JP5548176B2 - Translation optimization apparatus, method, and program - Google Patents

Description

  The present invention relates to a translation optimization device, method, and program, and more particularly, to a translation optimization device, method, and program that optimize weights for a plurality of translation feature models used in translation processing.

  A statistical translation system typically uses multiple knowledge bases. For example, in the case of translation from Japanese to English, a translation model (a knowledge base representing the most likely English translation for the input Japanese word) and a language model (representing the naturalness of the output English expression) The most likely translation for the input sentence (Japanese) after integrating various models such as knowledge base) and rearrangement model (knowledge base representing the likelihood of grammatical structure conversion between input and output languages) English).

  The likelihood of the translation is expressed by a weighted combination of model features. Here, one of the most important steps is the optimization of the weight (Weight Optimization). The conventional method uses only one translation quality evaluation metric as an objective function when optimizing the weights.

  The flow of weight optimization is as follows. First, for example, Japanese text (input) and correct English translation (output) are prepared as development data. (1) The machine translation system translates the input sentence and compares the output result with the correct translation. (2) Adjust the model weight. This (1) to (2) is repeated many times, and finally a weight is obtained when the output and the correct translation sufficiently match. The following two representative methods have been used for such optimization of statistical translation.

  One technique is MERT (Minimum Error Rate Training) (Non-Patent Document 1), and another technique is MIRA (Margin Infused Relaxed Algorithm) (Non-Patent Document 2). Both methods translate repeatedly and adjust the weights to produce the best translation in terms of a translation quality rating scale.

Thus, the evaluation scale plays an important role for optimizing the weight. A good rating scale gives high scores for correct translations and low scores for bad translations. The most widely used evaluation scale at present is a scale called BLEU (Non-patent Document 3). BLEU is calculated based on the degree of coincidence between successive partial word strings between the output and the correct translation (the BLEU value is high for good translation). Another representative evaluation scale is TER (Translation Edit Rate) (Non-patent Document 4). The TER is calculated based on how many steps are required when rewriting a translated sentence into a correct translation (the TER value is low when the translation is good).
Many evaluation scales have been proposed so far, but no evaluation scale that matches human evaluation is known. Actually, each evaluation scale evaluates translation quality from different viewpoints. In this sense, all evaluation scales have a certain level of rationality. Therefore, the reality is that machine translation developers measure system performance using multiple evaluation measures.

Franz Och, “Minimum Error Training in Statistical Machine Translation,” Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2003. Taro Watanabe, Jun Suzuki, Hajime Tsukada, Hideki Isozaki, “Oneline Large-Margin Training for Statistical Machine Translation,” Proceedings of the Conference on Empirical Methods for Natural Language Processing (EMNLP-CoNLL), 2007. Kishore Papineni, Salim Roukas, Todd Ward, Wei-jing Zhu, “BLEU: A method for automatic evaluation of machine translation,” Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2002. Matt Snover, Bonnie Dorr, Richard Schart, Linnea Micciulla, John Makhoul, “A study of Translation Edit Rate with Targeted Human Annotation,” Proceedings of Meeting of Association for Machine Translation for the Americas (AMTA), 2006.

  However, the optimization method using MERT and MIRA described in Non-Patent Documents 1 and 2 described above can use only one evaluation measure. Therefore, although it is optimized so that a good translation can be generated from a certain viewpoint, there is a problem that it is not guaranteed that the translation is appropriate from another viewpoint. This corresponds to the fact that the developer can use only BLEU or TER as an evaluation scale, for example. If optimized with BLEU, TER may not be optimal. Similarly, BLEU may worsen if optimized with TER. If there is an imbalance that one rating scale is good while the other rating scale is bad, it will never be a good translation when humans evaluate.

  The present invention has been made in view of the above circumstances, and provides a translation optimization device, method, and program capable of optimizing weights for a plurality of translation feature models by simultaneously using a plurality of evaluation scales. The purpose is to do.

In order to achieve the above object, a translation optimization apparatus according to the present invention converts a translation source language word string into a translation destination based on a plurality of types of translation feature models and weights for each of the plurality of types of translation feature models. multiple illustrating the translation processing means for translating the plurality of candidate h word string language, for each of a plurality of candidate h word string of the target language translated by said translation processing means, a plurality of types of translation rating scale score M _{1 (h),} a translation evaluation means for calculating the M 2 (h), the score of each of the plurality of calculated for a plurality of candidate h M _{1 (h),} based on M 2 (h) Using the following formula, while gradually increasing the variable α from 0 to infinity, a candidate h ^* satisfying Pareto optimality is extracted from the plurality of candidate sets L at each value of the variable α. The extracted Pareto identifying means for identifying a candidate whose score M ₂ (h) is smaller than the candidate h ^{* as} a candidate that does not satisfy the Pareto optimality, and the plurality of types of translation features according to the identification result by the Pareto identifying means Repeating the weight update means for updating the weight for each of the models, the translation by the translation processing means, the calculation by the translation evaluation means, the identification by the Pareto identification means, and the update by the weight update means, the plurality of types Optimizing means for optimizing the weight for each translation feature model.

A translation optimizing method according to the present invention is a translation optimizing method in a translation optimizing device including a translation processing unit, a translation evaluating unit, a Pareto identifying unit, a weight updating unit, and an optimizing unit, wherein the translation optimizing device It is by the translation processing unit, based on the weight for each of a plurality of types of translation feature model and the plurality of types of translation feature model, a word string of the source language, a plurality of candidate h word string of target language A plurality of scores M ₁ (h) indicating a plurality of types of translation evaluation scales for each of a plurality of candidates h of the word string of the translation target language translated by the translation processing means by the translation evaluating means; ), M ₂ (h) , and the plurality of scores M ₁ (h), M calculated for each of the plurality of candidates h by the Pareto identification means. ₂ Based on (h) , the Pareto optimality is satisfied from the set L of the plurality of candidates at each value of the variable α while gradually increasing the variable α from 0 to infinity using the following formula: extracts a candidate h ^*, the score M _{2 (h)} is smaller candidates from the extracted candidate h ^*, identifying as a candidate that does not meet the Pareto optimality, by the weight updating unit, wherein Updating the weights for each of the plurality of types of translation feature models according to the identification result by the Pareto identification means, the translation by the translation processing means by the optimization means, the calculation by the translation evaluation means, and the Pareto identification The weight for each of the plurality of types of translation feature models is optimized by repeating the identification by the means and the update by the weight update means. And executes includes a step, a.

  According to the present invention, the translation processing means converts the plurality of types of translation feature models and the plurality of types of translation feature models into weights of the source language word strings and the translation destination language word strings. Translate to candidate. The translation evaluation means calculates a plurality of scores indicating a plurality of types of translation evaluation scales for each of a plurality of candidates for the word string in the translation target language translated by the translation processing means.

  Then, the Pareto identifying means identifies whether each of the plurality of candidates satisfies Pareto optimality based on the plurality of scores calculated for each of the plurality of candidates. The weight updating unit updates the weight for each of the plurality of types of translation feature models according to the identification result by the Pareto identification unit.

  Then, the optimization unit repeats the translation by the translation processing unit, the calculation by the translation evaluation unit, the identification by the Pareto identification unit, and the update by the weight update unit. Optimize the weights.

  In this way, by calculating a plurality of scores indicating a plurality of types of translation evaluation scales for each of a plurality of candidates for the word string in the translation target language, it is identified whether or not Pareto optimality is satisfied, and according to the identification result By updating the weights for each translation feature model, it is possible to optimize the weights for a plurality of translation feature models by simultaneously using a plurality of evaluation scales.

The Pareto identification means according to the present invention sets the variable α to 0 using the following formula based on the plurality of scores M ₁ (h) and M ₂ (h) calculated for each of the plurality of candidates h. The candidate h ^* satisfying the Pareto optimality is extracted from the plurality of candidate sets L at each value of the variable α while gradually increasing from 1 to infinity, and the score M is extracted from the extracted candidate h ^*. the _{candidate 2 (h)} is smaller, you to be identified as candidates that do not satisfy the Pareto optimality.

  The translation processing means according to the present invention translates each of the plurality of word strings of the translation source language into a plurality of candidates of the word strings of the translation destination language, and the weight update means responds to the identification result by the Pareto identification means. When each candidate for a plurality of word strings in the target language is identified as satisfying Pareto optimality, a maximum value of the score is given as each score for the candidate, For each candidate, a cumulative score obtained by accumulating the total value of the scores can be calculated, and the weight for each of the plurality of types of translation feature models can be updated based on the calculated cumulative score.

  The weight updating unit according to the present invention is configured such that when at least one of the candidates for the word string in the translation destination language is identified as not satisfying the Pareto optimality according to the identification result by the Pareto identifying unit, The weight for each type of translation feature model can be updated.

  The program according to the present invention is a program for causing a computer to function as each unit of the translation optimization device.

  As described above, according to the translation optimization device, method, and program of the present invention, a plurality of scores indicating a plurality of types of translation evaluation scales are calculated for each of a plurality of candidates for a word string of a translation target language. , Identify whether or not Pareto optimality is satisfied, and update the weights for each translation feature model according to the identification result, thereby simultaneously optimizing the weights for multiple translation feature models using multiple evaluation measures The effect that it can be performed is acquired.

It is the schematic which shows the structure of the translation optimization apparatus which concerns on the 1st Embodiment of this invention. It is a figure for demonstrating Pareto optimality. It is a figure which shows the algorithm which calculates | requires a Pareto frontier. It is a flowchart which shows the content of the translation optimization process routine in the translation optimization apparatus which concerns on the 1st Embodiment of this invention.

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

<First Embodiment>
<System configuration>
The translation optimizing apparatus 100 according to the first embodiment of the present invention receives a plurality of development data composed of an input sentence in a source language and a correct output sentence in a destination language, and optimizes weights for a plurality of feature models. Turn into. This translation optimizing apparatus 100 is composed of a computer including a CPU, a RAM, and a ROM storing a program for executing a translation optimizing process routine to be described later, and functionally configured as shown below. Has been. As shown in FIG. 1, the translation optimization apparatus 100 includes an input unit 10 and a calculation unit 20.

  The input unit 10 accepts a plurality of data sets composed of a Japanese input sentence (word string) as a translation source and an output sentence (word string) as a correct English translation as input development data.

  The computing unit 20 includes a development data storage unit 21, a feature amount weight storage unit 22, a translation unit 23, a translation evaluation unit 24, a Pareto identification unit 25, and an optimization unit 26. The optimization unit 26 is an example of a weight update unit and an optimization unit.

  The development data storage unit 21 stores a plurality of development data received by the input unit 10.

  The feature amount weight storage unit 22 is a weight vector W (= (w1) comprising weights for each of a plurality of translation feature models (for example, a translation model, a language model, and a rearrangement model) referred to in the translation processing by the translation unit 23. , W2, w3)).

  For each development data, the translation unit 23 translates the input sentence of the development data as follows, and outputs a translation sentence output candidate list.

  First, the translation unit 23 performs a translation process on the input sentence of the development data to obtain a plurality of translation candidates. In addition, about a translation process, since a conventionally known method should just be used, detailed description is abbreviate | omitted. The translation unit 23 calculates a feature amount for each of a plurality of translation feature models for each translation candidate, and calculates a feature vector F (= (f1, f2, f3)) composed of the feature amounts of each translation feature model. The translation unit 23 uses, for each translation candidate, a weight vector W composed of weights for each translation feature model stored in the feature amount weight storage unit 22 to add a weighted addition value F · W of the feature amount of each translation feature model. (= W 1 · f 1 + w 2 · f 2 + w 3 · f 3)) is calculated as a translation feature score. The translation unit 23 arranges the translation candidates in descending order of the translation feature score, identifies the top k translation candidates, and outputs the identified top k translation candidates as a translation output candidate list.

  The translation evaluation unit 24 compares each translation candidate in the translation sentence output candidate list output by the translation unit 23 with a correct translation sentence included in the development data, and compares a plurality of types of evaluation scales (for example, BLEU and A plurality of types of evaluation scale scores (BLEU value, 1-TER value) are calculated for each of (TER).

  The Pareto identifying unit 25 identifies whether or not the Pareto optimality is satisfied for each translation candidate in the output candidate list of the translation sentence based on the plurality of types of evaluation scale scores calculated by the translation evaluating unit 24. Output a subset of the output candidate list consisting of translation candidates that satisfy Pareto optimality.

  Here, the concept of Pareto described in non-patent literature (Marler and Arora, “Survey of multi-objective optimization methods for Engineering”, in Journal of Structural and Multidisciplinary Optimization, vol. 26, 2004) will be described. In order to simplify the explanation, the translation unit 23 outputs only two translation candidates (H1, H2), and this will be described as a problem that determines which of these is an effective translation.

  The translation evaluation unit 24 calculates a score vector of each translation candidate included in the output candidate list (each element is composed of a BLEU value of the output candidate and (1-TER value)). Here, the score vector of translation candidate H1 is represented by M (H1), and the score vector of translation candidate H2 is represented by M (H2). The higher the value of each element of the score vector, the better the translation. If all the elements of M (H1) are larger than all the elements of M (H2), it can be said that the translation candidate H1 is a better translation than the translation candidate H2 on all evaluation scales. Conversely, if all the elements of the score vector M (H2) are larger than all the elements of the score vector M (H1), it can be said that the translation candidate H2 is a better translation than the translation candidate H1. However, when a certain element has a higher score vector M (H1) and other elements have a higher M (H2), it cannot be generally said which translation candidate H1 or H2 is a better translation. In such a case, the translation candidates H1 and H2 are considered to be translated sentences having the same quality. It is Pareto optimality that formulated such intuition. Below, the definition of weak pareto optimality (Weak Pareto-optimality) and pareto optimality (Pareto-optimality) is demonstrated.

  (Definition 1) Weak Pareto Optimality: A translation sentence output candidate list L is given, and there is no H ′ that satisfies M (H ′)> M (H) in all elements, and only in that case, H Is said to satisfy weak Pareto optimality.

(Definition 2) Pareto Optimality: A translation output candidate list L is given, M (H ′) ≧ M (H) for all elements, and M _i (H ′) for a certain element i It is said that H satisfies Pareto optimality when and only when H ′ that satisfies> M _i (H) does not exist.

  The above definition will be briefly described with reference to FIG. 2 using ten translation candidates and two evaluation scales (metric 1 and 2) as examples. In FIG. 2, the translation candidates indicated by “◯” satisfy Pareto optimality (and weak Pareto optimality). On the other hand, candidates indicated by “Δ” do not satisfy Pareto optimality but satisfy weak Pareto optimality. A translation candidate that does not satisfy the weak Pareto optimality indicated by “+” is referred to as a non-pareto point. For example, the point (0.4, 0.7) indicates a translation candidate that satisfies the Pareto optimality. This is because there are no other translation candidates that satisfy (metric1> = 0.4 and metric2> = 0.7). On the other hand, the point (0.6, 0.4) becomes metric 1 0.7> 0.6 and metric 2 0.6> 0.4 when compared with other points (0.7, 0.6). Therefore, it becomes a non-Pareto point.

  That is, the weak Pareto optimality is satisfied if there is no case where other translation candidates win on all evaluation scales with respect to a certain translation candidate. For a translation candidate, if there is another translation candidate that has an evaluation value that is equal to or greater than that of all evaluation scales and that has a higher (greater) evaluation value in at least one evaluation scale, Some translation candidates do not satisfy Pareto optimality. Conversely, when there is no such other translation candidate, the certain translation candidate satisfies the Pareto optimality.

  When a translation sentence output candidate list L is given, a set of all translation candidates satisfying the Pareto optimality is referred to as a Pareto Frontier.

  The purpose of the Pareto identification unit 25 is to efficiently find the Pareto frontier. Conventionally, several algorithms have been proposed, but in the present embodiment, as an example, an algorithm obtained by extending Convex Hull Algorithm (convex hull algorithm) is used. With this extension, an approximate Pareto frontier solution can be obtained at high speed. The algorithm used in this embodiment can be used when there are two evaluation measures.

In the algorithm used in the present embodiment, the maximum point (translation candidate) h ^* is obtained for each of the linear sums of the two evaluation measures according to the following equation (1).

Here, by changing the value of α from zero to infinity, it can be proved that all translation candidates h ^* obtained for each value of α satisfy the weak Pareto optimality.

  Since the equation (1) is a piecewise constant with respect to α, the Pareto identification unit 25 can use a high-speed algorithm to search for all values related to 0 ≦ α <∞. The pseudo-code is shown in FIG.

As shown in FIG. 3, the Pareto identification unit 25 uses the above formula (1) based on the evaluation scale scores M ₁ (h) and M ₂ (h) calculated for each of the plurality of translation candidates h. Then, a translation candidate h ^* satisfying the Pareto optimality is extracted from the translation sentence output candidate list L with the variable α = 0 as an initial value. Then, a translation candidate whose evaluation scale score M2 (h) is smaller than the extracted translation candidate h ^* is identified as a non-Pareto point.

For each remaining translation candidate h, α (h) is calculated using the following equation (2), and the above processing is repeated with the minimum value of α (h) as the next variable α. The above processing is repeated until the variable α becomes infinite. When the processing is completed, a set of translation candidates extracted as translation candidates h ^* satisfying Pareto optimality is defined as a Pareto frontier.

  As will be described below, the optimization unit 26 optimizes the weights for the plurality of translation feature models based on the evaluation scale score calculated by the translation evaluation unit 24 and the identification result by the Pareto identification unit 25.

  A translation candidate that satisfies the Pareto optimality is a translation obtained in consideration of a plurality of evaluation scale scores, and is therefore considered to be a good translation. Therefore, the optimization unit 26 adjusts the weight of each translation feature model so that a translation candidate satisfying the Pareto optimality in the translation candidate list has a high translation scale score. In the present embodiment, the optimization unit 26 uses a multi-objective optimization technique obtained by extending MERT. MERT is an optimization method for updating weights using a gradient-free method such as the POWELL method.

  The multi-objective optimization method used in the present embodiment operates according to the following procedure.

  First, a translation candidate list is obtained for each development data using a randomly determined weight vector W. Then, based on the evaluation scale score calculated by the translation evaluation unit 24 and the identification result by the Pareto identification unit 25 for each of a plurality of development data, a cumulative score is calculated as follows.

For translation candidates that do not satisfy Pareto optimality, a total score of M ₁ (H) + M ₂ (H) is given, while for translation candidates included in Pareto frontier (translation candidates that satisfy Pareto optimality) Gives the total score of M ₁ _MAX + M ₂ _MAX. Here, M ₁ _MAX is the maximum score of one evaluation scale in the translation candidate list, and M ₂ _MAX represents the maximum score of the other evaluation scale in the translation candidate list.

  For each candidate list, the above total score is calculated, and the total score in all the development data is totaled to calculate the total score.

  The weight vector W is repeatedly determined at random, the weight vector W is updated each time, and the cumulative score is calculated. The weight vector W with the maximum cumulative score is specified, and the weight vector stored in the feature amount weight storage unit 22 is updated as the optimized weight vector W.

  The translation optimization apparatus 100 repeatedly performs a series of processes by the translation unit 23, the translation evaluation unit 24, the Pareto identification unit 25, and the optimization unit 26. Every time the weight vector is updated, the maximum value of the cumulative score is gradually improved. When the improvement of the maximum value of the cumulative score stops, the repetition of the series of processes is terminated. For example, when the improvement of the cumulative score has stopped even after being repeated a predetermined number of times or more, the series of processing is terminated. Then, the weight vector that finally stores the feature amount weight storage unit 22 and gives the maximum cumulative score is set as the optimization result.

<Operation of translation optimization device>
Next, the operation of translation optimization apparatus 100 according to the present embodiment will be described. First, when a plurality of development data composed of a Japanese input sentence and a correct English translation is input to the translation optimization apparatus 100, the plurality of development data input by the translation optimization apparatus 100 is converted into the development data. It is stored in the storage unit 21. Then, the translation optimization apparatus 100 executes the translation optimization processing routine shown in FIG.

  First, in step S101, all of a plurality of development data is acquired from the development data storage unit 21. In step S <b> 102, a weight for each translation feature model is determined at random and stored in the feature amount weight storage unit 22. In step S103, the input sentence of each development data acquired in step S101 is subjected to translation processing, and a translation sentence composed of translation candidates whose translation score obtained by weighted addition of feature amounts to the translation feature model is the top k. Output candidate list is generated for each development data.

  In the next step S104, for each development data, a plurality of types of translation evaluation scale scores are calculated for each translation candidate included in the translation candidate output list generated in step S103 for the development data.

  In step S105, for each development data, each translation candidate included in the translation candidate list is Pareto based on the plurality of types of translation evaluation scale scores calculated in step S104 for each translation candidate of the development data. Identify whether or not the optimality is satisfied, and obtain a Pareto frontier for each development data.

  In the next step S106, for each development data, based on the translation evaluation scale score of each translation candidate calculated in step S104 and the Pareto frontier obtained in step S105, the total score of each translation candidate Is calculated, and the total score of each development data is accumulated to calculate the cumulative score. Based on the calculated cumulative score, the weight for each translation feature model that gives the maximum cumulative score is specified. In step S107, the weight of each translation feature model stored in the feature amount weight storage unit 22 is updated to the weight for each translation feature model specified in step S106.

  In step S108, it is determined whether or not to end the repetition process. For example, if the maximum value of the cumulative score calculated in step S106 is not continuously improved a predetermined number of times, it is determined that the iterative process is to be terminated, and the translation optimization process routine is terminated. At this time, the weight of each translation feature model stored in the feature amount weight storage unit 22 is an optimized weight.

  On the other hand, if it is determined not to end the repetitive process, the process returns to step S102, the weight for each translation feature model is determined at random, and the processes of steps S103 to S107 are repeated.

  The weight of each translation feature model stored in the feature amount weight storage unit 22 is output and input to, for example, a machine translation device (not shown), and machine translation processing is performed using the weight of each translation feature model. .

<Experimental result>
Next, the results of experiments performed on 16 pieces of development data will be described. Experiments were conducted using development data of four types (EMEA: medical translation, Opensub: movie subtitle translation, Europarl: translation of Diet speech, NIST: translation of newspapers and blogs). There were six translation languages (DE: German, ES: Spanish, FR: French, NL: Dutch, SV: Sweden, ZH: Chinese), and the target language was English. Five optimization methods were compared including the optimization method described in the first embodiment.

  As comparison targets, the MERT optimization method using the BLEU evaluation scale score as an objective function, the MERT optimization method using the TER evaluation scale score as an objective function, and the BLEU evaluation scale score and the TER evaluation scale score as an objective function Experiments using the MERT optimization technique were conducted. In addition, the Pareto identification unit 25 similarly performed an experiment on a multi-objective optimization method using a skyline algorithm described later.

  The results of experiments on medical translation development data are shown in Table 1 below. As a result of the experiment, BLEU values and 1-TER values (higher scores are better) were determined for each optimization method.

  In Table 1 above, “COMBINED” indicates a case where the MERT optimization method using the BLEU evaluation scale score and the TER evaluation scale score as an objective function is used. “SKYLINE” indicates the case where a multi-objective optimization method using the skyline algorithm is used.

“CVXHULL” indicates a case where the multi-objective optimization method using the convex hull algorithm described in the first embodiment is used. “MERT (BLEU)” indicates a case where a MERT optimization method using an evaluation function score of BLEU as an objective function is used. “MERT (TER)” indicates a case where a MERT optimization method using an evaluation scale score of TER as an objective function is used.

  Table 2 below shows the results of experiments conducted on development data for movie subtitle translation.

  Table 3 below shows the results of experiments conducted on development data for translations of Diet speech and development data for translations of newspapers and blogs.

  In Tables 1 to 3, * indicates the best system. Bold (bold) indicates the best of the systems (Combined, Skyline, CvxHull) using both evaluation scales.

  From the above experimental results, when using a plurality of objective functions, most of the combinations of the source language and the evaluation scale are the methods proposed in the first embodiment (including the method using the skyline algorithm), It turned out that the optimization which was superior to "Combined" was realized. In particular, it is worthy of special mention that the method proposed in the first embodiment produces a better result than “MERT (TER)” with a 1-TER value. From the above experimental results, it was found that the optimization based on the Pareto optimality proposed in the first embodiment is an effective method.

  As described above, according to the translation optimization device according to the present embodiment, whether or not the Pareto optimality is satisfied by calculating a plurality of types of evaluation scale scores for each translation candidate in the output candidate list of the translation sentence. By updating the weights for each translation feature model according to the identification result, the weights for the plurality of translation feature models can be optimized using a plurality of evaluation scale scores simultaneously.

  By adjusting the feature weights of the translation feature model so that translation candidates that satisfy Pareto optimality (Pareto points) for multiple evaluation scales have a high evaluation score, a balance can be achieved from the viewpoint of multiple evaluation scales. A machine translation system can be constructed.

  In the above embodiment, each time a series of processes by the translation unit 23, the translation evaluation unit 24, the Pareto identification unit 25, and the optimization unit 26 are repeated, a weight for the translation feature model is randomly determined. Although the case has been described as an example, the present invention is not limited to this. Each time a series of processes by the translation evaluation unit 24, the Pareto identification unit 25, and the optimization unit 26 are repeatedly performed, the weight for the translation feature model is optimized in the optimization direction based on the calculated cumulative score, for example, using the POWELL method. May be updated.

<Second Embodiment>
Next, a second embodiment will be described. Note that the translation optimizing device according to the second embodiment has the same configuration as that of the first embodiment, and thus the same reference numerals are given and description thereof is omitted.

  In the second embodiment, an algorithm for optimizing the weight for each translation feature model is different from that in the first embodiment.

  The optimization unit 26 of the translation optimization apparatus according to the second embodiment uses a multi-objective optimization technique that extends MIRA. The classifier assumed here is one that binary classifies translation candidates into a positive class and a negative class. The multi-objective optimization method, which is an extension of MIRA, updates the feature weights of the classifier online, and operates according to the following procedure.

  First, for each of a plurality of development data, based on the identification result by the Pareto identification unit 25, translation candidates that do not satisfy the Pareto optimality are classified into a negative class, and the translation candidates (Pareto) included in the Pareto frontier Translation candidates that satisfy the optimality are classified into positive classes. If each translation candidate included in the translation candidate list of each development data is positive, the weight is not updated. On the other hand, when there is a negative translation candidate, the weight of each translation feature model is updated. The weight update method may be a simple gradient method or a complicated method (for example, margin-infused update).

  The translation optimization apparatus 100 repeatedly performs a series of processes by the translation unit 23, the translation evaluation unit 24, the Pareto identification unit 25, and the optimization unit 26. A series of processing is repeated until the weight vector is not updated.

  In the translation optimization processing routine according to the second embodiment, first, all of a plurality of development data is acquired from the development data storage unit 21. Then, a weight for each translation feature model is randomly determined and stored in the feature amount weight storage unit 22. The input sentence of each development data acquired above is subjected to translation processing, and an output candidate list consisting of translation candidates with the top k translation scores obtained by weighted addition of feature amounts to the translation feature model is displayed for each development data. Generate about.

  Next, for each development data, a plurality of types of translation evaluation scale scores are calculated for each translation candidate included in the output candidate list generated above for the development data.

  For each development data, whether or not each translation candidate included in the translation candidate list satisfies the Pareto optimality based on the plurality of types of translation evaluation scale scores calculated above for each translation candidate of the development data And identify the Pareto frontier for each development data.

  Next, for each development data, classify each translation candidate class based on the Pareto frontier obtained above, and if there is a negative translation candidate, assign the optimal weight to each translation feature model Turn into. And the weight of each translation feature model memorize | stored in the feature-value weight memory | storage part 22 is updated to the weight with respect to each translation feature model optimized above.

  Then, it is determined whether or not to end the repetition process. For example, when the weight is not updated, it is determined that the iterative process is finished, and the translation optimization process routine is finished. At this time, the weight of each translation feature model stored in the feature amount weight storage unit 22 is an optimized weight.

  On the other hand, if it is determined not to end the repetitive process, the above process is repeated using the weight of each translation feature model updated as described above.

  Note that other configurations and operations of the translation optimization apparatus according to the second embodiment are the same as those of the first embodiment, and thus the description thereof is omitted.

  In this way, by using the identification result of whether or not the Pareto optimality identified for each translation candidate in the translated sentence output candidate list is used, a weight for each translation feature model is obtained by a multi-objective optimization method that extends MIRA. By updating, weights for a plurality of translation feature models can be optimized using a plurality of evaluation scale scores simultaneously.

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

  For example, the Pareto identification unit may obtain the Pareto frontier using a skyline algorithm. The skyline algorithm may be the same as that described in non-patent literature (Parke Godfrey, Ryan Shipley, Jarek Gyrz, “Algorithms and analyzes for maximal vector computation”, VLDB Journal, vol16, 2006). Therefore, detailed description is omitted.

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 10 Input part 20 Calculation part 21 Development data storage part 22 Feature-value weight storage part 23 Translation part 24 Translation evaluation part 25 Pareto identification part 26 Optimization part 100 Translation optimization apparatus

Claims (7)

A translation processing means for translating a word string in the source language into a plurality of candidates h in a word string in the target language, based on a plurality of types of translation feature models and weights for each of the plurality of types of translation feature models;
Translation for calculating a plurality of scores M ₁ (h) and M ₂ (h) indicating a plurality of types of translation evaluation scales for each of a plurality of candidates h of the word string in the target language translated by the translation processing means An evaluation means;
Based on the plurality of scores M ₁ (h) and M ₂ (h) calculated for each of the plurality of candidates h , the variable α is gradually increased from 0 to infinity using the following equation: For each value of the variable α, candidates h ^* satisfying Pareto optimality are extracted from the plurality of candidate sets L, and candidates whose score M ₂ (h) is smaller than the extracted candidates h ^* are selected. Pareto identifying means for identifying as a candidate that does not satisfy the Pareto optimality ;
Weight updating means for updating weights for each of the plurality of types of translation feature models according to the identification result by the Pareto identification means;
Optimization for optimizing the weight for each of the plurality of types of translation feature models by repeating the translation by the translation processing unit, the calculation by the translation evaluation unit, the identification by the Pareto identification unit, and the update by the weight update unit Means,
Translation optimization device including

The translation processing means translates each of the plurality of word strings in the translation source language into a plurality of candidates for the word string in the translation destination language,
The weight updating means, when each candidate for a plurality of word strings of the translation destination language is identified as satisfying Pareto optimality according to the identification result by the Pareto identification means, Given the maximum value of the score, for each candidate for the plurality of word strings, calculate a cumulative score by accumulating the total value of each score, and based on the calculated cumulative score, the plurality of types of translation feature models claim 1 Symbol placement translation optimization apparatus updates the weighting for each. The weight updating means determines the plurality of types of translations when at least one of the candidates for the translation target language word string is determined not to satisfy Pareto optimality according to the identification result by the Pareto identification means. claim 1 Symbol placement translation optimization apparatus updates the weights for each of the feature model. A translation optimization method in a translation optimization device including a translation processing means, a translation evaluation means, a Pareto identification means, a weight update means, and an optimization means,
The translation optimization device includes:
The translation processing means translates a word string in the source language into a plurality of candidate words h in the target language based on the weights for each of the plurality of types of translation feature models and the plurality of types of translation feature models. Steps,
A plurality of scores M ₁ (h) and M ₂ ( M ₂ ) indicating a plurality of types of translation evaluation scales for each of a plurality of candidates h of the word string in the target language translated by the translation processing unit by the translation evaluation unit. h) calculating;
Based on the plurality of scores M ₁ (h) and M ₂ (h) calculated for each of the plurality of candidates h by the Pareto identification unit , the variable α is changed from 0 to infinity using the following equation: while gradually increasing until, at each value of the variable alpha, the set L of the plurality of candidate, extracts the candidate h ^* satisfying Pareto optimality, the more the extracted candidate h ^* score M _{2 (h} ) For which the candidate is smaller as a candidate that does not satisfy the Pareto optimality ;
Updating the weight for each of the plurality of types of translation feature models by the weight updating means according to the identification result by the Pareto identifying means;
The optimization unit repeats the translation by the translation processing unit, the calculation by the translation evaluation unit, the identification by the Pareto identification unit, and the update by the weight update unit, whereby weights for each of the plurality of types of translation feature models are obtained. Steps to optimize
A translation optimizing method comprising:

The step of translating by the translation processing means translates each of the plurality of word strings in the source language into a plurality of candidates for the word strings in the target language,
The step of updating by the weight updating means is performed when each candidate for a plurality of word strings in the translation destination language is identified as satisfying Pareto optimality according to the identification result by the Pareto identifying means. For each candidate for the plurality of word strings, a cumulative score obtained by accumulating the total value of the scores is calculated as each score for the plurality of word strings, and the plurality of scores are calculated based on the calculated cumulative score. The translation optimization method according to claim 4 , wherein the weight for each of the types of translation feature models is updated. The step of updating by the weight updating means is performed when at least one of the candidates for the word string of the translation target language is identified as not satisfying the Pareto optimality according to the identification result by the Pareto identifying means. 4. Symbol placement translation optimization method to update the weight for each of the plurality of types of translation feature model. The program for functioning a computer as each means of the translation optimization apparatus of any one of Claims 1-3 .

Images