JP2006338261A - Translation device, translation method and translation program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To correct deviation of a likelihood derived from each dictionary when performing machine translation by use of the plurality of translation dictionary to perform high-accuracy translation processing. <P>SOLUTION: In this translation device, an intrinsic expression extraction means 11 extracts intrinsic expression included in an input sentence to draw up an already intrinsic expression-extracted sentence, a phrase parallel translation candidate enumeration means 25 retrieves a phrase parallel translation candidate corresponding to each phrase constituting the already intrinsic expression-extracted sentence from a translation model group storage part 21 and records it in a parallel translation candidate table 24 together with a probability value, a dictionary-classified weighting means 26 refers to a weighting table 23 to weight the probability value derived from each translation dictionary corresponding to each phrase parallel translation candidate, and an optimum route search means 27 finds a combination wherein a product of each weighted probability value and the probability values of two continuing words in a combination acquired from a language model storage part 22 becomes maximum among combinations by the phrase parallel translation candidates recorded in the parallel translation candidate table 24, and outputs it as a translation result. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a technique for translating an input sentence described in a first language into a second language.

  A machine translation technique for causing a computer to perform a translation process has been a very difficult technique and has been studied for a long time. In recent years, various translation software have appeared.

  Conventional machine translation technologies generally have several dictionaries and rules that represent parallel translations and grammars, and generally generate translated sentences while referring to them. It takes many years to accumulate rules and dictionary data. Often required. In recent years, research that accumulates rules and dictionary knowledge from a large amount of parallel translation data at low cost by statistical methods, that is, research on statistical machine translation technology, has been actively conducted.

  The feature of this technology is that the translation process is applied to a noisy channel model. For example, in Japanese-English translation, the Japanese input sentence to be translated is transformed into English by a noisy channel. It is thought that it is a thing, and restoring this to the original English is translation processing, and the translation processing system is considered to be a decompressor (decoder). That is, the process of decoding from Japanese J to English E uses the prior probability P (E) of the English sentence and the conditional probability P (J | E) of the Japanese sentence when the English sentence is given, Finding an English sentence that maximizes P (E) * P (J | E). Here, P (E) is called a language model, and P (J | E) is called a translation model.

  Initially, translation models performed in units of words were the mainstream translation models, but in recent years, statistical machine translation techniques based on phrase translation, in which translation models are designed in units of phrases that connect one or more words, have become mainstream. is there.

  The advantage of this technique is that it is relatively strong even when the position of a phrase moves back and forth in a certain unit, and that translation can be performed in a semantically unit. One of the conventional techniques of statistical machine translation based on this phrase translation is called “Pharaoh” (see Non-Patent Document 1). In this specification, a translation apparatus based on a statistical machine translation technique based on phrase translation such as “Pharaoh” is proposed.

  By the way, there are many kinds of proper nouns such as personal names, place names, and organization names, and new expressions are generated one after another. For this reason, there are features that are likely to become unknown words, and they often do not exist in the translation model. In addition, if these proper nouns are made into short units, there is a high possibility that each unit exists in the translation model, but even if these translation models are combined, the translation is not always correct. For example, the correct translation of the proper noun “Nippon Telegraph and Telephone Corporation” is “Nippon Telegraph and Telephone Corporation”. .

  Also, numerical expressions such as date, amount, and time often have their own customs depending on the country, and since there are countless types of expressions, they do not exist in the translation model and often become unknown words. Therefore, it is desirable to prepare conversion rules rather than translating from a translation model. These proper nouns and numerical expressions are collectively referred to as proper expressions.

  If the input sentence to be translated contains a specific expression, the specific expression is likely to become an unknown word as described above, or it is a notation based on its own customs, so if you try to translate with a normal translation model Translation may fail and the meaning of the whole sentence may become strange.

  Therefore, in addition to the usual phrase bilingual dictionary, there are various types of translations such as bilingual unique expression dictionary, conversion rules for numerical expressions, new word dictionary for modern words, bilingual dictionary for basic phrases. There is a need for a technique for performing statistical machine translation based on phrase translation while preparing dictionaries and conversion rules.

In addition, the construction of a dictionary is a very expensive operation, and it is often the case in actual technical development that the construction is not carried out manually but by means that are automated to some extent. For example, in the technique described in Patent Document 1, it is possible to collect proper expression parallel translation data from a large amount of parallel translation text data as a set in the order of the probability in descending order of probability as automatic translation. This likelihood can be regarded as the translation probability P (E | J) in the translation model. Moreover, it is a dictionary to which a probability value is not given, and it is possible to obtain a probability as an appearance probability using the frequency in the dictionary. It is reasonable to consider that each probability value indicates a relative position in each dictionary, and there is no logical basis for comparing the probability values of different dictionaries. Therefore, when statistical translation is to be performed using a plurality of dictionaries, the question is how to use the probability value of each dictionary in the phrase translation decoder.
Japanese Patent Application Laid-Open No. 2004-326584 (Japanese Patent Application No. 2003-122360) “Parent Translation Specific Expression Extraction Apparatus and Method, Parallel Translation Specific Expression Extraction Program” “Pharaoh: a Beam Search Decoder for Phrase-Based Statistical Machine Translation Models”, [online], Computer Science and Artificial Intelligence Lab Massachusetts Institute of Technology, [Search May 19, 2005], Internet <URL: http: / /www.iccs.informatics.ed.ac.uk/~pkoehn/publications/pharaoh-amta2004-slides.pdf> “Minimum Error Rate Training in Statistical Machine Translation”, [online], July 8, 2003, ACL anthology, [Search May 25, 2005], Internet <URL: http: //acl.ldc.upenn .edu / P / P03 / P03-1021.pdf>

  Conventionally, if the input sentence to be translated contains a specific expression, the specific expression is often an unknown word that does not originally exist in the translation model, which requires a specific expression parallel translation dictionary or unique translation technology. There was a problem that translation often failed when processed in the same system as other parts of the sentence.

  In addition to specific expressions, I want to use various types of dictionaries and rules together to increase the vocabulary separately from basic bilingual dictionaries that contain so-called basic words and phrases, such as new word bilingual translations and specialized word bilingual translations. There is a request. However, if you try to use a lot of dictionaries and rules together, it is difficult to set a clear standard for which information to prioritize, and automatically change the weight of the probability value of each dictionary to optimize translation. It is important to get the probability.

  In the translation apparatus of the present invention, the input sentence is subjected to a specific expression extraction process to extract a specific expression portion. Then, with respect to the extracted proper expression, the proper noun is referred to as a proper expression parallel translation dictionary, and the numerical expression is referred to as a conversion rule to obtain a parallel translation supplement. In the translation process for the entire sentence, since there is a case where a parallel translation supplement is obtained from another dictionary, a weight of a probability value for each dictionary is given.

  By assigning optimal weights, it is possible to obtain translation results that are likely to be translated even if various dictionaries are used in combination. As a technique for assigning optimum weights, a minimum error learning method can be used.

  According to the present invention, it is possible to perform translation processing using various dictionaries and rules together under optimum conditions for specific expressions, numerical expressions, technical terms, new words, etc. that have been difficult to translate in the past. Improves translation accuracy of the entire sentence.

  FIG. 1 is a block diagram showing an example of an embodiment of a translation apparatus of the present invention, FIG. 2 is a flowchart of translation processing, and FIG. 3 is a flowchart of processing when calculating weights for each dictionary. An outline will be described.

  The translation apparatus of the present invention includes a specific expression extraction unit 10, a phrase translation decoder 20, and a weight calculation unit 30. As will be described later, the weight calculation unit 30 does not operate in normal translation processing and is used only when calculating the weight for each dictionary.

  The specific expression extraction unit 10 includes a specific expression extraction unit 11 that performs morphological analysis on an input sentence, extracts a specific expression included in the input sentence, and creates a unique expression extracted text.

  The phrase translation decoder 20 stores a translation model group memory that stores a plurality of translation dictionaries including at least a specific expression parallel translation dictionary that describes a large number of parallel translation phrases in the second language corresponding to the specific expression phrases in the first language together with their probability values. Unit 21, a language model storage unit 22 that stores a large number of appearance probability values of two consecutive words in the second language, a weighting table 23 that stores weights for each of the plurality of translation dictionaries, and a predetermined parallel translation candidate table 24, a phrase parallel translation candidate corresponding to each phrase constituting the sentence having the extracted unique expression is searched from the translation model group storage unit 21, and the phrase parallel translation candidate and its probability value are recorded in the parallel translation supplementary table 24. With reference to the complement listing means 25 and the weighting table 23, the probability values derived from the respective translation dictionaries corresponding to the phrase parallel translation candidates recorded in the parallel translation supplementary table 24 are weighted. Among the combinations of phrase-by-phrase translation candidates recorded in the book-by-book weighting means 26 and the parallel translation compensation table 24, the respective weighted probability values and the two consecutive in the combination acquired from the language model storage unit 22 And an optimum route searching means 27 for obtaining a combination having the maximum product with the word probability value and outputting the result as a translation result.

  The weight calculation unit 30 compares the translation result by the optimum route search means 27 (here, the translation result Nbest from the highest probability value to the top N translation results) with the correct translation text input separately, and the translation accuracy And a weight calculation / table update means 31 for updating the weighting table 23 by changing the weight for each translation dictionary so as to improve the translation accuracy using the minimum error learning method.

  In the translation processing of the present invention, as shown in FIG. 2, the specific expression extraction means 11 performs morphological analysis on the input sentence, extracts the specific expression included in the input sentence, and creates a specific expression extracted sentence (s1 ), The phrase parallel translation supplement enumeration means 25 searches the translation model group storage unit 21 for a phrase parallel translation candidate corresponding to each phrase constituting the named sentence extracted sentence, Phrase translation candidates and their probability values are recorded in the bilingual supplementary table 24 (s2), and the dictionary weighting means 26 refers to the weighting table 23 and corresponds to the phrase parallel translation candidates recorded in the bilingual supplementary table 24. Weights are assigned to the probability values derived from the translation dictionaries (s3), and the optimum route searching means 27 selects the respective weighted probability values among the combinations of phrase parallel translation candidates recorded in the parallel translation compensation table 24. , Language model The product of the probability values of the two consecutive words in the combination obtained from 憶部 22 is performed by outputting as a translation results for combination with the maximum (s4).

  Further, when calculating the weight for each dictionary, as shown in FIG. 3, the weight calculation / table update means 31 performs the correct translation input separately from the translation result Nbest after the same processing of s1 to s4 as described above. The translation accuracy is calculated by comparing with the sentence (s7), the weight for each translation dictionary is changed so as to improve the translation accuracy using the minimum error learning method (s8), and the weighting table is updated (s9). This is repeated until the order of the translation result Nbest does not change (s5, s6).

  Hereinafter, the details of the processing in each unit described above will be described.

<Specific expression extraction means 11>
As described above, the specific expression extraction unit 11 performs morphological analysis on the input sentence, extracts the specific expression included in the input sentence to create a specific expression extracted text, and is a predetermined text (indicated by a broken line in the figure). Temporarily store in the memory.

  In the specific expression extraction process, a specific expression included in the input sentence is recognized. For example, if the input sentence is "Nippon Nippon Telegraph and Telephone Corporation (NTT West) applied for a business plan on March 1," the sentence was divided into words (morphological analysis), and then "Nippon Nippon Telegraph and Telephone" Recognizes that “Stock Company” and “NTT West Japan” are the organization name <ORG> and “March 1” is the date <DAT>.

  For example, the technique described in Japanese Patent Application Laid-Open No. 2004-46775 or Japanese Patent Application No. 2004-373532 is used for the specific expression extraction processing here. In general, morpheme analysis and specific expression extraction include word division of an input sentence to give language information such as part of speech and reading. However, in the present invention, only word division information is used. All the language information such as reading and reading is omitted. Therefore, in the word division, the input sentence is divided into words, and in the specific expression extraction, what kind of specific expression each divided word takes is recognized, and a range indicating a specific specific expression is specified.

  For example, xml markup may be used as the range specification method. Specific specific expressions include, for example, an organization name <ORG>, a person name <PSN>, a place name <LOC>, and a date <DAT>. Arbitrary types are set in advance according to the specific expressions to be extracted.

  FIG. 4 shows an example in which the range of specific expressions is specified using xml markup. The xml markup is one of markup languages in which various information is embedded between characters “<” and “>” in characters written in text.

<Translation model group storage unit 21>
In the translation model group storage unit 21, a number of parallel translation phrases in the second language (here, English) corresponding to the proper expression phrases in the first language (here, Japanese) as shown in FIG. Described unique-expression bilingual dictionary, basic bilingual dictionary describing a large number of bilingual words and phrases in the second language corresponding to common words and phrases in the first language, together with their probability values, and new words in the first language Probability of conversion rules between the first and second languages related to numerical expressions such as a new word dictionary, date, amount, time, etc., describing a large number of parallel words and parallel phrases in the second language corresponding to and phrases together with their probability values A conversion table described together with the value is stored in advance.

<Language model storage unit 22>
The language model storage unit 22 stores in advance a large number of probability values in which any two words in the second language (here, English) appear consecutively.

<Weighting table 23>
In the weighting table 23, weights for each of a plurality of translation dictionaries stored in the translation model group storage unit 21 are stored in advance.

<Phrase translation supplementary enumeration means 25>
The phrase translation candidate enumeration unit 25 refers to a plurality of translation dictionaries stored in the translation model group storage unit 21 by using as input the sentence which has been extracted by the specific expression extraction unit 21 and temporarily stored in the memory. However, enumerate all possible parallel translations.

  At this time, generation of phrases on the Japanese side is assumed to sequentially connect one or more consecutive words. However, for a portion that has already been extracted as a specific expression, the range is first set as one phrase. For example, “Nippon Nippon Telegraph and Telephone Corporation” is used as a group, and this is used as a search key to refer to a specific expression parallel translation dictionary to obtain a parallel translation and a probability value at that time. If the corresponding bilingual data does not exist, the grouping as the specific expression is canceled and a new phrase is generated.

  In addition, the locations of “Application for business plan” are “Business,” “Business plan,” “Business plan,” “Application for business plan,” “Application for business plan,” “Plan,” “Plan,”… Are sequentially generated, and all the generated phrases are searched in the order of the basic bilingual dictionary, the new word dictionary, etc., and all corresponding bilingual translations are listed.

  The parallel translations and probability values listed as described above are recorded in the parallel translation compensation table 24. An example of a recording example in the parallel translation compensation table 24 is shown in FIG. At this time, the position of the phrase in the input sentence, that is, the start and end positions are made known. Here, the number of words starting from the beginning and ending with the number of words is recorded. For example, “NTT West Japan” is a phrase composed of the eighth, ninth, and tenth words and becomes “8,10”.

<Dictionary weighting means 26>
In the translation apparatus of the present invention, the probability value is obtained by referring to a plurality of dictionaries, conversion rules, and the like. Therefore, it is not possible to judge the high likelihood by simply comparing the probability values P (E | J) of individual dictionaries. Therefore, the weighting means for each dictionary 26 refers to the weighting table 23 and normalizes the probability values by weighting the probability values derived from the translation dictionaries corresponding to the phrase parallel translation candidates recorded in the parallel translation compensation table 24. .

  For example, Pk (E | J), Pne (E | J), Ps (E | J), Ph (E | J) are the probability values derived from the basic bilingual dictionary, proper expression bilingual dictionary, new word bilingual dictionary, and conversion rule, respectively. And λk · Pk (E | J), λne · Pne (E | J), λs · Ps (E | J), λh · Ph (E | J) ).

<Optimum route search means 27>
The optimum route searching means 27 refers to the parallel translation supplementary table 24 in which all parallel translation candidates are written, and acquires a probability value P (E | J) for each Japanese phrase and the parallel translation of the Japanese phrase. Further, referring to the language model storage unit 22, a probability value P (E) that is the appearance probability of two consecutive words (here, English words) in the parallel translation is acquired. Ultimately, a combination is obtained in which all Japanese words are translated once and P (E | J) * P (E) is maximized. However, as described above, the probability value P (E | J) is a probability value obtained by weighting the probability value derived from each dictionary, that is, P (E | J) = λk · Pk (E | J) + λne · Pne (E | J)
+ Λs · Ps (E | J) + λh · Ph (E | J)
Finally, for example, “Nippon Telegraph and Telephone West Corporation (NTT West) submitted it business operations plan 1 May.” Is output as a translation result.

<Weight calculation / table update means 31>
The weight calculation / table update means 31 calculates the weight for each dictionary in advance. In this processing means, the optimum weight is obtained using a certain data set by using each component of the specific expression extraction unit 10 and the phrase translation decoder 20 described above and the minimum error learning method (see Non-Patent Document 2). decide.

  The minimum error learning method is to obtain an optimum weight of each probability model when calculating a probability value composed of a plurality of probability models as implemented in this specification. Here, the plurality of probability models indicate the translation models (probability values) Pk (E | J), Pne (E | J), Ps (E | J), and Ph (E | J) for each dictionary. The weights of the respective probability models are λk, λne, λs, and λh.

  In this method, a data set for obtaining weights is prepared in advance. The data set here is an input sentence (here, a Japanese sentence) and a corresponding translated sentence (here, an English sentence). The initial values of the weights of the respective probability models are set as λk0, λne0, λs0, and λh0, and the input sentence Ji is translated by using the translation device of the present invention, that is, each part of the specific expression extraction unit 10 and the phrase translation decoder 20. . Here, the translation result is obtained by the optimum route search means 27 from the maximum probability value to the top N translation results, that is, Nbest translation results Ei1, Ei2,... EiN (shown by broken lines in the figure). Temporarily stored in a predetermined memory.

  Here, each translation result Nbest temporarily stored in the memory is compared with a separately input correct translation, and the translation accuracy is calculated. For example, BLEU generally used as the accuracy of machine translation is used. BLEU is a geometric average of the ratio of the word ngram in the translation result that matches the correct translated sentence for n = 1 to 4, and indicates that the closer to 1, the better the accuracy (continuous with the word ngram) N word strings.)

  BLEU in the translation result Nbest is calculated, the weighting table 23 is updated little by little from the initial values λk0, λne0, λs0, and λh0 so that candidates with a large BLEU value are placed at the top. The optimum route search process is repeated, and the process ends when the change in the rank in the translation result Nbest stops. The value of the weight at that time becomes the final value, and this value is used to perform actual translation processing, that is, translation processing without a correct translation sentence.

  With the above configuration, the phrase translation decoder 20 having a plurality of dictionaries can use the translation probabilities derived from the respective dictionaries by converting them into appropriate probability values by dictionary weighting, and utilize a plurality of dictionary resources. Therefore, more accurate translation processing can be realized. For this weighting, an optimum value is set by the weight calculator 30. The translation model group can have various configurations such as an existing basic dictionary, a new word dictionary, a separately created unique expression parallel translation dictionary, and numerical value conversion rules.

  In particular, in the present invention, since the specific expression can be extracted in advance by the specific expression extraction unit 10, the unique expression that is likely to become an unknown word is originally translated by using the specific expression parallel translation dictionary generated by the bilingual specific expression extraction device. Can be performed with high accuracy.

  The translation apparatus of the present invention can be realized by a circuit (hardware) corresponding to each component described above, but can also be realized by installing a program on a known computer via a recording medium or a communication line. Is done.

The block diagram which shows an example of embodiment of the translation apparatus of this invention Flow chart of translation processing in translation apparatus of the present invention Process flow for calculating dictionary-specific weights An explanatory diagram showing an example of an input sentence and a sentence with a specific expression extracted corresponding to it Explanatory drawing which shows an example of a proper expression parallel translation dictionary Explanatory drawing which shows an example of a translation candidate table

Explanation of symbols

  10: proper expression extraction unit, 11: specific expression extraction means, 20: phrase translation decoder, 21: translation model group storage unit, 22: language model storage unit, 23: weighting table, 24: parallel translation candidate table, 25: phrase parallel translation Candidate enumeration means, 26: weighting means by dictionary, 27: optimum route search means, 30: weight calculation section, 31: weight calculation / table update means.

Claims (6)

An apparatus for translating an input sentence written in a first language into a second language,
A translation model group storage unit that stores a plurality of translation dictionaries including at least a specific expression parallel translation dictionary that describes a number of parallel translation phrases of the second language corresponding to the specific expression phrases in the first language together with their probability values;
A language model storage unit that stores a large number of appearance probability values of two consecutive words in the second language;
A weighting table storing weights for each of the plurality of translation dictionaries;
A specific expression extraction unit that performs morphological analysis on the input sentence, extracts a specific expression included in the input sentence, and creates a specific expression extracted sentence;
Phrase parallel translation candidate corresponding to each phrase constituting the named sentence extracted from the language model storage unit, and the phrase parallel translation candidate and enumeration means for recording the phrase parallel translation candidate and its probability value in the parallel translation compensation table;
A dictionary-specific weighting unit that refers to the weighting table and weights the probability value derived from each translation dictionary corresponding to the phrase parallel translation candidate recorded in the parallel translation compensation table;
Among the combinations of phrase parallel translation candidates recorded in the parallel translation compensation table, the product of each weighted probability value and the probability value of two consecutive words in the combination acquired from the language model storage unit is the maximum. A translation apparatus comprising: an optimum route searching means for obtaining a combination of the following and outputting as a translation result. The translation accuracy is calculated by comparing the translation result obtained by the optimum route searching means and the correct translated sentence inputted separately, and the weight for each translation dictionary is changed to improve the translation accuracy by using the minimum error learning method. The translation apparatus according to claim 1, further comprising weight calculation / table updating means for updating the weighting table. A method of translating an input sentence written in a first language into a second language using a computer,
A computer that stores a plurality of translation dictionaries including at least a specific-expression parallel translation dictionary that describes a number of parallel-translation phrases in the second language corresponding to the specific-expression phrases in the first language together with their probability values; A language model storage unit that stores a large number of occurrence probability values of two consecutive words in the second language, and a weighting table that stores weights for each of the plurality of translation dictionaries,
The computer is
Morphological analysis of the input sentence, extracting a specific expression included in the input sentence to create a specific expression extracted sentence;
Searching for a phrase parallel translation candidate corresponding to each phrase constituting the named sentence extracted from the language model storage unit, and recording the phrase parallel translation candidate and its probability value in a parallel translation compensation table;
Referring to the weighting table, weighting the probability values derived from each translation dictionary corresponding to the phrase translation candidate recorded in the parallel translation compensation table;
Among the combinations of phrase parallel translation candidates recorded in the parallel translation compensation table, the product of each weighted probability value and the probability value of two consecutive words in the combination acquired from the language model storage unit is the maximum. And a step of obtaining a combination to be output as a translation result. The translation result is compared with the correct translated sentence input separately to calculate the translation accuracy. Using the minimum error learning method, the weight for each translation dictionary is changed to improve the translation accuracy, and the weighting table is updated. The translation method according to claim 3, further comprising the step of: A program that causes a computer to translate an input sentence written in a first language into a second language,
A translation model group storage unit that stores a plurality of translation dictionaries including at least a specific expression parallel translation dictionary that describes a number of parallel translation phrases of the second language corresponding to the specific expression phrases in the first language together with their probability values; A computer comprising a language model storage unit that stores a large number of appearance probability values of two consecutive words in a language, and a weighting table that stores weights for each of the plurality of translation dictionaries,
Morphological analysis of the input sentence, extracting a specific expression included in the input sentence to create a specific expression extracted sentence;
Searching for a phrase parallel translation candidate corresponding to each phrase constituting the named sentence extracted from the language model storage unit, and recording the phrase parallel translation candidate and its probability value in a parallel translation compensation table;
Referring to the weighting table, weighting the probability values derived from each translation dictionary corresponding to the phrase translation candidate recorded in the parallel translation compensation table;
Among the combinations of phrase parallel translation candidates recorded in the parallel translation compensation table, the product of each weighted probability value and the probability value of two consecutive words in the combination acquired from the language model storage unit is the maximum. A translation program characterized by causing a combination to be obtained and executing a step of outputting as a translation result. The translation result is compared with the correct translated sentence input separately to calculate the translation accuracy. Using the minimum error learning method, the weight for each translation dictionary is changed to improve the translation accuracy, and the weighting table is updated. The translation program according to claim 5, further comprising the step of:

Images