JP5555542B2 - Automatic word association apparatus, method and program thereof - Google Patents

Description

  The present invention relates to an automatic word association apparatus and method for automatically extracting a correspondence relationship between words from a parallel sentence corpus in which a parallel translation of a source language and a target language after translation is associated in units of sentences. Regarding the program.

  The correspondence between words in the source language and the target language is called word alignment. Conventionally, there are word alignments based on co-occurrence information and those based on a noise channel model. In the method based on co-occurrence information, for example, in the case of word alignment between Japanese and English, the number of times a certain Japanese word appears on the bilingual corpus, the number of times a certain English word appears on the corpus, The Dice coefficient and mutual information are calculated from the number of simultaneous occurrences in the sentence, and the most likely word correspondence is extracted.

  The method based on the noise channel model assumes that the source language is stochastically converted to the target language. For example, when the source language is Japanese and the target language is English, the generation probability of each Japanese word, the translation probability from a Japanese word to an English word, and the alignment probability that indicates the replacement of the word order are learned from the bilingual corpus. Then, word alignment is extracted from those probabilities.

  The word alignment method based on the co-occurrence information and the word alignment method based on the noise channel model are disclosed in Non-Patent Document 1, for example.

F.J Och and H. Ney. 2003.A systematic comparison of various statistical alignment models.Computational Linguistics, 29 (1): 19-51.

  There are two problems with conventional automatic word association devices. One of the problems is that the accuracy of low-frequency word association is not high. For example, in the word alignment method based on statistical information, the more frequently a word in any source language and a word in any target language appear in the parallel translation, the more likely they are in a correspondence relationship. judge. However, a word that appears only once in the bilingual corpus is determined to have the same possibility of being in a bilingual relationship with any word included in the bilingual sentence. For this reason, it is difficult to determine a parallel translation word having a corresponding relationship as the number of words included in the parallel corpus is small.

  The second issue is the problem of word diversity. For example, the English word “head” has a plurality of meanings such as “chairman” and “head”, and the meaning represented by “head” varies depending on the context. Therefore, if the context in which “head” is used is not taken into account, there is a risk of incorrect association.

  The present invention has been made in view of such problems. For example, a synonym dictionary model in which topics such as “company” and “economy” are introduced is constructed, and the synonym dictionary model and the conventional It is an object of the present invention to provide an automatic word association apparatus, a method and a program for learning word alignment using a word association model at the same time.

The automatic word association apparatus according to the present invention includes a training data storage unit, an alignment probability learning unit, and an automatic association unit. The training data storage unit is composed of a bilingual corpus composed of pairs of source language and target language parallel translations separated by words, and a synonym dictionary that is a set of synonyms of the target language. The alignment probability learning unit maximizes, for each topic, the weighted sum of the log likelihood of the bilingual corpus data of the bilingual sentence corpus and the log likelihood of the synonym dictionary probability model of the synonym dictionary data of the synonym dictionary. Learn the parameters to be The automatic association unit receives the target translation sentence and its parameters as input, and generates an alignment between the source language and target language words of the target translation sentence.

  In the automatic word association device according to the present invention, the alignment probability learning unit learns a parameter that maximizes the weighted sum of the log likelihood of the bilingual corpus and the log likelihood of the synonym dictionary. The accuracy of attachment can be improved.

The figure which shows the function structural example of the automatic word matching apparatus 100 of this invention. The figure which shows the operation | movement flow of the automatic word matching apparatus 100. The figure which shows the function structural example of the alignment probability learning part 20. FIG. The figure which shows the operation | movement flow of the alignment probability learning part 20. FIG. The figure which shows the example of the word production | generation probability table of the source language according to a topic. The figure which shows the example of the word translation probability table according to topic. The figure which shows the example of the probability table of a synonym dictionary.

Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated. Prior to the description of the embodiments, the basic concept of the present invention will be described.
[Basic concept of this invention]
This invention is new in that the idea of probability is applied to a synonym dictionary and topic information is further introduced into the synonym dictionary. In the present invention, “meaning” u of synonyms is expressed by a combination u = (k, e) of a topic k and a source language word e. And it is disclosed in Reference Document 1 (Zhao, B. and Xing, EP 2007. HM-BiTAM: Bilingual topic exploration, word alignment, and translation. Twenty-second annual conference on neural information processing systems, Vancouver BC, Canada). Using the parameter set Θ≡ ({α _k }, {β _{k, e} }, {B _{f, e, k} }, {T _{i, i '} }), the synonym dictionary probability model p _m (D _m ; Θ) is calculated as shown in equation (1).

The value of the synonym dictionary probability model p _m (D _m ; Θ) proportional to the sum of the probabilities of the source language e, the object f _s and the synonym f ′ _s for all topics k and source language e. Define as Here, α _k is a parameter of a probability model that generates a mixture ratio of topic k, β _{k, e} is a word generation probability in the source language, B _{fs, e, k} is a word translation probability, and B _{f's, e, k} is synonymous. Word probability.

On the other hand, the probability model p (D _b ; Θ) of the bilingual data is calculated using Expression (2) used in the prior art.

Here, E is a source language sentence, F is a target language sentence, z is a topic, θ is a topic mixing ratio, a is a word alignment, and z, θ, and a are values determined stochastically by the value of the parameter Θ. is there. T _{i, i ′} is a parameter of the probability model for generating the alignment a.

The log likelihood of the bilingual data probability model p (D _b ; Θ) is calculated by equation (4).

  Here, since word alignment a and topic posterior probability p (z, θ, a) cannot be solved analytically, p (z, θ, a) ≈q (θ | approximated as follows: γ) q (z | φ) q (a | λ). Further, q (θ | γ) is assumed to be a direct distribution, q (z | φ) is assumed to be a multinomial distribution, and q (a | λ) is assumed to be a first order hidden Markov model. Expressions (2) to (4) are all conventional techniques.

As described above, the present invention is new in that the idea of probability is applied to the synonym dictionary and topic information is further introduced into the synonym dictionary. The log likelihood of the synonym dictionary probability model p _m (D _m ; Θ) (equation (1)) is calculated by equation (5).

The alignment probability learning unit according to the present invention includes a weighted sum log of the log likelihood of the probability model p (D _b ; Θ) of the bilingual data and the log likelihood of the synonym dictionary probability model p _m (D _m ; Θ). A parameter Θ that maximizes L (Θ) = log p (D _b ; Θ) + ζ log p (D _m ; Θ) is learned. The lower limit value of log L (Θ) is calculated by equation (6).

Here, ζ is a weight given to the synonym dictionary probability model p _m (D _m ; Θ). The automatic word association apparatus of the present invention learns a parameter that maximizes log L (Θ).

  FIG. 1 shows a functional configuration example of the automatic word association apparatus 100 of the present invention. The operation flow is shown in FIG. The automatic word association device 100 includes a training data storage unit 10, an arrangement probability learning unit 20, and an automatic association unit 30.

  The training data storage unit 10 is composed of a bilingual corpus 11 composed of pairs of source language and target language bilingual sentences separated by words, and a synonym dictionary 12 that is a set of synonyms of target words. . The alignment probability learning unit 20 learns a parameter that maximizes the weighted log likelihood of the bilingual data and the synonym dictionary for each topic (step S20). The automatic association unit 40 receives the target translation sentence X and the parameter Θ as input and generates word alignment of the target translation sentence X (step S40).

  The automatic word association apparatus 100 is an apparatus that automatically estimates word alignment from a bilingual sentence. The function of each unit described above is obtained by reading a predetermined program into a computer including, for example, a ROM, a RAM, and a CPU. This is realized by the CPU executing the program.

  The parallel translation corpus 11 is data composed of parallel translation sentence pairs of the source language and the target language. The synonym dictionary 12 is a set of synonym pairs of the target language. For example, when the target language is Japanese, synonym pairs such as “carbon dioxide” and “carbon dioxide”, “student” and “student” are displayed. Collected dictionary data.

The alignment probability learning unit 20 is a weighted sum of the log likelihood of the bilingual corpus 11 and the log likelihood of the synonym dictionary 12 log L (Θ) = log p _b (D _b ; Θ) + ζ log p _m (D _The parameter Θ that maximizes _m ; Θ) is learned (step S20). However, D _b, respectively D _m, represent the data of the data, a synonym dictionary 12 bilingual corpus 11 is prepared as training data. P _b (D _b ; Θ) is a probabilistic model of the bilingual corpus 11 represented by the parameter Θ, and takes a value from 0 to 1. Similarly, p _m (D _m : Θ) is a probability model of the synonym dictionary 12. The ζ is a weight representing the degree of importance percentage of the translation data D _b and synonym dictionary data D _m. The learned parameters are stored in a memory (not shown) in the alignment probability learning unit 20.

The probabilistic models p _b (D _b ; Θ) and p _m (D _m : Θ) are formulated with unknown information from the training data as a latent variable Z. For example, the present invention uses a probability model in which word correspondence a and topic z information are introduced as latent variables. Therefore, the probability model p _b (D _b ; Θ) of the bilingual sentence data is p _b (D _b ; Θ) = Σ _Z p _b (D _b , Z _b ; Θ), Z _b = (a, z) Become. Similarly, the probability model synonym dictionary data p _m (D _m: Θ) _{is, p m (D m: Θ} ) = Σ Z p m (D m, Z m; Θ) becomes.

  The automatic association unit 30 estimates the correspondence relationship between the source language and the target language of the target translated sentence X using the parameter Θ learned by the alignment probability learning unit 20 (step S40).

  The automatic word association apparatus 100 that operates as described above constructs a synonym dictionary model in which topics are introduced. Then, automatic word association can be performed with high accuracy by simultaneously learning the word alignment using the synonym dictionary model and the conventional word association model.

  For example, by using information that “carbon dioxide gas” and “carbon dioxide” are synonyms, “carbon dioxide” is associated with the same translation as “carbon dioxide”. At this time, even if the bilingual sentence data contains almost no “carbon dioxide”, if the bilingual sentence data contains many “carbon dioxide” and is associated with the correct bilingual word “carbon dioxide”, then “carbon dioxide” "Can also be associated with the correct bilingual word" carbon dioxide ".

  In addition, by introducing topic information, it is possible to deal with the issue of ambiguity. For example, the word “head” is synonymous with both the words “forefront” or “chief”, but the meanings of “forefront” and “chief” are different. That is, the meaning of “head” varies depending on the context. By using a synonym dictionary model that introduces the concept of a topic, it automatically learns which is the synonym of “head” according to the topic of the whole sentence. As a result, synonyms can be used for learning word alignment, and an improvement in alignment accuracy can be expected.

  FIG. 3 shows a specific functional configuration example of the alignment probability learning unit 20, and the operation of the automatic word association apparatus 100 will be described in more detail. The alignment probability learning unit 20 includes a reference value calculation unit 21, a word alignment probability calculation unit 22, a synonym dictionary probability calculation unit 23, a parameter update unit 24, and a convergence determination unit 25.

The alignment probability learning unit 20 obtains the optimum parameter Θ ^ by sequentially updating the parameters Θ of the probability models p _b (D _b ; Θ) and p _m (D _m : Θ) until they converge. The operation flow is shown in FIG.

The reference value calculation unit 21 reads the bilingual sentence data D _b and the synonym dictionary data D _m stored in the training data storage unit 10, and sets an initial value Θ ⁽⁰⁾ (step S21).

The word alignment probability calculation unit 22 calculates the posterior probability p _b (Z _b | D _b : Θ ⁽⁰⁾ ) of the latent variable Z with the initial value Θ ⁽⁰⁾ of the parameter as input, and then inputs from the convergence determination unit 25. The posterior probability p _b (Z _b | D _b : Θ ^(t) ) of the latent variable Z is calculated from the parameter Θ ^(t) (step S22). Θ ^(t) represents a parameter obtained in the t-th update step.

The synonym dictionary probability calculation unit 23 calculates the posterior probability p _m (Z _m | D _m : Θ ^(t) ) of the latent variable Z from the initial value Θ ⁽⁰⁾ of the parameter and the parameter Θ ^{(t) being} updated. (Step S23).

The parameter updating unit 24 uses the posterior probability p _b (Z _b | D _b : Θ ^(t) ) of the latent variable Z calculated by the word alignment probability calculating unit 22 and the latent calculated by the synonym dictionary probability calculating unit 23. Using the posterior probability p _b (Z _b | D _b : Θ ^(t) ) of the variable Z as an input, the word translation probability recorded in the topic-specific word translation probability table 241 and the topic-specific source language word generation probability table 242 and the word generation probability, which is recorded in, and synonym dictionary probability that has been recorded in the synonym dictionary probability table 243, and word order swapping probability of a word to the target language from the original language, and the generation probability of the mixture ratio of topics refer. Then, the word generation probability p (E _n | z _n ; β) of the source language, the word translation probability p (F _m | E _n , z _n , a _n ; B) from the source language to the target language, and the translation of the training data The topic z _{n of} the sentence, the mixing ratio θ of the topics of the entire training data, the word alignment a of the source language and the target language are estimated, and a new parameter Θ ^{(t + 1)} is calculated based on the estimated value (step ⁾ S24).

  FIG. 5 shows an example of the word generation probability table 242 of the source language for each topic. FIG. 6 shows an example of the word translation probability table 241 for each topic. FIG. 7 shows an example of the synonym dictionary probability table 243.

The convergence determination unit 25 calculates the log likelihood log L (Θ ^{(t + 1)} ) calculated from the parameter Θ ^{(t + 1)} input from the parameter update unit 24 and the convergence condition log L (Θ ^{(t (t +1)} ) -log If L (Θ ^(t) ) <ε is satisfied, the parameter estimate is updated as Θ ^ ← Θ ^{(t + 1)} . When the convergence condition is not satisfied, the update step of the parameter Θ ^{(t + 1)} is set to t ← t + 1, and the updated parameter Θ ^{(t + 1)} is calculated again with the word alignment probability calculation unit 22 and the synonym dictionary probability calculation. To the unit 23 (step S25, unconvergence).

The processes in steps S22 to S25 are repeatedly executed until the convergence condition is satisfied. The automatic association unit 30 extracts the optimum association an _n ^ between the source language and the target language of the target translation sentence X using the learned optimum parameter Θ ^ (formula (7)).

  A specific example of each process described above will be shown and described in more detail. In the word alignment problem, the source language sentence is probabilistically converted to generate the target language sentence. At this time, a bilingual sentence set (formula (8)) of the source language and the target language is given as the bilingual data, and a synonym pair set (formula (9)) of the target language is given as the synonym dictionary.

Here, E _n is the nth source language sentence in the parallel translation data, and F _n is the nth target language sentence. (f _s , f ' _s ) represents a synonym pair of the target language.

The case where HM-BiTAM disclosed in Reference Document 1 is used as an existing word alignment extraction technique will be described. HM-BiTAM gives a probabilistic model of the bilingual set of source and target languages. The latent variable Z _b = (z, a) represents the word alignment a and the topic z. In HM-BiTAM, a prior distribution is set with respect to the topic mixture ratio θ which is a parameter of topic z. Therefore, the mixture ratio θ of the topic does not appear explicitly in the learning of the model, and instead, a prior distribution parameter (hyper parameter) is estimated, and therefore, Z _b = (z, a, θ) may be set for convenience. The variable a _jn = i representing the word alignment indicates that in the parallel translation n, the j _nth word in the target language and the i th word in the source language are in a correspondence relationship. One topic z is assigned to each bilingual sentence (E _n , F _n ). The topic mixing ratio θ is a frequency distribution representing the generation probability of each topic.

Next, the synonym dictionary probability model of the present invention will be described. The synonym dictionary probability model gives a probability to a synonym dictionary that is a set of synonym pairs (f _s , f _s ) of the target language. A synonym is a representation of the “meaning” of a word in a different expression and is defined as shown in equation (10).

  Here, u represents the meaning of a synonym. In this embodiment, the meaning u of a synonym is expressed by a combination of a topic k and a source language word e. That is, u = (k, e). FIG. 7 shows an example of a synonym dictionary probability table expressed by a synonym dictionary probability model (synonym dictionary probability table 243 in FIG. 3).

Under the above assumptions, the bilingual sentence data probability model p (D _b ; Θ) can be expressed by the above equation (2), and the synonym dictionary probability model p _m (D _m ; Θ) can be expressed by the equation (1). The alignment probability learning unit 20 is a weighted sum log L () of the log likelihood of the bilingual data probability model p (D _b ; Θ) and the log likelihood of the synonym dictionary probability model p _m (D _m ; Θ). A parameter Θ that maximizes Θ) = log p (D _b ; Θ) + ζ log p (D _m ; Θ) is learned. The lower limit value of log L (Θ) is calculated by the above equation (6).

The word alignment probability calculation unit 22 sets the parameter γ _k of the topic mixture ratio θ that maximizes the equation (6) to the equation (11), and the parameter φ _{n, k} of the topic z probability model to the equation (12). Then, the parameter λ _{n, j, i} of the probability model of word alignment a is obtained by equation (13).

The synonym dictionary probability calculation unit 23 calculates the posterior probability p _m (k, e | D _m ; Θ ^(t) ) because the latent variable Z _m = (k, e) (formula (14)).

Here, β _{k, e} is the generation probability of the source language word for each topic, and is calculated by equation (15). B _{f, e, k} is a word translation probability from the source language to the target language for each topic, and is calculated by Expression (16). α _k is a parameter of the probability model of the topic mixing ratio θ, and is calculated by Expression (17). T _{i ′, i} is the word order switching probability of the words in the source language and the target language, and is calculated by equation (18).

  The parameters shown in Expression (11) to Expression (18) are obtained by partial differentiation of Expression (6) indicating the log likelihood log L (Θ) of training data with each parameter.

The convergence determination unit 25 determines whether or not the updated parameter Θ ^{(t + 1)} has converged. For example, the log likelihood of the training data is compared before and after the update, and if the difference is less than ε, it is determined that the training has converged (formula (19)).

〔Experimental result〕
An evaluation experiment was conducted for the purpose of confirming the effect of the automatic word association method of the present invention. Reference translation 2 (R. Mihalcea and T. Pedersen. 2003. An evaluation exercise for word alignment. In Proceedings of the HLT-NAACL 2003 Workshop on Building and using parallel texts: data driven machine translation and beyond-Voiume 3, page 10. Association for Computational Linguistics. This is an English-French bilingual corpus. As a synonym dictionary, WordNet 2.1 shown in Reference 3 (GA Miller. 1995. WordNet: alexical database for English. Communications of the ACM, 38 (11): 41.) Was used.

  As evaluation data, 10,000 translations were randomly extracted from the Hansards data set. In addition, among the synonyms posted on WordNet, synonym pairs that appear at least once in the evaluation parallel translation data set were used as synonym dictionaries. Using these heterogeneous data and synonym dictionaries as training data, word alignment between English and French was estimated and evaluated. The results are shown in Table 1. As an evaluation index, accuracy, recall, F value, and AER (Alignment Error Rate) generally used in word alignment were used.

この発明の自動単語対応付け方法の精度が0.941と最も高く、この発明の効果が確認できた。

The accuracy of the automatic word association method of the present invention was the highest at 0.941, and the effect of the present invention was confirmed.

  When the processing means in the above apparatus is realized by a computer, the processing contents of functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

Claims (9)

A training data storage unit comprising a bilingual corpus composed of a pair of bilingual sentences of a source language and a target language separated by words, and a synonym dictionary that is a set of synonyms of the target language;
Alignment probability for learning a parameter that maximizes the weighted sum of the log likelihood of the probability model of the parallel sentence data of the parallel sentence corpus and the log likelihood of the synonym dictionary probability model of the synonym dictionary data of the synonym dictionary The learning department,
An automatic association unit that generates an alignment between a source language and a target language word of the target translation sentence, using the target translation sentence and the parameter as input,
An automatic word association apparatus comprising: In the automatic word matching device according to claim 1,
The synonym dictionary Sho確 rate model,
An automatic word association device comprising a generation probability for each meaning of a pair of synonyms. In the automatic word matching device according to claim 1,
The synonym dictionary Sho確 rate model of the target language is,
Suppose the meaning of synonyms can be represented by a combination of words belt pick the source language used in word alignment, the word translation probabilities by the topic, a word generation probability of the source language, the generation of the mixing ratio of the topic An automatic word association device, characterized in that the automatic word association device is defined as a value proportional to a sum product of a probability and a synonym probability . In the automatic word matching apparatus according to claim 1 or 3,
The alignment probability learning unit
The word translation probability recorded in the topic-specific word translation probability table, the word generation probability recorded in the topic-specific source language word generation probability table, the synonym dictionary probability recorded in the synonym dictionary table, and the original Refer to the probabilities of changing the order of words from the language to the target language and the probability of generating the mixing ratio of topics. The topic of the parallel translation of the training data, the mixing ratio of the topics of the entire training data, the source language and the target language Estimate alignment between words,
The parameter Θ is updated based on the estimated value,
Read the bilingual sentence data and the synonym dictionary data, and calculate the reference value of the parameter Θ that can be expressed as a weighted sum of the log likelihood of the probability model of the bilingual sentence data and the log likelihood of the synonym dictionary probability model A reference value calculation unit,
A word alignment probability calculation unit for calculating a posteriori probability of a latent variable included in the bilingual sentence data probability model , given the reference value of the parameter Θ and the current parameter Θ ^(t) ;
A synonym dictionary probability calculation unit that calculates a posteriori probability of a latent variable included in the synonym dictionary probability model , given the reference value of the parameter Θ and the current parameter Θ ^(t) ,
A parameter updater that calculates a new parameter Θ ^{(t + 1)} from the current parameter Θ ^(t) ;
The said parameter theta to calculate the log-likelihood using a ^{(t + 1),} the parameter Θ ^{(t + 1)} is the convergence determining unit determines whether the convergence conditions are satisfied optimum parameter estimates theta ^,
The automatic word matching apparatus characterized by comprising. A set of log likelihoods of the bilingual corpus bilingual text data probabilistic model and synonyms of the target language described above, in which the alignment probability learning unit is composed of pairs of source language and target language bilingual sentences separated by words. An alignment probability learning process for learning a parameter that maximizes a weighted sum of logarithmic likelihoods of the synonym dictionary probability model of the synonym dictionary data of the synonym dictionary ,
An automatic association process in which an automatic association unit generates an alignment between a source language and a target language word of the target translation sentence by inputting the target translation sentence and the parameter;
Automatic word matching method including In the automatic word matching method according to claim 5,
The synonym dictionary Sho確 rate model,
An automatic word association method comprising a generation probability for each meaning of a synonym pair. In the automatic word matching method according to claim 5,
The synonym dictionary Sho確 rate model of the target language the meaning of synonyms,
Suppose it can be expressed by a combination of words belt pick the source language used in word alignment, the word translation probabilities by the topic, a word generation probability of the source language, and generation probability of the mixing ratio of the topic, synonyms An automatic word association method, characterized in that the automatic word association method is defined as a value proportional to a sum product with a probability . In the automatic word matching method according to claim 5 or 7,
The above alignment probability learning process
The word translation probability recorded in the topic-specific word translation probability table, the word generation probability recorded in the topic-specific source language word generation probability table, the synonym dictionary probability recorded in the synonym dictionary table, and the original Refer to the probabilities of changing the order of words from the language to the target language and the probability of generating the mixing ratio of topics. The topic of the parallel translation of the training data, the mixing ratio of the topics of the entire training data, the source language and the target language Estimate alignment between words,
The parameter Θ is updated based on the estimated value,
A parameter that the reference value calculation unit reads the bilingual sentence data and the synonym dictionary data and can represent as a weighted sum of the log likelihood of the probability model of the bilingual sentence data and the log likelihood of the synonym dictionary probability model A reference value calculating step for calculating a reference value of Θ;
A word alignment probability calculating step for calculating a posteriori probability of a latent variable included in the bilingual sentence data probability model , given the reference value of the parameter Θ and the current parameter Θ ^(t) ;
A synonym dictionary probability calculation step for calculating a posteriori probability of a latent variable included in the synonym dictionary probability model , given the reference value of the parameter Θ and the current parameter Θ ^(t) ,
A parameter update step in which the parameter update unit calculates a new parameter Θ ^{(t + 1)} from the current parameter Θ ^(t) ;
Convergence determination unit, the parameter Θ ^{(t + 1)} to calculate the log-likelihood using a convergence determination determines whether the parameter Θ ^{(t + 1)} is the optimum parameter estimates theta ^ or convergence condition is satisfied Steps,
The automatic word matching method characterized by including this. Programs for each part of the function of the automatic word associating device according to either one of claims 1 to 4, causes the computer to execute.

Images