 343, and a monitor 344.

  In FIG. 10, in addition to the FD drive 3411 and the CD-ROM drive 3412, the computer 341 stores an MPU 3413, a bus 3414 connected to the CD-ROM drive 3412 and the FD drive 3411, and a program such as a bootup program. A RAM 3416 for temporarily storing application program instructions and providing a temporary storage space; and a hard disk 3417 for storing application programs, system programs, and data. Although not shown here, the computer 341 may further include a network card that provides connection to the LAN.

  A program for causing the computer system 340 to execute the functions of the bilingual phrase learning device or the like of the above-described embodiment is stored in the CD-ROM 3501 or FD 3502, inserted into the CD-ROM drive 3412 or FD drive 3411, and further a hard disk. 3417 may be transferred. Alternatively, the program may be transmitted to the computer 341 via a network (not shown) and stored in the hard disk 3417. The program is loaded into the RAM 3416 at the time of execution. The program may be loaded directly from the CD-ROM 3501, the FD 3502, or the network.

  The program does not necessarily include an operating system (OS) or a third-party program that causes the computer 341 to execute the functions of the bilingual phrase learning apparatus according to 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 340 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, 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 parallel phrase learning device according to the present invention has an effect that a large number of appropriate phrase pairs can be learned, and is useful as a parallel phrase learning device, a machine translation device, and the like.

DESCRIPTION OF SYMBOLS 1 Parallel phrase learning apparatus 2 Phrase-based statistical machine translation apparatus 101 Phrase table 102 Phrase appearance frequency information storage part 103 Symbol appearance frequency information storage part 104 Generation | occurrence | production phrase pair acquisition part 105 Phrase appearance frequency information update part 106 Symbol acquisition part 107 Symbol appearance Frequency information update unit 108 Partial phrase pair generation unit 109 New phrase pair generation unit 110 Control unit 111 Score calculation unit 112 Phrase table update unit 201 Reception unit 202 Phrase acquisition unit 203 Sentence configuration unit 204 Output unit

Claims (6)

In a first language phrase having one or more words in the first language, pairs der the second language phrase having one or more words in the second language is, pairs of the first language sentence and the second language sentence One or more scores having a phrase pair acquired from a bilingual corpus storing one or more parallel translation information having a certain bilingual sentence and a tree structure of the bilingual sentence, and a score that is information relating to the appearance probability of the phrase pair A phrase table that can store phrase pairs with
A phrase appearance frequency information storage unit capable of storing one or more phrase appearance frequency information having a phrase pair and F appearance frequency information that is information related 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;
F appearance frequency information included in the one or more phrase appearance frequency information from one or more phrase pairs constituting a tree structure of one or more bilingual sentences stored in the bilingual corpus, A generated phrase pair acquisition unit that attempts to acquire a phrase pair using the probability distribution of each of the one or more phrase pairs acquired using F appearance frequency information corresponding to each phrase pair ;
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;
A symbol acquisition unit that acquires one symbol using the probability distribution of the one or more symbols acquired using the S appearance frequency information included in the one or more symbol appearance frequency information when the phrase pair cannot be acquired; ,
A symbol appearance frequency information update unit that is S appearance frequency information corresponding to the symbol acquired by the symbol acquisition unit and increases the S appearance frequency information stored in the symbol appearance frequency information storage unit by a predetermined value. When,
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;
A process corresponding to a symbol acquired by the symbol acquisition unit, wherein a new phrase pair is generated using a prior probability of a phrase pair acquired using F appearance frequency information corresponding to one or more phrase pairs . A new process in which two first language phrases constituting the generated two phrase pairs are sequentially connected using one process or two smaller phrase pairs and using the one or more phrase appearance frequency information A second process of generating one phrase pair having a first language phrase and a new second language phrase sequentially concatenating the two second language phrases that make up the two phrase pairs, or two smaller It generates a phrase pair, using the one or more phrases appearance frequency information, sequentially connecting the two first language phrase that constitutes the two phrases pairs generated A new first language phrase, and either the third process of generating a single phrase pair having a new second language phrase linked in reverse order with two second language phrase that constitutes the two phrases pairs A new phrase pair generator that performs
For each of the two phrase pairs acquired by the process of the second process by the new phrase pair generation unit, and for each of the two phrase pairs acquired by the process of the third process by the new phrase pair generation unit , A control unit that recursively processes 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;
Using one or more phrase appearance frequency information stored in the phrase appearance frequency information storage unit, the phrase pair acquired by the generated phrase pair acquisition unit , the new phrase pair generation unit acquired by the first process The score for the new phrase pair, the phrase pair acquired by the new phrase pair generation unit by the second process, and the phrase pair acquired by the new phrase pair generation unit by the third process are calculated. A score calculator to
A bilingual phrase learning device comprising: a phrase table updating unit that associates each score calculated by the score calculation unit with each phrase pair and accumulates the score in the phrase table. The generated phrase pair acquisition unit
F appearance frequency information included in the one or more phrase appearance frequency information from one or more phrase pairs constituting a tree structure of one or more bilingual sentences stored in the bilingual corpus, Using the probability distribution of phrase pairs acquired using F appearance frequency information corresponding to each phrase pair, a generated phrase pair having a first language phrase and a second language phrase is acquired,
The symbol acquisition unit
If the phrase pair could not be obtained, the symbol probability distribution is used to obtain one symbol,
The partial phrase pair generation unit
If the phrase pair could not be obtained, the base measure is used to generate two phrase pairs that are smaller than the phrase pair that was to be generated,
The first process is a process for generating a new phrase pair using a phrase pair basis measure,
The score calculation unit
The parallel phrase learning according to claim 1, wherein a score for each phrase pair in the phrase table is calculated based on a nonparametric Bayes method using one or more phrase appearance frequency information stored in the phrase appearance frequency information storage unit. apparatus. A phrase table learned by the parallel phrase learning device according to claim 1;
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 phrase-based statistical machine translation device comprising: an output unit that outputs a sentence formed by the sentence composing unit. On the storage medium,
In a first language phrase having one or more words in the first language, pairs der the second language phrase having one or more words in the second language is, pairs of the first language sentence and the second language sentence One or more scores having a phrase pair acquired from a bilingual corpus storing one or more parallel translation information having a certain bilingual sentence and a tree structure of the bilingual sentence, and a score that is information relating to the appearance probability of the phrase pair A phrase table with a phrase pair,
One or more phrase appearance frequency information having a phrase pair and F appearance frequency information that is information related to the appearance frequency of the phrase pair;
Storing 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 relating to the appearance frequency of the symbol,
Realized by 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, new phrase pair generation unit, control unit, score calculation unit, and phrase table update unit A parallel phrase learning method to be performed,
The generated phrase pair acquisition unit includes an F appearance frequency included in the one or more phrase appearance frequency information from one or more phrase pairs constituting the tree structure of one or more parallel sentences stored in the bilingual corpus. A generated phrase pair acquisition step for acquiring a phrase pair using the F appearance frequency information corresponding to each of the one or more phrase pairs ;
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;
If the symbol acquisition unit fails to acquire a phrase pair, a symbol acquisition step of acquiring one symbol using S appearance frequency information included in one or more symbol appearance frequency information;
The symbol appearance frequency information update unit is S appearance frequency information corresponding to the symbol acquired in the symbol acquisition step, and the S appearance frequency information stored in the symbol appearance frequency information storage unit is predetermined. A symbol appearance frequency information update step that increases by a 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 is a process corresponding to the symbol acquired in the symbol acquisition step, generates a new phrase pair, or generates two smaller phrase pairs, the one or more Using the phrase appearance frequency information, the first first language phrase formed by sequentially connecting the two first language phrases constituting the generated two phrase pairs and the two second language phrases constituting the two phrase pairs A second process for generating one phrase pair having a new second language phrase connected in order, or two smaller phrase pairs, and generated using the one or more phrase appearance frequency information A new first language phrase that connects two first language phrases that compose one phrase pair in sequence, and two that compose 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 bilingual phrase,
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 A control step for recursively performing the processing of the new phrase pair generation step;
The score calculation unit calculates a score for the phrase pair acquired by the generated phrase pair acquisition unit using one or more phrase appearance frequency information stored in the storage medium; and
The phrase table update unit, comprising: a phrase table update step in which the score calculated in the score calculation step is associated with each phrase pair and accumulated in the phrase table. On the storage medium,
In a first language phrase having one or more words in the first language, pairs der the second language phrase having one or more words in the second language is, pairs of the first language sentence and the second language sentence One or more scores having a phrase pair acquired from a bilingual corpus storing one or more parallel translation information having a certain bilingual sentence and a tree structure of the bilingual sentence, and a score that is information relating to the appearance probability of the phrase pair A phrase table with a phrase pair,
One or more phrase appearance frequency information having a phrase pair and F appearance frequency information that is information related to the appearance frequency of the phrase pair;
Storing 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 relating to the appearance frequency of the symbol,
Realized by 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, new phrase pair generation unit, control unit, score calculation unit, and phrase table update unit A method for producing a translated phrase,
The generated phrase pair acquisition unit includes an F appearance frequency included in the one or more phrase appearance frequency information from one or more phrase pairs constituting the tree structure of one or more parallel sentences stored in the bilingual corpus. A generated phrase pair acquisition step for acquiring a phrase pair using the probability distribution of the one or more phrase pairs acquired using the F appearance frequency information corresponding to each of the one or more phrase pairs. When,
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 cannot acquire a phrase pair, the probability distribution of the one or more symbols acquired using the S appearance frequency information included in the one or more symbol appearance frequency information is used to determine one symbol. A symbol acquisition step to acquire;
The symbol appearance frequency information update unit is S appearance frequency information corresponding to the symbol acquired in the symbol acquisition step, and the S appearance frequency information stored in the symbol appearance frequency information storage unit is predetermined. A symbol appearance frequency information update step that increases by a 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 said new phrase pair production | generation part is a process corresponding to the symbol acquired at the said symbol acquisition step, and uses the prior probability of the phrase pair acquired using F appearance frequency information corresponding to each 1 or more phrase pair Te, first process of generating a new phrase pair, or to generate two smaller phrase pair, using the one or more phrases appearance frequency information, the two constituting two phrases pairs of said generated first A second that generates a single phrase pair having a new first language phrase in which one language phrase is sequentially connected and a new second language phrase in which two second language phrases constituting two phrase pairs are sequentially connected Or two smaller phrase pairs are generated, and the two generated phrase pairs are constructed using the one or more phrase appearance frequency information. One phrase having a new first language phrase in which two first language phrases are sequentially connected and a new second language phrase in which two second language phrases constituting two phrase pairs are connected in reverse order A new phrase pair generation step for performing one of the third processes for generating a pair;
Each of the two phrase pairs acquired by the process of the second process in the new phrase pair generation step and the two of the phrase pairs generated by the process of the third process are generated by the control unit. for each phrase pair, the phrase appearance frequency information updating step, the symbol acquisition step, said symbol frequency information updating step, the partial phrase pair generation step, and the recursively performs control processing of the new phrase pair generation step Steps,
The score calculation unit calculates a score for the phrase pair acquired by the generated phrase pair acquisition unit using one or more phrase appearance frequency information stored in the storage medium; and
The phrase table update unit includes a phrase table update step in which the score calculated in the score calculation step is associated with each phrase pair and accumulated in the phrase table. On the storage medium,
In a first language phrase having one or more words in the first language, pairs der the second language phrase having one or more words in the second language is, pairs of the first language sentence and the second language sentence One or more scores having a phrase pair acquired from a bilingual corpus storing one or more parallel translation information having a certain bilingual sentence and a tree structure of the bilingual sentence, and a score that is information relating to the appearance probability of the phrase pair A phrase table with a phrase pair,
One or more phrase appearance frequency information having a phrase pair and F appearance frequency information that is information related to the appearance frequency of the phrase pair;
Storing 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 relating to the appearance frequency of the symbol,
Computer
F appearance frequency information included in the one or more phrase appearance frequency information from one or more phrase pairs constituting a tree structure of one or more bilingual sentences stored in the bilingual corpus, A generated phrase pair acquisition unit that attempts to acquire a phrase pair using the probability distribution of each of the one or more phrase pairs acquired using F appearance frequency information corresponding to each phrase pair ;
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;
A symbol acquisition unit that acquires one symbol using the probability distribution of the one or more symbols acquired using the S appearance frequency information included in the one or more symbol appearance frequency information when the phrase pair cannot be acquired; ,
A symbol appearance frequency information update unit that is S appearance frequency information corresponding to the symbol acquired by the symbol acquisition unit and increases the S appearance frequency information stored in the symbol appearance frequency information storage unit by a predetermined value. When,
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;
A process corresponding to a symbol acquired by the symbol acquisition unit, wherein a new phrase pair is generated using a prior probability of a phrase pair acquired using F appearance frequency information corresponding to one or more phrase pairs . One process or two smaller phrase pairs are generated, and using the one or more phrase appearance frequency information, two first language phrases constituting the two generated phrase pairs are sequentially connected. A second process of generating a single phrase pair having a single language phrase and a new second language phrase sequentially concatenating the two second language phrases that comprise the two phrase pairs, or two smaller phrase pairs Using the one or more phrase appearance frequency information, the two first language phrases constituting the two generated phrase pairs are sequentially linked. One of the third processes for generating one phrase pair having a new first language phrase that is connected and a new second language phrase that is a concatenation of the two second language phrases constituting the two phrase pairs. A new phrase pair generation unit to perform,
For each of the two phrase pairs acquired by the process of the second process by the new phrase pair generation unit, and for each of the two phrase pairs acquired by the process of the third process by the new phrase pair generation unit , A control unit that recursively processes 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;
Using one or more phrase appearance frequency information stored in the storage medium, a phrase pair acquired by the generated phrase pair acquisition unit , a new phrase pair acquired by the new phrase pair generation unit by the first process, A score calculation unit that calculates a score for the phrase pair acquired by the new phrase pair generation unit connected through the second process and the phrase pair acquired by the new phrase pair generation unit through the third process. When,
A program for causing each score calculated by the score calculation unit to correspond to each phrase pair and to function as a phrase table update unit that accumulates in the phrase table.