JP2007072927A - Translation apparatus and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a translation apparatus for reducing calculation costs and using a greedy decoding algorithm. <P>SOLUTION: The translating apparatus 30 for translating a first language sentence 46 into a second language sentence 72 includes: MT engines 52 for translating the first language sentence 46 into hypotheses 56 in the second language; a classifier 58 and a hypotheses selection part 74 for selecting, on a statistical basis, a hypothesis from the hypotheses 56 that will yield a better hypothesis in a predetermined algorithm for generating hypothesis from an initial hypothesis; and a decoder 68 for generating the translation sentence 72 using the predetermined algorithm from the classifier 58, the hypothesis selection part 74, and the hypothesis selected by classifier 58 and a hypothesis selection part 74. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to machine translation, and more particularly, to statistical machine translation (SMT) using a greedy decoding algorithm.

[Greedy decoding for statistical machine translation]
The Statistical Machine Translation (SMT) framework formulates the problem of translating source language S sentences into target language T as a conditional probability maximization problem as follows.

ただしｐ（Ｔ）は「言語モデル」（ｌａｎｇｕａｇｅ ｍｏｄｅｌ：ＬＭ）と呼ばれ、ターゲット言語の尤度を表す。ｐ（Ｓ｜Ｔ）は「翻訳モデル」（ｔｒａｎｓｌａｔｉｏｎ ｍｏｄｅｌ：ＴＭ）と呼ばれ、ＳからＴへの生成確率を表す。言語モデルと翻訳モデルとの詳細については、非特許文献１を参照されたい。

However, p (T) is called a “language model” (LM) and represents the likelihood of the target language. p (S | T) is called a “translation model” (TM) and represents the generation probability from S to T. Refer to Non-Patent Document 1 for details of the language model and the translation model.

  The probabilities of LM and TM are automatically trained (learned) from a monolingual corpus and a bilingual corpus, respectively. These represent general translation knowledge used to map source language word sequences to target languages.

  An example of mapping an English (E) word sequence to Japanese (J) is shown in FIG.

  In the conventional SMT approach, all of the generated translation sentences are scored based on the probabilities of the translation model (fertility), NULL generation, lexicon, distortion, and language model. . The translation sentence candidate having the highest TM * LM score is selected as the translation output.

翻訳処理（デコーディング）の間には、適切な翻訳文を見出すために、起こりうる多数の単語の挿入、ソース言語とターゲット言語との単語シーケンス間の複雑な単語のアライメント（ファーティリティ）、及び単語の順序の変更（ディストーション）を考慮しなければならない。
Ｐ．ブラウン及びＳ．デラ ピエトラ及びＶ．デラ ピエトラ及びＲ．メルサー、「統計的機械翻訳の数学：パラメータの推定」、コンピュータ言語、１９（２）、２６３−３１１ページ、１９９３年（P. Brown and S. Della Pietra and V. Della Pietra and R. Mercer: "The mathematics of statistical machine translation: Parameter estimation", Computational Linguistics, 19(2), pp. 263-311, 1993） Ｍ．ポール及びＥ．スミタ及びＳ．ヤマモト、「統計的機械翻訳出力の用例ベースの再スコアリング」、HLT/NAACL2004予稿集、姉妹編、９−１２ページ、ボストン、USA、２００４年（M. Paul and E. Sumita and S. Yamamoto: "Example-based Rescoring of Statistical Machine Translation Output", In Proc. of the HLT/NAACL2004, Companion Volume, pp. 9-12, Boston, USA, 2004） ルールクエスト「データマイニングツールＣ５．０」、http://rulequest.com/see5-info.html （Rulequest "Data mining tool C5.0", http://rulequest.com/see5-info.html） Ｔ．ワタナベ及びＥ．スミタ、「統計的機械翻訳の用例ベースのデコーディング」、MTサミットIX予稿集、４１０−４１７ページ、２００３年（T. Watanabe and E. Sumita: "Example-based Decoding for Statistical Machine Translation", Proceedings of MT Summit IX, pp. 410-417, 2003.）

During the translation process (decoding), in order to find an appropriate translation, many possible word insertions, complex word alignments between the source and target language word sequences (fertility), and Changes in word order (distortion) must be taken into account.
P. Brown and S. Della Pietra and V. Della Pietra and R. Mercer, "Mathematics of Statistical Machine Translation: Parameter Estimation," Computer Language, 19 (2), pp. 263-311, 1993 (P. Brown and S. Della Pietra and V. Della Pietra and R. Mercer: " The mathematics of statistical machine translation: Parameter estimation ", Computational Linguistics, 19 (2), pp. 263-311, 1993) M.M. Paul and E. Sumita and S. Yamamoto, "Example-based rescoring of statistical machine translation output," HLT / NAACL2004 proceedings, sisters, pages 9-12, Boston, USA, 2004 (M. Paul and E. Sumita and S. Yamamoto: "Example-based Rescoring of Statistical Machine Translation Output", In Proc. Of the HLT / NAACL2004, Companion Volume, pp. 9-12, Boston, USA, 2004) Rule quest "Data mining tool C5.0", http://rulequest.com/see5-info.html (Rulequest "Data mining tool C5.0", http://rulequest.com/see5-info.html) T.A. Watanabe and E. Sumita, “Example-Based Decoding for Statistical Machine Translation”, MT Summit IX Proceedings, pages 410-417, 2003 (T. Watanabe and E. Sumita: “Example-based Decoding for Statistical Machine Translation”, Proceedings of MT Summit IX, pp. 410-417, 2003.)

  The translation output depends on the initial translation hypothesis to initiate the search, which will lead to a locally optimal translation, but not necessarily a global optimal translation.

  In order to solve this problem, the conventional measures have proposed using a sentence obtained by preliminary translation of input or a sentence similar to the input as an initial translation hypothesis (see Non-Patent Document 4). .)

  In this method, a translation example whose source part is similar to an input sentence is extracted from a bilingual corpus and used. The target part of each example is slightly modified so that each pair is an actual translation. The advantage of this example-based approach is that the search for good translations starts with the searched translation example rather than with the guessed translation, so the search error is small.

  However, since this method uses the same greedy search algorithm as the basic method, the search error cannot be completely avoided.

  In order to overcome the local optimization problem of greedy search and obtain the best translation, the method disclosed in Non-Patent Document 2 uses various starting points generated by a number of search engines. These translated sentences are called “initial translation hypotheses” or “seed”.

  Since initial translation hypotheses are generated by an independently developed translation engine using different dictionaries, different grammars, and different translation rules, a great variety of initial translation hypotheses can be obtained. Therefore, the greedy search algorithm can search a wider part of the search space, increasing the possibility of obtaining a global optimal solution.

  The best translation among the translations generated during the greedy search is selected by a rescoring method using edit distance. This corrects the statistical score of each of the generated translation candidates with information on how much the initial translation hypothesis was modified during decoding.

  This method exhibits performance superior to an approach based on greedy decoding based only on a statistical model, such as Non-Patent Document 4.

  However, the drawback of this approach is that the decoder must be applied to all initial translation hypotheses. For this reason, a high calculation cost is required to specify an optimal translated sentence.

  Accordingly, an object of the present invention is to provide a translation apparatus that uses a greedy decoding algorithm while reducing the calculation cost.

  Another object of the present invention is to provide a translation apparatus using a greedy decoding algorithm while maintaining the quality of the translation result and reducing the calculation cost.

  A translation device according to one aspect of the present invention is a translation device for translating a sentence in a first language into a sentence in a second language, wherein the sentence in the first language is converted into a plurality of sentences in the second language. A means for translating into a hypothesis, a means for selecting a plurality of hypotheses from which a better hypothesis is generated by a predetermined algorithm based on a statistical basis, and a predetermined algorithm And means for generating a translation from the hypothesis selected by the means for selecting.

  The means for translating translates the sentence in the first language into a plurality of hypotheses in the second language. The means for selecting selects a hypothesis that produces a better hypothesis when processed by a predetermined algorithm. The means for generating uses the selected hypothesis as an initial hypothesis, and generates a translation sentence by a predetermined algorithm. The result is expected to be better than the selected initial hypothesis because the preferred hypothesis is used in a predetermined algorithm by means for generating. Since the number of initial hypotheses used in the predetermined algorithm is substantially reduced, the computational cost of the means for generating is low.

  Preferably, the means for selecting classifies the plurality of hypotheses of the second language into a plurality of classes based on a statistical basis, and the respective scores of the plurality of hypotheses are based on the statistical basis. Including means for calculating. The plurality of classes includes a first class and a second class. The means for selecting further includes means for selecting one of the hypotheses classified into the first class based on the score calculated by the means for classifying.

  More preferably, the decision tree classifier wherein the means for classifying is trained by pairs of sentences in the first language and their corresponding hypotheses, and further results obtained from the hypotheses by a predetermined algorithm. including.

  More preferably, a set of training sentences in the first language, a set of training hypotheses obtained from the set of training sentences, and a second language obtained by a predetermined algorithm from the set of training hypotheses Means for training means for selecting using a set of sentences, each sentence being classified into a good translation sentence and a bad translation sentence.

  A computer program according to the second aspect of the present invention causes a computer to perform all the functions of any of the above translation apparatuses.

  The following system relates to translation from Japanese to English, but the invention is not limited to such language combinations. Any combination and any translation direction can be used as long as a bilingual corpus is available for that language combination.

<System overview>
FIG. 1 shows an overall block diagram of a translation system 30 according to the first embodiment of the present invention. The translation system 30 translates Japanese sentences into English sentences.

  The translation system 30 is a bilingual corpus 40 in which Japanese sentences and English sentences are pre-aligned, and an English-to-Japanese statistical translation model and English statistics derived from a training data set of the bilingual corpus 40. It works in conjunction with the global language model 70. Note that the translation direction of the translation model is the opposite of the translation direction of the translation system (from Japanese to English).

  Translation system 30 operates in two phases of operation. It is a training phase and an operation phase. The translation system 30 operates in the training phase or the operation phase according to the external operation mode signal 78.

  The translation system 30 includes a reading module 42 for reading the training sentence 44 from the bilingual corpus 40 in the training phase, an input selecting part 48 for selecting the training sentence 44 or the input sentence 46 according to the operation mode, and an input selecting part. And a set of MT engines 52 for translating the sentence selected by 48 and outputting a predetermined number of hypotheses 56 for each sentence from the input selector 48. The set of MT engines 52 includes m MT engines, and the MT engine 52 outputs m hypotheses for each sentence from the input selection unit 48.

  The translation system 30 further includes a selector 50 for selecting the training sentence 44 or the input sentence 46 according to the operation mode, and whether a given input sentence and hypothesis pair can be improved by an approach using greedy decoding. A decision tree (DT) classifier 58 to determine. The DT classifier 58 classifies hypotheses based on statistical evidence obtained during the training phase using training data from the bilingual corpus 40.

  The translation system 30 further includes a hypothesis selection unit 74 that selects one of the hypotheses based on the statistics and the reliability value of the classification result of the DT classifier 58, and two inputs that are the output of the hypothesis selection unit 74 and the hypothesis. 56, and the output selection unit 66 connected to each other to decode the output of the output selection unit 66 using the greedy algorithm shown in Non-Patent Document 2, and the best according to the rescoring method using the edit distance And a decoder 68 for selecting the translation result. The translation result is output as a decoder output 72.

  For the purpose of training the DT classifier 58, the translation system 30 further provides an automatic evaluation scoring scheme (“word error rate”, Word Error Rate (abbreviated for short) for each hypothesis pair in the hypothesis 56 and its corresponding decoder output 72. DT attribute extraction module 64 for assigning a “good” or “bad” attribute to the pair, and analyzing each input-hypothesis pair to produce a DT classifier 58. DT classification using the DT feature quantity extraction module 54 for extracting the feature quantity necessary for training, the feature quantity extracted by the DT feature quantity extraction module 54 and the attribute extracted by the DT attribute extraction module 64 DT classifier learning unit 62 for training the device 58 and the output of the DT feature quantity extraction module 54 according to the operation mode And a selection unit 60 for selectively connecting the output of the DT classifier learning unit 62 to the DT classifier 58.

  The translation system 30 further controls the input selectors 48, 50, 60 and 66, the DT classifier 58, and the reading module 42 to operate the translation system 30 in the training phase and the operation phase as follows. An operation phase control unit 76 is included.

I. Training Phase In the training phase, the operation phase controller 76 controls the selectors 48, 50, 60 and 66 to set the translation system 30 to train the DT classifier 58 as shown in FIG.

  In the training phase, the operation phase control unit 76 controls the translation system 30 as follows. The reading module 42 reads each of the Japanese training sentences in the bilingual corpus 40 and gives the training sentence 44 to the input selection unit 48 (FIG. 1). The input selection unit 48 selects the training sentence 44 and gives each of the training sentences to the MT engine 52. The MT engine 52 translates each of the training sentences 44 and outputs m hypotheses 56 for each of the training sentences 44.

  The selection unit 50 (FIG. 1) selects the training sentence 44 and provides the training sentence 44 to the DT feature amount extraction module 54. For each MT engine in the MT engine 52, each input-hypothesis pair pair is analyzed by the DT feature extraction module 54 to extract the features necessary to train the DT classifier 58, and this feature is This is given to the DT classifier learning unit 62.

  The output selection unit 66 (FIG. 1) selects the output (hypothesis 56) of the MT engine 52 and supplies each hypothesis to the decoder 68. The decoder 68 decodes a given hypothesis, that is, uses the given hypothesis as a seed to search for a translation using a greedy algorithm, selects the best translation according to a re-scoring scheme using edit distance, and gives A decoder output 72 is generated for each hypothesized.

  For each pair of hypothesis in hypothesis 56 and its corresponding output from decoder 68, DT attribute extraction module 64 applies an automatic evaluation scoring scheme to assign a “good” or “bad” attribute to the pair; The attribute is given to the DT classifier learning unit 62.

  The selector 60 (FIG. 1) connects the output of the DT classifier learning unit 62 to the training input of the DT classifier 58. The extracted feature quantity from the DT feature quantity extraction module 54 and the attribute from the DT attribute extraction module 64 are input to the DT classifier learning unit 62. As a result, the DT classifier 58 has a given input-hypothesis pair as greedy. It will be judged whether it can improve by the approach using decoding. The DT classifier 58 is a decision tree classifier, and a reliability measure is calculated for each leaf node of the DT classifier 58. The reliability measure indicates whether or not the possibility of judgment (“good” or “bad”) by the DT classifier 58 is correct in the form of statistical probability obtained in the training phase.

II. Operation Phase In the operation phase, the operation phase control unit 76 controls the selection units 48, 50, 60 and 66 so that the translation system 30 is set as shown in FIG. In this phase, a given input sentence 46 is translated using the MT engine 52 and m initial hypotheses 56 are generated. Applying a DT classifier to each of the input-hypothesis pairs, all input-hypothesis pairs classified as “bad” are deleted from the set of initial hypotheses 56.

  Based on the statistical score and the confidence value of the classification result, a single input-hypothesis pair is selected by the hypothesis selector 74 and used as input to the greedy decoder 68, where the resulting best translation is obtained. The (decoder output 72) is selected according to an edit distance based rescoring method.

  In this embodiment, five MT engines are used as the MT engine 52. Two were example-based MT (EBMT), which were trained by the same training set as the greedy decoder 68. The remaining three are commercially available rule-based MT (RBMT) systems, which are based on dictionaries, grammar and translation rules.

  A commercially available data mining tool was used to learn the decision tree classifier. An example of such a tool is in Non-Patent Document 3. The feature quantity extracted by the DT feature quantity extraction module 54 from the initial input-hypothesis pair set is as follows.

  (1) “Similarity features” between initial translation hypotheses generated by different MT engines, including:

-Number of identical initial translation hypotheses-Average edit distance (ED) between a given hypothesis and that of other MT engines. The edit distance is defined as the sum of the costs of insert (INS), delete (DEL), and replace (SUB) operations, ED = INS + DEL + SUB.

  -Sentence length differences from shortest / longest initial translation hypotheses.

  (2) “Statistical features” extracted from the training corpus. This includes:

TM * LM statistical model score-frequency of words in specific source and target languages.

  (3) “Language perplexity” feature quantity of source language input and initial translation hypothesis calculated based on trigram language model.

  (4) “Linguistic features”, which includes:

-Sentence length-Sentence type-Sentence structure (number of nodes in the parse tree, etc.)
-The size of the component-Density features, ie the ratio of function words to content words-Semantic categories of content words.

  FIG. 4 shows a table 90 as an example of a set of feature amounts extracted by the DT feature amount extraction unit 54 of this embodiment. In FIG. 4, “INPUT” means an input sentence, and “HYP” represents a hypothesis corresponding to “INPUT”.

  FIG. 5 shows a table 100 as an example of a set of feature amounts. Here, a pair extracted with respect to a pair of “INPUT” and one MT engine 52 (MT1) is shown as a set 110, and a pair extracted with respect to a pair of “INPUT” and another MT engine 52 (MT2). 112.

  The DT attribute (“good” or “bad”), that is, the feature quantity used for training of the classifier is determined as follows.

  Sentences in the bilingual corpus 40 are translated using the MT engine 52 to generate an initial hypothesis set 56. These initial hypothesis sets 56 are decoded by edit distance based rescoring. An automatic evaluation scoring method (WER) is applied to each of the outputs of the decoder 68 to determine the following translation quality.

(A) Initial hypothesis generated by MT engine (b) Decoder output.

  The attributes (“good” or “bad”) for each input-hypothesis pair are assigned as follows:

  -If the translation quality is improved, i.e. WER (decoder output) <WER (initial hypothesis), "good".

  -Otherwise it is "bad".

  The DT classifier 58 is trained by the feature quantities and attributes respectively extracted by the DT feature quantity extraction module 54 and the DT attribute extraction module 64 for each input-hypothesis pair.

Referring to FIG. 6, the trained DT classifier 58 includes nodes 140, 142, 144, 150, 152, 160, 162, 180, 182, 200, including questions about the input-hypothesis pair features. 210, 212, 230, 232, 240, 242, 260 and 262. Starting from the node route 140, the decision tree is traced according to whether or not the feature quantity is true, and the leaf including the classification result is reached. For example, the leaf node
A classified attribute (“good” or “bad”) and a numerical score; and a reliability of the classification result indicated by the tags 170, 190, 192, 220, 250, 270 and 272 (the same path of the decision tree) Can be assigned as a percentage of training samples correctly classified.

  For example, when the tag 170 traces the tree to reach the node 160, it indicates that the attribute is “bad” and the reliability value thereof is 1.0. This represents that all hypotheses leading to this leaf node 160 did not show any improvement in the training phase.

  In operation, a given input-hypothesis pair is classified by the DT classifier 58. If a given input-hypothesis pair is classified as “bad”, it is excluded from the initial hypothesis set.

  The hypothesis selection unit 74 converts a single input-hypothesis pair from the input-hypothesis pair selected by the DT classifier 58 into a score (TM · LM) based on a statistical model calculated using the translation and language model 70. The selection is made based on the information and the reliability value (CONF) of the decision tree classification result.

The hypothesis selection unit 74 rescores the hypothesis selected by the DT classifier 58 by the CONF · TM · LM score that is a function of CONF, TM, and LM.

CONF / TM / LM score = func (CONF, TM, LM)

The criteria used by the hypothesis selection unit 74 are as follows.

  -The higher the statistical model score TM · LM, the higher the quality of translation.

  -The higher the confidence score derived from the classification results, the better the starting point may have been found.

  The input-hypothesis pair with the highest CONF.TM.LM score is selected for decoding by decoder 68.

  In the above-described embodiment, the following scoring function is used.

ここで、「Ｉ」と「Ｈ」とは、それぞれ「入力」と「仮説」を表す。

Here, “I” and “H” represent “input” and “hypothesis”, respectively.

  In attribute extraction in the DT attribute extraction module 64, ranks according to the following criteria are assigned to each of the outputs of the decoder 68 in the training phase.

ランクは、英語を母国語とする人によって決定される。ＡＢＣスコアが次のように計算される。

The rank is determined by a person whose native language is English. The ABC score is calculated as follows:

例：トレーニングフェーズ（日本語から英語への翻訳）
トレーニングフェーズでのトレーニング文の一例「えと 名前 が タナカ ヨシコ と 申し ます それ で 予約 御 願い し ます」。この文をＭＴエンジン５２に与える。ＭＴ１及びＭＴ２がこの入力文（ＩＮＰＵＴ）を翻訳し、以下の二個の仮説ＨＹＰ１及びＨＹＰ２をそれぞれ生成する。

Example: Training phase (Translation from Japanese to English)
An example of the training text in the training phase, “My name is Yoshiko Tanaka, so please make a reservation.” This sentence is given to the MT engine 52. MT1 and MT2 translate this input sentence (INPUT) and generate the following two hypotheses HYP1 and HYP2, respectively.

これらの仮説に対するデコーダ６８の出力（ＯＵＴＰＵＴ）は次のようになる。

The output (OUTPUT) of the decoder 68 for these hypotheses is as follows.

ＤＴ特徴量抽出モジュール５４は各対（ＩＮＰＵＴ，ＨＹＰ１）、（ＩＮＰＵＴ，ＨＹＰ２）から特徴量を抽出し、一方ＤＴ属性抽出モジュール６４は対（ＨＹＰ１，ＯＵＴ１）、（ＨＹＰ２，ＯＵＴ２）の属性を抽出する。

The DT feature quantity extraction module 54 extracts feature quantities from each pair (INPUT, HYP1) and (INPUT, HYP2), while the DT attribute extraction module 64 extracts attributes of the pair (HYP1, OUT1) and (HYP2, OUT2). To do.

Example: Operation phase (Translation from Japanese to English)
In the operation phase, the following sentence is given to the translation system 30.

INPUT: “Yes, what cards do you have?”
The MT engine 52 outputs a hypothesis 56 shown in the following table together with each TM / LM score.

ＤＴ分類器５８は以下の表に示されるように仮説を分類した。この表からわかるように、仮説ＨＹＰ２及びＨＹＰ３は「不良」と分類されるため、削除される。仮説ＨＹＰ１、ＨＹＰ４及びＨＹＰ５のみが仮説選択部７４に与えられる。

The DT classifier 58 classified the hypotheses as shown in the following table. As can be seen from this table, hypotheses HYP2 and HYP3 are classified as “bad” and are therefore deleted. Only hypotheses HYP1, HYP4, and HYP5 are provided to the hypothesis selection unit 74.

先行技術では、全ての初期仮説にＴＭ・ＬＭスコアを用いて優先順位をつけた。しかし、全ての仮説がデコードのために選択される。

In the prior art, all initial hypotheses were prioritized using TM / LM scores. However, all hypotheses are selected for decoding.

この結果、デコーダ出力は以下の表に示されるようになり、翻訳文ＯＵＴ３の“ｄｏ ｙｏｕ ｈａｖｅ ｃａｒｄｓ ｄｏ ｙｏｕ ｈａｖｅ”が選択されることになる。

As a result, the decoder output is as shown in the following table, and “do you have cards do you have” in the translated text OUT3 is selected.

一方、この実施の形態では、「良」と分類された仮説のみがデコーディングのために選択され、ＣＯＮＦ・ＴＭ・ＬＭスコアを用いて優先順位がつけられる。とくに、この実施の形態では、最も高いＣＯＮＦ・ＴＭ・ＬＭスコアを有する仮説一つのみがデコーディングのために選択される。

On the other hand, in this embodiment, only hypotheses classified as “good” are selected for decoding and prioritized using CONF / TM / LM scores. In particular, in this embodiment, only one hypothesis having the highest CONF.TM.LM score is selected for decoding.

デコーダ６８の出力は、ＯＵＴ５の“ｙｅｓ ｗｈａｔ ｃａｒｄ ｄｏ ｙｏｕ ｈａｖｅ”となり、そのＴＭ・ＬＭスコアは１．０６６００１０１４８１ｅ−４６である。翻訳システム３０は“ｙｅｓ ｗｈａｔ ｃａｒｄ ｄｏ ｙｏｕ ｈａｖｅ”を出力する。

The output of the decoder 68 is “yes what card do you have” of OUT5, and its TM · LM score is 1.0666101481e-46. The translation system 30 outputs “yes what card do you have”.

  A subjective evaluation of the decoded outputs OUT1 to OUT5 was performed. The results are as follows.

この実施の形態の翻訳システム３０の出力は「Ａ」にランクされ、一方先行技術の方法の結果は「Ｄ」にランクされた。

The output of translation system 30 in this embodiment was ranked “A”, while the results of the prior art method were ranked “D”.

<Experimental result>
The inventors conducted experiments. The translation direction is from Japanese to English. The test data set contained 500 sentences.

(1) Accuracy of translation (a) MT engine

（ｂ）編集距離を用いた再スコアリングによる貪欲デコーディング

(B) Greedy decoding by rescoring using edit distance

（２）計算コスト

(2) Calculation cost

この実施の形態の翻訳システム３０は、先行技術の方法と比較して、計算コストを８５．７％減じつつ、より良好な翻訳精度を達成することが明らかである。

It is clear that the translation system 30 of this embodiment achieves better translation accuracy while reducing the computational cost by 85.7% compared to prior art methods.

  Although the hypothesis selection unit 74 of the above-described embodiment selects only one hypothesis, the present invention is not limited to such an embodiment. The hypothesis selection unit 74 may select two or more hypotheses as long as the number of hypotheses to be selected is smaller than the number of initial hypotheses 56. Compared to the embodiment described above, the computational cost is higher but still lower than the prior art.

  In the above-described embodiment, a decision tree is used to select a hypothesis for decoding. However, the selection of hypotheses may be done by other strategies as long as it can be learned to determine the hypotheses that would yield a better output of the greedy decoder. Instead of the decision tree, a support vector machine (Support Vector Machine: SVM), a neural network (NN), or a multi-layer perceptron (MLP) may be used.

  FIG. 7 shows the appearance of a computer system 330 that executes the above-described program and realizes the translation system 30 of this embodiment, and FIG. 8 shows the structure of the system 330 in a block diagram.

  Referring to FIG. 7, the computer system 330 includes a computer 340 having an FD (Flexible Disk) drive 352 and a CD-ROM (Compact Disc Read Only Memory) drive 350, a keyboard 346, a mouse 348, and a monitor 342. Including.

  Referring to FIG. 8, in addition to FD drive 352 and CD-ROM drive 350, computer 340 further includes a CPU (Central Processing Unit) 356 and a bus connected to CPU 356, CD-ROM drive 350, and FD drive 352. 366, a ROM (Read-Only Memory) 358 for storing a program such as a boot-up program, and a RAM (Random Access) that is connected to the CPU 356 and temporarily stores application program instructions and provides a temporary storage area. Memory) 360 and a hard disk 354 for storing application programs, system programs and data. Although not shown here, the computer 340 may further include a network adapter board that provides a connection to a local area network (LAN).

  A program for causing the computer system 330 to execute the functions of the translation system 30 of this embodiment is stored in the CD-ROM 362 or FD 364 inserted in the CD-ROM drive 350 or FD drive 352, and further transferred to the hard disk 354. Good. Alternatively, the program may be transmitted via a network (not shown) and stored in the hard disk 354. The program is loaded into the RAM 360 when executed. The program may be loaded directly into the RAM 360 from the CD-ROM 362, the FD 364, or the network.

  The program includes a plurality of instructions for causing the computer 340 to execute the functions of the translation system 30 of this embodiment. Since some of the required basic functions are provided by an operating system (OS) or a third party program running on the computer 340, or a module installed on the computer 340, the program is not necessarily of this embodiment. It is not necessary to include all of the basic functions required to implement the translation system 30. The program only needs to include a part of an instruction that calls an appropriate function in a controlled manner so as to obtain a desired result. The general operation of computer system 330 is well known and will not be repeated here.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each of the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

It is a block diagram of the translation system of one embodiment of this invention. 1 is a block diagram illustrating a translation system 30 configured to operate in a training phase. FIG. 1 is a block diagram illustrating a translation system 30 configured to operate in an operation phase. FIG. It is a figure explaining the feature-value extracted from the input-hypothesis pair of embodiment. It is a figure which shows the value of the feature-value of the input-hypothesis pair calculated by experiment. It is a figure which shows the detailed structure of the decision tree used by embodiment. It is a figure which shows the external appearance of the computer system 330 which executes the above-mentioned program and implement | achieves the system 30 of embodiment. It is a figure which shows the structure of the computer 340 shown in FIG. It is a figure which shows the example which maps the word sequence of English (E) to Japanese (J).

Explanation of symbols

30 translation system 40 bilingual corpus 52 MT engine 54 DT feature extraction module 56 hypothesis 58 DT classifier 62 DT classifier learning unit 64 DT attribute extraction module 68 decoder 70 translation and language model 72 decoder output 74 hypothesis selection module

Claims (5)

A translation device for translating a sentence in a first language into a sentence in a second language,
Means for translating the sentence in the first language into a plurality of hypotheses in the second language;
Means for selecting, based on statistical evidence, a plurality of hypotheses that produce a better hypothesis by a predetermined algorithm for generating a hypothesis from an initial hypothesis;
Means for generating a translated sentence from a hypothesis selected by the means for selecting using the predetermined algorithm. The means for selecting comprises:
Means for classifying the plurality of hypotheses of the second language into a plurality of classes based on a statistical basis, and calculating a score of each of the plurality of hypotheses based on the statistical basis; The plurality of classes includes a first class and a second class;
The means for selecting further comprises means for selecting one of hypotheses classified into the first class based on the score calculated by the means for classifying. The translation device described. A decision tree classifier trained by a pair of sentences of the first language and hypotheses corresponding to the first language, and a result obtained by the predetermined algorithm from the hypothesis; The translation apparatus according to claim 2, further comprising: A set of training sentences in the first language, a set of training hypotheses obtained from the set of training sentences, and the second language obtained by the predetermined algorithm from the set of training hypotheses And further comprising means for training the means for selecting using a set of sentences, each sentence being classified as a good translation sentence and a bad translation sentence. The translation apparatus according to claim 3. A computer program that, when executed on a computer, causes the computer to perform all of the functions according to any one of claims 1 to 4.

Images