JP2010244385A - Machine translation device, machine translation method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To achieve machine translation without making it necessary to prepare an accurate translation rule to a grammar phenomenon which is difficult to describe or can not be fully described as a translation rule. <P>SOLUTION: A vocabulary having a grammar phenomenon which is difficult to describe or can not be fully described as a translation rule is registered as a specific pattern in a specific pattern DB103 in advance, and when the specific pattern is included in an input sentence in first language, the vocabulary or syntax information in second language corresponding to the analysis results of the specific pattern is acquired from a translation DB108, and the error of a second language translation sentence to be obtained by translating first language is detected by an error detection/proofreading part 107 by using the vocabulary or syntax information in the second language and statistical co-occurrence information stored in a second language statistical model 104, and the second language translation sentence is proofread. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a translation technique, and more particularly to a technique for machine translation of a first language into a second language using a translation dictionary or translation rules.

  Conventionally, in order to improve the translation accuracy of machine translation using a rule base, a technique for performing semi-automatic or fully automatic editing based on a pre-edit rule or a post-edit rule has been proposed in, for example, Patent Documents 1-3.

As shown in FIG. 15, the post-translational editing support device described in Patent Document 1 includes a display unit, a display control unit, a text storage unit, an input control unit, an input unit, and an automatic post-editing process. And an automatic post-editing rule storage unit.
In this patent document 1, as a post-translation editing support device, a sentence that is determined to be an automatic post-edit sentence based on a determination result of an automatic post-edit sentence determination unit and an automatic post-edit sentence using an automatic post-edit rule. The sentence that is not determined to be is distinguished from the sentence and displayed on the display unit.

In addition, as shown in FIG. 16, an automatic text pre-editing device for machine translation described in Patent Document 2 includes a text input unit, a text buffer for storing input text, a morpheme analysis unit, a dictionary, and pre-editing. A control group that detects a predetermined search pattern from a rule group stored in the pre-edit rule group, performs a predetermined process, and performs pre-edit processing in a form suitable for machine translation; a sentence buffer; And an output unit.
In Patent Document 2, an automatic pre-editing device is realized by applying a pre-edit rule group to a machine translation system.

The rule generation device for machine translation described in Patent Document 3 includes a first natural language expression as a parallel translation example, an input unit for inputting the second natural language expression as a parallel translation, and the input parallel translation example. A conversion rule identifying means for identifying a conversion rule for converting the first natural language into the second natural language, and an expression that matches the condition of the first natural language of the identified conversion rule from the first natural language corpus. A second natural language that generates a candidate for a second natural language expression corresponding to the first natural language by applying the identified natural rule to the first natural language expression; An expression generator, a second natural language expression verification means for verifying the validity of the generated second natural language expression using the second natural language corpus, and outputting a valid expression; and a first natural language expression extraction The first natural language corpus by means And extracted representation, and a rule addition unit for adding a set of representations of the second natural language which is output by the second natural language expressions verification means as a new translation rules.
In Patent Document 3, a translation rule is automatically acquired from a parallel translation example.

JP-A-7-28818 Japanese Patent Laid-Open No. 06-139274 Patent No. 003329371

Martin, S., Liermann, J. and Ney, H., 1998, "Algorithms for bigram and trigram word clustering", Speech Communication, 24 (1998), 19-37.

However, with such a conventional technique, the cost of creating an accurate translation rule for a grammatical phenomenon that is difficult or impossible to describe as a pre-edit rule, a post-edit rule, or a translation rule in a rule-based machine translation device. There is a problem of becoming higher.
The reason for this is that, according to the prior art, for the grammatical phenomenon that is difficult or impossible to describe as a translation rule, pre-editing rules, post-editing rules, and complete translation example data for creating translation rules are provided. This is because it is necessary to collect a lot.

In other words, for grammatical phenomena that are difficult or impossible to describe as translation rules, there are a large number of translation rules. The pre-edit rule, post-edit rule, and translation rule cannot be acquired.
However, it is expensive to collect examples in which correspondence between two languages can be comprehensively reflected on grammatical phenomena that are difficult to describe or cannot be described as translation rules.

  Using a Japanese-Chinese machine translation system as an example, Chinese constants generation processing for Japanese combination modification, Chinese character generation processing for Japanese combination modification, and generation from Japanese to Chinese complements In order to perform the processing, there are an infinite number of corresponding Chinese generation translation rules. In particular, in order to build a high-precision Japanese-Chinese translation system, when there are multiple Chinese constants, adjectives, and complements, the production rules related to the order of the order and the Chinese structure particles that coexist with each component In order to determine whether or not to generate “target”, “ground”, and “profit”, it is necessary to prepare a Chinese generation rule in detail after conducting a stricter grammatical analysis. However, in order to create these rules, it is necessary to create an enormous number of conversion rules according to complicated grammatical phenomena such as manually abstracting grammatical phenomena and creating translation rules, which is expensive.

  The present invention is for solving such problems, and machine translation that can perform machine translation without requiring an accurate translation rule for a grammatical phenomenon that is difficult or impossible to describe as a translation rule. To provide technology.

  In order to achieve such an object, a machine translation apparatus according to the present invention includes a morpheme / syntax analyzer that performs morpheme / syntax analysis on an input sentence expressed in a first language, and a second language as a first language. A translation database composed of a translation dictionary or translation rule used for translation into a language, the analysis result of the morpheme / syntactic analysis unit, and the translation database are referred to, and a second language translation sentence is generated as a translation result corresponding to the input sentence. The specific pattern stored in the specific pattern database is detected from the analysis result of the second language generation unit, the specific pattern database storing the specific vocabulary used in the first language as the specific pattern, and the morpheme / syntactic analysis unit A specific pattern detection unit that acquires vocabulary or syntax information of the second language corresponding to the obtained analysis result of the specific pattern from the translation database; A second language statistical model that stores statistical co-occurrence information related to the co-occurrence of vocabulary used in a word, and a second language vocabulary or syntax information obtained by a specific pattern detection unit and a second language statistical model An error detection / calibration unit that detects an error in the second language translation sentence generated by the second language generation unit using the statistical co-occurrence information that has been generated, and calibrates the second language translation sentence. ing.

  A machine translation method according to the present invention is a machine translation method used in a machine translation device that translates a first language into a second language, and the morpheme / syntax analyzer is an input sentence expressed in the first language. A morpheme / syntax analysis step for performing morpheme / syntax analysis on a gramme, a result of analysis by a morpheme / syntax analysis unit, and a translation dictionary or translation used for translating the first language into the second language A second language generation step for generating a second language translation sentence as a translation result corresponding to the input sentence with reference to a translation database composed of rules, and a specific pattern detection unit from the analysis result of the morpheme / syntax analysis unit A specific language stored in a specific pattern database that stores a specific vocabulary used in one language as a specific pattern, and a second language word corresponding to the obtained analysis result of the specific pattern Alternatively, the specific pattern detection step of acquiring syntax information from the translation database, and the error detection / proofreading unit co-occurrence of the vocabulary or syntax information of the second language obtained by the specific pattern detection unit and the vocabulary used in the second language And using the statistical co-occurrence information stored in the second language statistical model for storing the statistical co-occurrence information regarding the second language translation unit to detect an error in the second language translation sentence, An error detection / proofreading step for proofreading the second language translation sentence.

  Further, the program according to the present invention allows a morpheme / syntax analyzer to perform a morpheme / syntax analysis on an input sentence expressed in a first language on a computer of a machine translation device that translates a first language into a second language. The morpheme / syntax analysis step to be performed, the second language generation unit refers to the analysis result of the morpheme / syntax analysis unit, and a translation database comprising translation dictionaries or translation rules used to translate the first language into the second language And a second language generation step for generating a second language translation sentence as a translation result corresponding to the input sentence, and a specific pattern detection unit using a specific vocabulary used in the first language based on an analysis result of the morpheme / syntax analysis unit. The specific pattern stored in the specific pattern database is stored as a specific pattern, and the lexical or syntax information of the second language corresponding to the obtained analysis result of the specific pattern is obtained. The specific pattern detection step acquired from the translation database, and the error detection / proofreading unit perform statistical co-occurrence regarding the co-occurrence of the vocabulary or syntax information of the second language obtained by the specific pattern detection unit and the vocabulary used in the second language. Using the statistical co-occurrence information stored in the second language statistical model for storing the origin information, detecting an error in the second language translation generated by the second language generation unit, and detecting the second language An error detection / proofreading step for proofreading the translation is executed.

According to the present invention, in a rule-based machine translation device from a first language to a second language, machine translation without requiring accurate translation rules for grammatical phenomena that are difficult or impossible to describe as translation rules. It can be performed.
This makes it possible to reduce the cost required to create a translation rule for a grammatical phenomenon that is difficult to describe or cannot be described as a translation rule.

It is a block diagram which shows the structure of the machine translation apparatus concerning the 1st Embodiment of this invention. It is explanatory drawing which shows the example of Japanese syntax analysis. It is explanatory drawing which shows specific pattern DB used with the machine translation apparatus concerning the 1st Embodiment of this invention. It is a flowchart which shows the machine translation process concerning the 1st Embodiment of this invention. It is a block diagram which shows the structure of the machine translation apparatus concerning the 2nd Embodiment of this invention. It is explanatory drawing which shows specific pattern DB used with the machine translation apparatus concerning the 2nd Embodiment of this invention. It is explanatory drawing which shows the structural example of specific pattern DB used with the machine translation apparatus concerning the 2nd Embodiment of this invention. It is a flowchart which shows the machine translation process concerning the 2nd Embodiment of this invention. It is a block diagram which shows the structure of the machine translation apparatus concerning the 3rd Embodiment of this invention. It is a block diagram which shows the structure of the machine translation apparatus concerning the 4th Embodiment of this invention. It is a block diagram which shows the structure of the machine translation apparatus concerning 1st Example of this invention. It is a flowchart which shows the machine translation process concerning 1st Example of this invention. It is a block diagram which shows the structure of the machine translation apparatus concerning the 2nd Example of this invention. It is a flowchart which shows the machine translation process concerning the 2nd Example of this invention. It is a block diagram which shows the structure of the conventional post-machine-translational editing support apparatus. It is a block diagram which shows the structure of the conventional machine translation automatic pre-editing system.

Next, embodiments of the present invention will be described with reference to the drawings.
[First Embodiment]
First, a machine translation apparatus according to a first embodiment of the present invention will be described with reference to FIG. FIG. 1 is a block diagram showing the configuration of the machine translation apparatus according to the first embodiment of the present invention.

The machine translation apparatus 100 includes an information processing apparatus such as a computer as a whole, and has a function of machine-translating an input first language into a second language.
The machine translation apparatus 100 includes, as main functional units, a first language input unit 101, a morpheme / syntax analysis unit 102, a specific pattern database (hereinafter referred to as a specific pattern DB) 103, a second language statistical model 104, a specific A pattern detection unit 105, a second language generation unit 106, an error detection / calibration unit 107, a translation database (hereinafter referred to as translation DB) 108, and a second language output unit 109 are provided.

  The first language input unit 101 includes an operation input device such as a keyboard and an input interface circuit that acquires data from an external device (not shown), and has a function of inputting an input sentence of a first language to be translated. ing.

The morpheme / syntax analysis unit 102 has a function of performing morpheme analysis processing and syntax analysis processing on the input sentence of the first language input by the first language input unit 101.
FIG. 2 is an explanatory diagram showing an example of Japanese syntax analysis. Here, a Japanese parsing example is shown for the Japanese input sentence "The nearest restaurant parking lot is full". Japanese input sentences are broken down into vocabulary of morphemes, and the parts of speech and combinations of these vocabularies are shown.

  The specific pattern DB 103 is a database in which a specific pattern indicating a specific vocabulary used in the first language is described and stored as information on the first language side given as an analysis result of the morpheme / syntactic analysis unit 102, Are stored in a storage unit including a storage device.

  The second language statistical model 104 is a statistical model that stores statistical co-occurrence information related to vocabulary co-occurrence used in the second language, and is stored in the storage unit. For example, in the second language monolingual corpus, it can be constructed using statistical methods such as N-gram, decision tree, SVM (Support Vector Machine), maximum entropy, HMM (Hidden Markov Model), and Bayesian learning. . Of course, it is not limited to these statistical methods. The second language statistical model 104 includes, as statistical co-occurrence information, a notation, a prototype, a usage form, a part of speech, a case frame, a tense, a state, a phase, a semantic classification, or a vocabulary or syntax information of the second language. Statistical co-occurrence information regarding any one or more of the co-occurrence patterns having the dependency relationship is stored.

The specific pattern detection unit 105 corresponds to the function of detecting the specific pattern stored in the specific pattern DB 103 from the analysis result obtained by the morpheme / syntax analysis unit 102 and the analysis result of the specific pattern obtained by this search. The second language vocabulary or syntax information is acquired from the translation DB 108.
The second language generation unit 106 refers to the analysis result obtained by the morpheme / syntax analysis unit 102 and the translation DB 108, translates the input sentence of the first language into the second language, and obtains the first translation result as the obtained translation result. It has a function to generate bilingual translated sentences.

  The error detection / calibration unit 107 uses the second language vocabulary or syntax information obtained by the specific pattern detection unit 105 and the statistical co-occurrence information stored in the second language statistical model 104 to generate a second It has a function of detecting an error in the second language translation sentence generated by the language generation unit 106 and a function of automatically calibrating the error of the second language translation sentence. As the error detection processing in the error detection / calibration unit 107, using the second language statistical model, detection of unnecessary components, detection of missing components from the second language translation sentence generated by the second language generation unit, Alternatively, any one of the error detections in the word order is detected.

Further, the error detection / calibration unit 107 corresponds to the second language statistical model from the second language translation sentence based on the learning features and feature quantities of the statistical model registered in the second language statistical model 104. If components such as learning features and feature quantities are extracted and shaped into a format that can be applied to the model, the optimal solution can be estimated by applying it to the model.
For example, a second language constructed by weighting vocabulary notation information, parts of speech, semantic classification, patterns having dependency relationships, etc. as learning features and feature quantities, with co-occurrence information quantities such as frequency values, probability values, and information quantities When the statistical model 104 is used to perform processing in the error detection / calibration unit 107, error detection / calibration processing is performed from the second language translation sentence using a feature extraction method and algorithm suitable for the learning feature and feature amount of the model. Just do it.

  Further, when an N-gram language statistical model or a Class N-gram model constructed with the second language monolingual corpus is used as the second language statistical model 104, the second language translation sentence is used in accordance with a specific pattern. As an example of the simplest calculation method for detecting whether there is an unnecessary component or missing component, or detecting whether there is an error in the word order of sentence components, the second language statistical model In 104, the one that maximizes the global statistical information amount of the generated sentence composed of a combination of the second language translation sentence and the vocabulary and syntax information of the second language acquired by the specific pattern detection unit 105 is optimized. It can be a solution. Further, when calculating the sentence occurrence probability, it may be calculated using the co-occurrence information of the vocabulary of the vocabulary of the second language generation sentence or the co-occurrence information of the part of speech class. In particular, when using an N-gram, a forward-facing N-gram model and a backward-facing N-gram model may be used simultaneously.

  Alternatively, when a model constructed using SVM with positive or negative learning data is used as the second language statistical model 104, the error detection / calibration unit 107 is used when constructing the SVM model. Based on the learning feature, the sentence that matches the learning feature is extracted from the second language translation sentence and shaped in a certain format, and the SVM model analyzes the unnecessary sentence components and missing parts of the specific target from the second language translation sentence. A proofreading process can be performed by detecting an error in sentence or word order.

  The translation DB 108 includes a translation rule, a translation dictionary, or both a translation rule and a translation dictionary. The translation rule generally stores the analysis rules on the first language side used in the rule-based translation system, the generation rules on the second language side, and the corresponding rules on both sides of the first language and the second language. It is. A translation dictionary is generally a dictionary in which correspondence between a vocabulary of a first language and a vocabulary of a second language used in a rule-based translation system is obtained.

  The second language output unit 109 includes a screen display device such as an LCD or a PDP, or an output interface circuit that outputs data to an external device (not shown). When the specific pattern is not detected by the specific pattern detection unit 105 , The function of outputting the second language translation from the second language generator 106, and the error detection of the second language translation from the second language generator 106 when the specific pattern is detected from the specific pattern detector 105 A function of outputting a result calibrated by the calibration unit 107;

  FIG. 3 is an explanatory diagram showing a specific pattern DB used in the machine translation apparatus according to the first embodiment of the present invention. This example is assumed to be a daytime machine translation system, and shows an example in which a specific pattern database is constructed with information in only a single Japanese language. Here, a specific pattern consisting of part-of-speech information and notation such as a Japanese case particle “NO”, a formal noun “NO”, and an auxiliary verb “DA” is recorded.

For example, there are two Chinese translations corresponding to the case particle “no”, a translation that does not generate anything and a Chinese structure particle “target”, depending on the context. However, the clear Japanese and Chinese grammatical phenomena that can correctly translate and translate these Chinese translations are extremely complex. When creating machine translation rules, it is generally necessary to create strict translation rules. Have difficulty.
In addition, the formal noun “NO” and the auxiliary verb “DA” have similar problems. There are two Chinese translations of the formal noun “NO”: “target” and “target east-west”. The Chinese translations of the auxiliary verb “Da” are “present”, “screaming”, “sore”, “present”, etc., and the translation of the Chinese translations of these vocabularies is extremely complicated.
Therefore, the rules for translation into Chinese corresponding to these vocabularies are complicated and cannot be generally created. Therefore, in the present invention, these rules are recorded in the specific pattern DB 103 and the specific pattern detection unit 105 performs the specific pattern detection. Is detected in the input sentence.

  Further, the specific pattern detection unit 105 detects whether or not the specific pattern stored in the specific pattern DB 103 is included from the analysis result of the first language input sentence obtained from the morpheme / syntax analysis unit 102. For a specific pattern, information on adjacent vocabulary contained in the first language sentence or syntax information having a dependency relationship with the detected specific pattern is acquired, and the vocabulary on the second language side of these information is acquired. Alternatively, a mechanism may be used in which the syntax information is acquired from a translation dictionary or translation rules and can be used by the error detection / calibration unit 107.

Here, the processing of the specific pattern detection unit 105 will be described with reference to FIG. 3. First, pattern matching is performed using the analysis result of the morpheme / syntax analysis unit 102 and the specific pattern DB 103 shown in FIG. Thus, it can be detected whether it is included in the Japanese input sentence of the Japanese-Chinese translation system.
Next, when the specific pattern shown in FIG. 3 is detected, the error detection / calibration unit 107 performs a dictionary lookup process for the daytime translation dictionary in the translation DB 108 using the part of speech and notation information of the detected specific pattern as search keys. Go to acquire Chinese translation candidates of the detected specific pattern.

Subsequently, the error detection / proofreading unit 107 sets the parallel translation candidate for error detection / proofreading processing for the Chinese translation obtained by the second language generation unit 106. Also, by acquiring vocabulary information that has a dependency relationship with a Japanese vocabulary of a specific pattern and using the corresponding vocabulary information on the Chinese side at the same time, the accuracy can be improved if it is used for error detection and proofreading processing. .
For example, the symbol “Φ” indicating “no translation” and the Chinese structure particle “target” can be acquired from the Japanese-Chinese translation dictionary using the Japanese case particle “no” part of speech and notation shown in FIG. 3 as search keys. With these pieces of information as the identification target on the Chinese generation side, the suitability of candidates can be determined from the Chinese translation using a Chinese statistical model.

  Among the functional units of the machine translation apparatus 100, the morpheme / syntax analyzing unit 102, the specific pattern detecting unit 105, the second language generating unit 106, and the error detecting / calibrating unit 107 may be realized by a dedicated information processing circuit. It may be realized by an arithmetic processing unit that has a CPU and its peripheral circuits, and implements various processing units by reading and executing a program from a storage unit (not shown).

[Operation of First Embodiment]
Next, the operation of the machine translation apparatus according to the first embodiment of the present invention will be described in detail with reference to FIGS. FIG. 4 is a flowchart showing machine translation processing according to the first embodiment.

  First, the first language input unit 101 inputs an input sentence of the first language to be translated (step S11), and the morpheme / syntax analysis unit 102 performs morphological analysis or word division on the input sentence. Processing is performed, syntax analysis of the input sentence is performed using the morpheme information, processing for obtaining a vocabulary having a dependency relationship between words included in the input sentence is performed, and the analysis result is stored (step S12). ).

  Subsequently, the specific pattern registered in the specific pattern DB 103 is detected by the specific pattern detection unit 105 from the morpheme / syntax analysis result obtained by the morpheme / syntax analysis unit 102 (step S13), and the specific pattern is detected from the input sentence. If NO is detected, the process proceeds to step S14.

On the other hand, when a specific pattern is detected from the input sentence in step S13, the specific pattern detection unit 105 uses the analysis result corresponding to the detected specific pattern to search the second language from the translation rule or translation dictionary of the machine translation system. , And the detected specific pattern and the vocabulary and syntax information of the second language are stored, and the process proceeds to step S14.
At this time, as described above, when a specific pattern is detected from the input sentence, the specific pattern detection unit 105 detects the vocabulary information adjacent to the specific pattern included in the first language sentence, or the detected specific pattern. Syntactic information having a dependency relationship with the pattern may be acquired, and the vocabulary and syntax information on the second language side corresponding to the information may be acquired from a translation dictionary or a translation rule.

Next, the second language generation unit 106 uses the result of the morpheme / syntax analysis obtained by the morpheme / syntax analysis unit 102 and the translation dictionary and translation rules for translation from the first language to the second language. The input sentence in the first language is translated into the second language, and a second language translation sentence that is the translation result is generated (step S14).
Thereafter, when the specific pattern is not detected in step S13, the process proceeds to step S16, and the second language translation sentence generated in step S14 is output by the second language output unit 109 (step S16). End the translation process.

  On the other hand, when the specific pattern is detected in step S13, the vocabulary and syntax information of the second language corresponding to the specific pattern stored in step S13 and the statistical co-occurrence information of the second language statistical model 104 are used. After the error detection / proofreading unit 107 detects an error in the second language translation sentence generated in step S14 and corrects this error (step S15), the second language output unit 109 corrects the second language. The translated text and the proofreading result are output (step S16), and the series of machine translation processing is terminated.

[Effect of the first embodiment]
As described above, in this embodiment, a vocabulary having a grammatical phenomenon that is difficult to describe or cannot be described as a translation rule is registered in advance in the specific pattern DB 103 as a specific pattern, and this specific pattern is an input sentence in the first language. The second language vocabulary or syntax information corresponding to the analysis result of the specific pattern is obtained from the translation DB 108, and the second language vocabulary or syntax information and the second language statistical model 104 are acquired. The error detection / calibration unit 107 detects an error in the second language translation sentence obtained by translating the first language using the statistical co-occurrence information stored in the second language translation sentence. Thus, machine translation can be performed without requiring accurate translation rules for grammatical phenomena that are difficult to describe or cannot be described as translation rules.

  This makes it possible to reduce the cost required to create a translation rule for a grammatical phenomenon that is difficult to describe or cannot be described as a translation rule. For example, in this embodiment, when the example of the specific pattern DB shown in FIG. 3 is implemented in a Japanese-Chinese translation system, the Japanese case particle “NO”, the Japanese formal noun “NO”, and the Japanese auxiliary verb “DA” There is no need to create a translation rule for translation selection.

  Further, in the present embodiment, as the second language statistical model, notation, prototype, inflection form, part of speech, case frame, tense, state, phase, semantic classification, or relation regarding vocabulary or syntax information of the second language You may use the model which stores any one or more statistical co-occurrence information among the co-occurrence patterns which have a receiving relationship.

  In the present embodiment, the error detection / calibration unit uses the second language statistical model to detect unnecessary components and missing components from the second language translation sentence generated by the second language generation unit. Alternatively, the error detection process of any one of the error detections in the word order and the automatic correction process of the obtained error may be performed.

  Further, in the present embodiment, when the second language is Chinese, error detection is performed by using the syntax component of the first language corresponding to the Chinese constant word component as the specific pattern stored in the specific pattern database.・ In the proofreading part, for the Chinese constants of the Chinese translation generated by the second language generation part, the error detection processing of the word order of the Chinese constants and the structure particle “target” and the obtained error The automatic calibration process may be performed.

  Further, in the present embodiment, when the second language is Chinese, error detection is performed by using the syntax component of the first language corresponding to the Chinese character component as the specific pattern stored in the specific pattern database.・ In the proofreading part, for the Chinese-like word component of the Chinese translation sentence generated by the second language generation part, the error detection processing of the word order of the Chinese-like word and the structure particle “earth”, and the obtained error The automatic calibration process may be performed.

  In the present embodiment, when the second language is Chinese, the syntax component of the first language corresponding to the Chinese complement component is used as the specific pattern stored in the specific pattern database, and error detection / In the proofreading section, for the Chinese complement component of the Chinese translation sentence generated by the second language generation section, the error detection processing of the word order of the Chinese complement and the structure particle “Koto”, and automatic correction of the obtained error Processing may be performed.

  In the present embodiment, when the second language is Chinese, the syntax component of the first language corresponding to the Chinese verbal component is used as the specific pattern stored in the specific pattern database, and error detection / The proofreading unit may perform error detection processing of the Chinese quantifier and automatic correction processing of the obtained error for the Chinese quantifier component of the Chinese translation sentence generated by the second language generation unit. Good.

  Further, in the present embodiment, when the second language is Chinese, the syntax of the first language corresponding to the vocabulary and syntax information representing the state information of Chinese is used as the specific pattern stored in the specific pattern database. Vocabulary that expresses Chinese modal information with respect to vocabulary and syntactic information components of the Chinese translation sentence generated by the second language generator in the error detection / proofreading section using the component Alternatively, error detection processing of the syntax information component and automatic correction processing of the obtained error may be performed.

  Further, in the present embodiment, when the second language is Chinese, the syntax component of the first language corresponding to the infix which is a Chinese preposition is used as the specific pattern stored in the specific pattern database, and an error occurs. The detection / calibration unit performs error detection processing of the Chinese language verb and automatic correction processing of the obtained error on the Chinese language component of the Chinese translation generated by the second language generation unit. May be.

[Second Embodiment]
Next, a machine translation device according to a second embodiment of the present invention will be described with reference to FIG. FIG. 5 is a block diagram showing a configuration of a machine translation apparatus according to the second embodiment of the present invention, and the same or equivalent parts as those in FIG. 1 are denoted by the same reference numerals.

As shown in FIG. 5, the present embodiment includes a specific pattern DB 201 instead of the specific pattern DB 103 in the machine translation apparatus according to the first embodiment shown in FIG. 1. Other configurations are the same as those of the first embodiment shown in FIG. 1, and a detailed description thereof is omitted here.
In the specific pattern DB 201, a specific pattern indicating a specific vocabulary used in the first language is added to the information on the first language given as the analysis result of the morpheme / syntax analyzing unit 102, and the information on the first language side. It is a database described in a pair with corresponding information of a specific target on the second language side.

  FIG. 6 is an explanatory diagram showing a specific pattern DB used in the machine translation apparatus according to the second embodiment of the present invention. This example assumes a Japanese-Chinese machine translation system, and what is shown in FIG. 6 is a table for associating translation rule usage information with Chinese structure particles in the Japanese-Chinese machine translation system.

  For example, in the pattern shown in the search ID No. 0 in FIG. 6, the translation rule “capture Japanese language modification” is applied as the analysis rule on the Japanese side when performing morpheme / syntactic analysis on the Japanese input sentence. Means that Further, the Chinese side generation information corresponding to the search ID No. 0 in FIG. 6 means that a fixed word slot is generated. Then, “target” of the Chinese structure particle corresponding to the search ID No. 0 in FIG. 6 means that it is a specific object for determining the appropriateness of the structure particle “target” from the Chinese translation result.

A slot is a component of a case frame used in syntax analysis and is also called a case slot. In parsing based on case grammar, there is a method to analyze by focusing on the language structure (case) required by the verb, and a tree structure describing what kind of case is and what kind of property it has Using the case frame, the morpheme obtained by the morpheme analysis or the combination phrase thereof is applied to the case slot to perform the syntax analysis.
A fixed word is one of the modifiers of Chinese grammatical components and corresponds to a Japanese collocation modifier. Chinese constants and Japanese modifiers have a structure of “modifier (noun, pronoun, adjective) + modifier (noun, noun phrase)”.

In the search ID No. 0 in FIG. 6, the translation rule “import Japanese linkage modification” is applied in the analysis on the Japanese side, and the Chinese fixed word slot is set on the Chinese generation side corresponding to this rule. When generated, it means to determine the appropriateness of the Chinese structure particle “target” for the Chinese translation of the Japanese input sentence.
Similarly, the conditions for determining the appropriateness of the Chinese structure particle “ground” are shown in the specific pattern of the search ID No. 1 in FIG. The conditions for determining the appropriateness of the Chinese structure particle “Koto” are shown in the specific pattern of the search ID No. 2 in FIG.

  FIG. 7 is an explanatory diagram showing a configuration example of the specific pattern DB used in the machine translation apparatus according to the second embodiment of the present invention. This example shows a configuration example of the specific pattern DB 201 for checking the word order of Chinese translation sentences in the daytime machine translation system.

Similar to FIG. 6, the search ID No. 0 in FIG. 7 will be described. As a whole, the search ID No. 0 is applied with the translation rule of “import Japanese language modification” in the analysis on the Japanese side, and If the Chinese generation side corresponding to this rule generates a Chinese fixed word slot and the Chinese fixed word slot has two or more places, the Chinese translation of the Japanese input sentence It means that the appropriateness of the generated word order is determined for the fixed word component.
Similarly, the conditions for determining the appropriateness of Chinese characters are shown in the search ID No. 1 in FIG. The conditions for determining the appropriateness of the Chinese complement are shown in the search ID No. 2 in FIG.

Here, the concepts of Chinese constants, words, and complements are Chinese syntax elements. In general, a Chinese syntactic component is described using a subject, predicate, object, constant, condition, and complement. In machine translation, this information is added to the Chinese case frame.
And the Chinese structural particle “ma” is associated with Chinese constants, the Chinese structural particle “ji” is associated with Chinese characters, and the structural particle “Toku” is associated with Chinese complements. Is. Since these structural particles may or may not exist depending on the Chinese context, it takes a great deal of cost to create strict translation rules in machine translation.

As mentioned above, select the target for error detection and proofreading from the Chinese translations, and determine the appropriateness of the corresponding specific target among the Chinese translations using the Chinese statistical model Thus, error detection / calibration processing such as checking unnecessary components, missing components and word order can be performed.
In addition, when it is necessary to perform error detection and proofreading processing of other components in the Chinese translation, it is described in the specific pattern DB 201 using information such as vocabulary information and translation rules corresponding to the first language. Then, it may be applied to the present invention.

For example, when it is desired to specify the appropriateness of a participle in a Chinese translation, it is only necessary to associate the Chinese participle with a Japanese pronoun and a part of speech pattern of the Japanese noun.
In addition, if you want to specify the appropriateness of particles that represent Chinese modal information, you can create a table with correspondence between Japanese particles or vocabulary that represent Japanese modal information and particles that represent Chinese modal information. Create it.
In addition, when it is desired to specify the appropriateness of an injunction that is a Chinese preposition, a table in which the correspondence between Japanese case particles and adjunct particles and Chinese injunction can be created.

[Operation of Second Embodiment]
Next, the operation of the second embodiment for carrying out the present invention will be described in detail with reference to FIGS. FIG. 8 is a flowchart showing machine translation processing according to the second embodiment of the present invention, and the same or equivalent parts as those in FIG. 4 are denoted by the same reference numerals.
As shown in FIG. 8, in the machine translation processing according to the present embodiment, step S21 is provided instead of step S13 as compared with FIG. 4, and step S22 is added between step S14 and step S15. Has been.

  First, the first language input unit 101 inputs an input sentence of the first language to be translated (step S11), and the morpheme / syntax analysis unit 102 performs morphological analysis or word division on the input sentence. Processing is performed, syntax analysis of the input sentence is performed using the morpheme information, processing for obtaining a vocabulary having a dependency relationship between words included in the input sentence is performed, and the analysis result is stored (step S12). ). At this time, the analysis rules applied at each stage of the morpheme / syntax analysis process are also stored as a part of the analysis result.

Subsequently, from the morpheme / syntax analysis results obtained by the morpheme / syntax analysis unit 102, all of the morpheme / syntax analysis results that match the specific pattern made up of the information in the first language registered in the specific pattern DB 201 by the specific pattern detection unit 105 Are detected as specific pattern candidates (step S21).
Next, the second language generation unit 106 uses the result of the morpheme / syntax analysis obtained by the morpheme / syntax analysis unit 102 and the translation dictionary and translation rules for translation from the first language to the second language. The input sentence in the first language is translated into the second language, and a second language translation sentence that is the translation result is generated (step S14).

  After that, the specific pattern made up of the information on the second language side registered in the specific pattern DB 103 is detected by the specific pattern detection unit 105 from the second language translation sentence obtained in step S14, and the specific pattern obtained Among these, a specific pattern that matches the specific pattern candidate obtained in step S21 is detected, and information on the second language side corresponding to this specific pattern is selected as a specific target for error detection / proofreading (step S22).

Here, when the specific pattern is not detected in step S22, the process proceeds to step S16, and the second language translation sentence generated in step S14 is output by the second language output unit 109 (step S16). End the translation process.
On the other hand, when a specific pattern is detected in step S22, using the specific pattern of the specific target and the specific target of the second language, and the statistical co-occurrence information stored in the second language statistical model 104, After detecting an error in the second language translation sentence generated in step S14 and correcting this error (step S15), the second language output unit 109 outputs the proofread second language translation sentence and the proofreading result. (Step S16), a series of machine translation processing ends.

[Effects of Second Embodiment]
Thus, in this embodiment, a vocabulary having a grammatical phenomenon that is difficult to describe or cannot be described as a translation rule is specified in advance as a set of information on the first language side and information on the second language side as a specific pattern. Registered in the pattern DB 103, information on the first language side of the specific pattern is included in the input sentence of the first language, and information on the second language side of the specific pattern is obtained by translating the input sentence. If it is included in the second language translation sentence, it is stored in the second language statistical model and the specific object consisting of the second language translation sentence corresponding to the specific pattern and the information on the second language side. Since the error of the second language translation sentence obtained by translating the first language is detected using the statistical co-occurrence information and the second language translation sentence is proofread, it is described as a translation rule. Hard or hard to write Without the need for accurate translation rules for the phenomenon, it is possible to perform machine translation.

In this embodiment, since the specific pattern DB 103 in which the specific pattern is described as a set of information on the first language side and information on the second language side is used, information on the first language side of the specific pattern and Corresponding information on the second language side can be set appropriately, and error detection and calibration can be performed with high accuracy.
For example, in the present embodiment, when the example of the specific pattern DB shown in FIGS. 6 and 7 is used, the structure particle “target” in the translation result of Japanese example sentence 1 below is missing, the translation result of Japanese example sentence 2 is It is possible to automatically proofread the correct result by detecting the lack of the structural particle “ground” and the error of the adjective component “Tokyo” in the Japanese example sentence 3.

Example sentences (in parentheses in the examples indicate error locations)
Japanese example sentence 1: “The parking lot at the nearest restaurant is full.”
Chinese translation result: “Recently full stop at a hotel stop”
Chinese correct answer: “Recently, the hotel is full of restaurants.”

Japanese example sentence 2: "What are you thinking about?"
Chinese translation result: “Consideration of the presence of senile dementia?”
Chinese correct answer: “Are you aware of the sensation?”

Japanese example sentence 3: “M Airlines A flight to Tokyo will start boarding at Gate 1 now.”
Chinese translation result: “Old M Airlines A voyage group Tokyo start now.
Chinese correct answer: “Old (Tokyo) M Airlines A voyage starter No. 1 climber entrance lift.”

[Third Embodiment]
Next, a machine translation apparatus according to a third embodiment of the present invention will be described with reference to FIG. FIG. 9 is a block diagram showing a configuration of a machine translation apparatus according to the third embodiment of the present invention. The same reference numerals are given to the same or equivalent parts as FIG.

As shown in FIG. 9, the present embodiment is the same as the components of the machine translation apparatus according to the first embodiment shown in FIG. 1, but is configured by changing the connection relationship of each functional unit. Specifically, in accordance with the detection result of the specific pattern in the specific pattern detection unit 105, the second language generation unit 106 determines whether or not to perform error detection / configuration processing in the error detection / calibration unit 107. Yes. Other points are the same as those of the first embodiment.
Even with such a connection relationship, the same effects as those of the first embodiment can be obtained.

[Fourth Embodiment]
Next, a machine translation apparatus according to the fourth embodiment of the present invention will be described with reference to FIG. FIG. 10 is a block diagram showing a configuration of a machine translation apparatus according to the fourth embodiment of the present invention. The same or equivalent parts as those in FIG.

  As shown in FIG. 10, the present embodiment is the same as the constituent elements of the machine translation apparatus according to the second embodiment shown in FIG. 5, but is configured by changing the connection relationship of each functional unit. Specifically, in accordance with the detection result of the specific pattern in the specific pattern detection unit 105, the second language generation unit 106 determines whether or not to perform error detection / configuration processing in the error detection / calibration unit 107. Yes. In addition, a translation selection unit 401 is added between the morpheme / syntax analysis unit 102 and the specific pattern detection unit 105. Other points are the same as those of the first embodiment.

The translation word selection unit 401 uses the morpheme / syntax analysis result for the input sentence in the first language obtained by the morpheme / syntax analysis unit 102 to acquire a translation word candidate of each morpheme from the translation dictionary of the translation DB 108, The translation word candidate is applied to the translation rules of the translation DB 108 and translation rule selection processing rules to obtain the optimal translation word candidate for each morpheme.
Accordingly, when the specific pattern detection unit 105 detects a specific pattern made up of information on the second language side registered in the specific pattern DB 103 from the second language translation generated by the second language generation unit 106, A specific pattern can be searched with high accuracy, and a specific target for error detection / calibration can be selected more accurately.

  Next, the first embodiment of the present invention will be described with reference to FIG. FIG. 11 is a block diagram showing the configuration of the machine translation apparatus according to the first embodiment of the present invention. The same reference numerals are given to the same or equivalent parts as in FIG.

  This example corresponds to the first and third embodiments described above. As shown in FIG. 11, the present example is the first language input unit 101, the specific pattern DB 103, the second language statistical model 104, the first of the machine translation devices according to the first embodiment shown in FIG. 1. Instead of the bilingual generation unit 106 and the second language output unit 109, a Japanese input unit 501, a specific pattern DB 500, a Chinese statistical model 502, a Chinese generation unit 503, and a Chinese output unit 504 are provided. Yes. These specialize the first language and the second language of the first embodiment in Japanese and Chinese, and are substantially equivalent to the components of the first embodiment.

  FIG. 12 is a flowchart showing the machine translation process according to the first embodiment of the present invention. The same reference numerals are given to the same or equivalent parts as those in FIG. Here, an example will be described in which an input sentence “It is between three and five columns” input in Japanese is machine-translated into Chinese.

When the input sentence input by the Japanese input unit 501 is analyzed by the morpheme / syntax analyzing unit 102 (step S12), “it / ha /, / three / column / and / five / column / The result of morpheme analysis is obtained. Each morpheme obtained by morpheme analysis has a unique attribute value. The attribute value includes information such as original form, part of speech, notation, utilization form, semantic classification, state, and phase.
For example, the morphemes “no” and “a” in the example sentence have the following attributes.
Notation Kana Proper part Prototype Part of speech…
Nono's case particle ...
A certain a certain certain verb ...

  In addition, the translation dictionary of the translation DB 108 includes information such as a translation of the morpheme, a part of speech, a semantic classification, a case frame in case of a predicate, a state, and phase information in the Chinese generation block of each morpheme. The constituent elements of the Chinese case frame describe sentence structure information such as a subject, predicate, object, constants, adjectives, and complements as elements of the case frame for Chinese generation.

The morpheme / syntax analyzer 102 performs syntax analysis of the input sentence together with the syntax analysis rules using the morpheme analysis result described above (step S12). As a result of the parsing process, it is possible to obtain a dependency relationship between words having a dependency relationship in the input sentence.
The specific pattern detection unit 105 uses the result of the morpheme analysis / syntax analysis process in the morpheme / syntax analysis unit 102 to perform a specific pattern detection process by collating with the specific pattern registered in the specific pattern DB 500 (step S52). .

  For example, the case particle “no” is included in the result of the morpheme / syntactic analysis process of the input sentence “it is between the third and fifth columns”. When the specific pattern detection unit 105 performs pattern matching with the input sentence using the specific pattern DB for intraday translation shown in FIG. 3, the case particle “NO” of search ID 0 is detected. Then, with respect to the detected case particle “no”, a symbol “Φ” indicating “no translation” as a candidate word of the case particle “no” and a Chinese structure particle “target” are found from the translation dictionary of the translation DB 108. To be acquired.

Further, when the specific pattern detection unit 105 detects a specific pattern described in the specific pattern DB 500 from the result of the morpheme / syntax analysis process of the input sentence, information on the morpheme adjacent to the detected specific pattern is displayed. Obtained from the result of morpheme / syntactic analysis.
For example, from the morpheme / syntax analysis result of the example sentence, the information on the vocabulary “sequence” and “between” adjacent to the case particle “no” can be cut out and acquired. In addition, information on the Chinese side corresponding to “column”, “no”, and “between” can be acquired from the dictionary lookup process of the Sino-Japanese translation dictionary.

  Thereby, attribute information such as Japanese vocabulary notation and part of speech and the corresponding Chinese vocabulary notation and part of speech attribute information can be acquired. For example, the Japanese part-of-speech “classifier” in the vocabulary “column” before the case particle “no” and the corresponding Chinese part-of-speech “quantifier”, and the word “ma” after the case particle “no” It can acquire “noun”, which is part of Chinese corresponding to “noun”, which is part of Japanese. Therefore, as part-of-speech information in the Chinese generated vocabulary corresponding to Japanese “column + no + ma”, patterns such as “quantifier + Φ + noun” or “quantifier + target + noun” are used for morpheme / syntax analysis of the input sentence. It can be obtained from the processing result.

Next, the Chinese generation unit 503 generates a Chinese input sentence, that is, a Chinese translation, using the result of the morpheme / syntax analysis, the Japanese-Chinese translation dictionary, and the translation rules (step S53). In this Chinese translation, attribute information such as Chinese part of speech is assigned to each Chinese morpheme.
Here, when the specific pattern is not detected by the specific pattern detection unit 105, the Chinese translation sentence obtained by the Chinese generation unit 503 is shaped without performing error detection / proofreading processing, and the Chinese output unit 504 From Chinese (step S54).

  On the other hand, when the specific pattern is detected by the specific pattern detection unit 105 as in the example sentence, the error detection / calibration unit 107 performs Chinese statistical analysis on the specific target in the Chinese translation sentence specified based on the specific pattern. Using the Chinese statistical model 502 storing the co-occurrence information, an error is detected from the Chinese translation and proofreading is performed (step S15), and the proofread result is output from the Chinese output unit 504. (Step S54).

A specific error detection and proofreading process in the error detection / proofreading unit 107 will be described with respect to a Japanese example sentence using a Chinese N-gram model.
As described above, when the translation of the case particle “no” is acquired with the symbol “Φ” of “no translation” as a translation candidate, when error detection / configuration processing is performed on this symbol “Φ” , Convert to empty character. In the case of “quantifier + Φ + noun”, it may be converted into “quantifier + noun”.

  For example, if the Japanese input sentence “It is between the 3rd and 5th rows”, the Chinese translation obtained by the Chinese generation unit 503 is “Na / pronouns / quantifiers / adverbs 3 / numerals”. Column / quantifier sum / adjunct five / numerical string / quantifier / structure particle noma / noun ”. Here, as an example of the Chinese statistical model 502, when an N-gram model constructed with word notation and part-of-speech information is used in a Chinese monolingual corpus, the suitability of the Chinese structure particle “target” is determined. For this algorithm, the following calculation method can be used.

First, there is a method of determining the appropriateness of the Chinese identification target w _de using the generation probability that is approximated by 3-gram. In the following, w represents a vocabulary notation or a combination of notation and part of speech.
For example, sentence translation result _{1 = w 0, w 1,} w 2, ..., w i-1, w i, w i + 1, and ... w _n, this to those considering the position of the w _de, sentences _{2 = w 0, w 1,} w 2, ..., w i-1, w de, w i, w i + 1, it is assumed that the ... w _n.

Here, when the generation probability of sentence 1 is approximated by 3-gram, the following expression (1) is obtained. Similarly, when the generation probability of sentence 2 is approximated by 3-gram, expression (2) is obtained.

Therefore, the calculation formula that can determine the appropriateness of w _de in (w _i−1 , w _de , w _i ) can be expressed by the following formula (3).

Furthermore, sentence _{= w 0, w 1, ...} , for the generation probability of the w _n, may be used the calculation method shown in the following using 3-gram formula (4).

  Of the Chinese translations obtained by the Chinese generation unit 503, the specific pattern detection unit 105 uses two words, “quantifier + noun” and “quantifier + target + noun”, which are Chinese vocabulary and syntax information. When one pattern is specified as the target of the translation result, “sequence / quantitative / structural particle noma / noun” is acquired as a part that matches the pattern of “quantifier + target + noun”.

In this case, using the Chinese N-gram language statistical model with parts of speech, the following two conditional probability values are calculated using the algorithms shown in equations (1) to (3). Compare.
P1 = P (target / structural particle | five / numerical string / quantifier) / P (noma / noun | string / quantifier / structural particle)
P2 = P (Noma / Noun | Five / Numeric string / Quantifier)

Here, the probability value is calculated in Chinese 3-gram,
P1 = 8.333336e-005
P2 = 1.25000e-001
When a result like this is obtained, the probability value of P2 is higher than P1, and therefore the error “target / structural particle” in (sequence / quantitative / structural noun / noun) can be detected from the translation result.
At this time, if the P2 pattern “5 / numeral string / quantifier noma / noun” is cut out and replaced with “sequence / quantifier / structure particle noma / noun” in the translation result sentence, the Chinese translation sentence The proofreading result is "Na / pronoun individual / quantifier in / injunction three / numerical string / quantifier sum / adjunct five / numerical string / quantifier noma / noun".

In addition, when calculating the probability value of P1, if 3-gram that contains an error (column / quantifier / structure particle noma / noun) is not detected, P (noma / noun | sequence / quantifier) The value of / structural particle) is 2-gram and 1-gram, and the approximate probability can be calculated by interpolation processing by smoothing.
Specifically, the smoothing method of the N-gram model includes countable smoothing, linear interpolation, back-off smoothing, Witon Bell smoothing, uncounting method, and the like.

In addition, when using the algorithm shown in Expression (4) in the processing in the error detection / calibration unit 107, the Chinese vocabulary and syntax information detected by the specific pattern detection unit 105 and the Chinese generation unit 503 All sentence occurrence probabilities that co-occur with the Chinese translations obtained can be calculated, and the sentence with the maximum occurrence probability can be determined as the correct answer.
In the example sentence, two candidate words for the case particle “no” are obtained, and all sentence occurrence probabilities when “target” is generated for the Chinese translation and when “target” is not generated are calculated. The maximum result may be used as the calibration result. When an error location is to be detected, the difference between the generated sentence with the maximum sentence occurrence probability and the Chinese translation may be obtained.

For example, if the Chinese translation for a Japanese input sentence is “Na / pronouns / quantifiers in / adjunct three / numerical strings / quantifier sum / adjunct five / numerical strings / quantifier-like / structural particles noma / nouns”. The occurrence probability of the following sentence can be approximated.
■ Sentences that do not contain “target / structural particles”:
Sentence 3 “Na / pronouns individual / quantifiers present / intern three / numerical strings / quantifiers sum / adjunct five / numerical strings / quantifiers noma / nouns”
■ Sentences that contain “target / structural particles”:
Sentence 4 “Na / pronouns individual / quantifiers present / adjunct three / numerical strings / quantifiers sum / adjunct five / numerical strings / quantifier-like / structural particles noma / nouns”
Sentence 5 “Na / pronouns / structural particles individual / quantifiers present / adjunct three / numerical strings / quantifiers sum / adjunct five / numerical strings / quantifiers noma / nouns”
Sentence 6 “Na / pronoun individual / quantifier-like / structural particle resident / adjunct three / numerical string / quantifier sum / adjunct five / numerical string / quantifier noman / noun”
Sentence 7 “Na / pronoun individual / quantifier in / injective / structure particle three / numerical string / quantifier sum / adjunct five / numerical string / quantifier noma / noun”
Sentence 8 “Na / pronouns individual / quantifiers present / adjunct three / numerical / structural particle sequence / quantifier sum / adjunct five / numerical string / quantifier noma / noun”
Sentence 9 “Na / pronouns individual / quantifiers present / adjunct three / numerical strings / quantifier-like / structural particles sum / adjunct five / numerical strings / quantifiers noma / nouns”
Sentence 10 “Na / pronouns individual / quantifiers present / adjunct three / numerical strings / quantifiers sum / adjunctive / structural particles five / numerical strings / quantifiers noma / nouns”
Sentence 11 “Na / pronouns individual / quantifiers in / in adjunct three / numerical strings / quantifiers sum / adjunctive five / numerical / structural particle strings / quantifiers noma / nouns”
Sentence 12 “Na / pronouns individual / quantifiers present / adjunct three / numerical strings / quantifiers sum / adjunct five / numerical strings / quantifier-like / structural particles noma / nouns”
Sentence 13 “Na / pronouns individual / quantifiers present / adjunct three / numerical strings / quantifiers sum / adjunct five / numerical strings / quantifier strings / nounative / structural particles”

The sentence occurrence probabilities of these sentences 3 to 13 are calculated using Chinese 3-gram, and the sentence with the maximum probability value can be set as the optimum translation result. When calculating the sentence occurrence probability here, the above-described smoothing method of the N-gram model can be used.
Assume that the following calculation result of the sentence occurrence probability in the back-off / smoothing smoothing process is obtained for the example sentence by the Witten-Bell disc.
Sentence probability of sentence 3: 1.49286e-008
Sentence occurrence probability of sentence 4: 1.51677e-011
Sentence occurrence probability of sentence 5: 1.01744e-010
Sentence occurrence probability of sentence 6: 5.67270e-011
Sentence occurrence probability of sentence 7: 1.05371e-010
Sentence occurrence probability of sentence 8: 3.20494e-010
Sentence occurrence probability of sentence 9: 1.53964e-011
Sentence occurrence probability of sentence 10: 8.87810e-010
Sentence occurrence probability of sentence 11: 3.079999e-010
Sentence probability of sentence 12: 1.49252e-011
Sentence occurrence probability of sentence 13: 1.03949e-009

  In this case, the probability values are sorted, and the result of sentence 3 is maximized, so that sentence 3 can be the correct answer. Further, when it is desired to detect an error in the Chinese structure particle “target”, it is obtained from the difference between the sentence 3 and the Chinese translation sentence.

Also, a Class N-gram model constructed with a Chinese monolingual corpus can be used with the Chinese part of speech as the word class.
As a model construction method, a Class bigram model or a Class trigram model described in Non-Patent Document 1 can be used.

An algorithm for detecting and correcting an error in the target vocabulary from the second language translation sentence using the Class N-gram model is the second algorithm using the Class bigram model and the Class trigram model recorded in Non-Patent Document 1. Of the language translation sentences, the sentence having the maximum sentence generation probability considering the presence or absence of “target” may be estimated as the most likely proofreading result.
The Chinese statistical model may be constructed by any learning method of decision list, SVM, maximum entropy, HMM, and Bayesian learning. Of course, it is not limited to these learning methods.

In this way, if the Chinese translation of the Japanese input sentence “It is between the third and fifth rows” becomes “Natashi, the third row, the fifth row, and the fifth row”, The correct result is obtained as “Nana three-row sum five-row nooma”.
Similarly to the first embodiment, it is possible to solve the problem of selecting a translation of a formal noun and an auxiliary verb “DA” shown in FIG.

  Next, a second embodiment of the present invention will be described with reference to FIG. FIG. 13 is a block diagram showing the configuration of the machine translation apparatus according to the second embodiment of the present invention. The same or equivalent parts as those in FIG. 5 are given the same reference numerals.

  This example corresponds to the second and fourth embodiments described above. As shown in FIG. 13, the present example is the first language input unit 101, the specific pattern DB 103, the second language statistical model 104, the first of the machine translation devices according to the fourth embodiment shown in FIG. 10. Instead of the bilingual generation unit 106 and the second language output unit 109, a Japanese input unit 602, a specific pattern DB 601, a Chinese statistical model 603, a Chinese generation unit 604, and a Chinese output unit 605 are provided. Yes. These specialize the first language and the second language of the fourth embodiment in Japanese and Chinese, and are substantially equivalent to the constituent elements of the fourth embodiment.

FIG. 14 is a flowchart showing a machine translation process according to the second embodiment of the present invention, and the same or equivalent parts as those in FIG.
Here, a case where the following input sentence input in Japanese is machine-translated into Chinese will be described as an example.
“The closest restaurant parking lot is full”
“What are you thinking about?”
“M Airlines A flight to Tokyo will start boarding at Gate 1”

When the input sentence input by the Japanese input unit 501 is analyzed by the morpheme / syntax analyzing unit 102 (step S61),
"The most / close / restaurant / no / parking lot / ha / full /"
"What / What / Vague / Thinking / Te / I ///"
“M Airlines / A flight / Tokyo / bound / ha / just now / from / first / gate / de / boarding / start / do / do”
The result of morphological analysis is obtained. Each morpheme obtained by morpheme analysis has a unique attribute value as described above.

The morpheme / syntax analyzing unit 102 performs the syntax analysis processing of the input sentence together with the syntax analysis rules using the morpheme analysis result described above. As a result of the parsing process, it is possible to obtain a dependency relationship between words having a dependency relationship in the input sentence.
For example, FIG. 2 shows the result of analyzing the input sentence “The closest restaurant parking lot is full” based on the context-free grammar rules.
In addition, when performing Japanese morpheme / syntactic analysis processing, analysis rules applied at each analysis stage can be assigned to morphemes or those having a dependency relationship.

  Next, the translation selection unit 401 acquires translation word candidates of each morpheme of the input sentence using the result of the morpheme / syntax analysis processing of the input sentence and the translation dictionary of the translation DB 108, and stores it in the storage unit (step S1). S62). At this time, it is applied to the translation word selection rules of the translation DB 108, and the optimal translation candidate for each morpheme is acquired and stored in the storage unit. Also, the translation selection rules applied when the Chinese translation selection process is performed are stored in the storage unit as part of the analysis result in the Chinese translation of the target vocabulary.

Subsequently, the specific pattern detection unit 105 uses the result of the morpheme / syntax analysis processing of the input sentence and the translation selection processing result of each Japanese morpheme in the translation selection unit 401 to identify the specific pattern registered in the specific pattern DB 601. When a pattern matching process with a pattern is performed and a specific pattern is detected, the specific pattern is stored as a candidate in the storage unit (step S63).
Next, the Chinese generation unit 604 generates a Chinese input sentence, that is, a Chinese translation, using the result of the morpheme / syntactic analysis, the daytime translation dictionary, and the translation rules (step S64). In this Chinese translation, attribute information such as Chinese part of speech is assigned to each Chinese morpheme.

After that, the specific pattern detection unit 105 detects a specific pattern including information indicating the Chinese generation rule registered in the specific pattern DB 601 from the obtained Chinese translation, and among the obtained specific patterns, A specific pattern that matches the specific pattern stored as a candidate is detected, and the corresponding Chinese translation is specified as an error detection / proofreading target.
Here, when the specific pattern is not detected by the specific pattern detection unit 105, the Chinese translation sentence obtained by the Chinese generation unit 604 is shaped without performing error detection / calibration processing, and the Chinese output unit 605. To output Chinese (step S65).

  On the other hand, when the specific pattern is detected by the specific pattern detection unit 105, the error detection / calibration unit 107 outputs Chinese statistical co-occurrence information for the specific target in the Chinese translation specified based on the specific pattern. Using the stored Chinese statistical model 603, an error is detected from the Chinese translation and proofreading is performed (step S15), and the proofread result is output from the Chinese output unit 605 (step S65). .

  In the specific pattern detection unit 105 described above, for example, using the result of the morpheme / syntax analysis processing of the input sentence “the closest parking lot of the restaurant is full” is compared with the example shown in the specific pattern DB shown in FIG. Process. In the input sentence, “incorporation modification incorporation” can be acquired from the analysis rule applied to the analysis of “close” + “restaurant”. In addition, the Chinese generation rules corresponding to Japanese “incorporation of collocation modification” in Japanese-Chinese machine translation generally generate “constant word slots”. By detecting the pattern with the pattern search ID 0, it is possible to determine the appropriateness of the Chinese structure particle “target” from the Chinese translation sentence with the Chinese structure particle “target” given thereto as a specific target. .

In addition, in order to guarantee the detection accuracy, the words “Near” and “Haiten” are derived from the generated Chinese results using the words “Near” and “Part of speech” of “Near”, and the word “Restaurant” and part of speech of “Restaurant”. The ambiguity of the vocabulary on the Chinese side can be improved by performing the process of determining whether or not to insert “target” in between.
For example, if the Chinese translation of the input sentence “Closest restaurant parking lot is full” becomes “Recently full restaurant stop”, the translation means “Recently full restaurant stop” in Japanese. Recently, the restaurant parking lot is full. " The part-of-speech of “recent” in this incorrect Japanese sentence is a temporal noun or adverb, and “recent” in the Chinese translation corresponding to this recently becomes a temporal adverb. Therefore, detection accuracy can be improved by performing error detection processing in consideration of correspondence between vocabulary between two languages and part-of-speech information.

As an example of the Chinese statistical model 603, for example, an N-gram model constructed with a Chinese monolingual corpus with the surface of the face or part-of-speech information, the appropriateness of the Chinese structure particle “target” is determined. As the algorithm for the above, the calculation method of the above-described formula (1) to formula (3) or formula (4) can be used.
Also, a Class N-gram model constructed with a Chinese monolingual corpus can be used with the Chinese part of speech as the word class.

As a model construction method, a Class bigram model or a Class trigram model described in Non-Patent Document 1 can be used. The algorithm for detecting and correcting word order errors from Chinese translations using the Class N-gram model is based on the class bigram model and class trigram model recorded in non-patent literature. What has become the most probable proofreading result should be the one with the highest sentence generation probability considering the word order.
The Chinese statistical model 603 may be constructed by any learning method of decision list, SVM, maximum entropy, HMM, and Bayesian learning. Of course, it is not limited to these methods.

In addition, when a Chinese word is generated, the structure particle “ji” corresponding to the word is generated in the same manner as described above, and when the Chinese word is generated, the structure particle “ Can be determined and the calibration process can be performed.
Further, in the case where there are two or more constant words, words, and complements in the Chinese translation, the generated Chinese word order can be checked in the same manner. For example, if the specific pattern DB shown in FIG. 7 is used, a Chinese translation of “M Airlines A flight to Tokyo will start boarding at Gate 1 from now” in the Japanese sentence example, In the case of “M aviation A aviation group Tokyo starting current No. 1 climax”, according to the algorithm described above, the correct China “Old Tokyo M aviation A flight group currently starting No. 1 climax” A word translation can be generated.

Similarly to the second embodiment, error detection / calibration processing is performed by considering components such as Chinese subject, predicate, object, fixed word, synonym, and complement from the Chinese translation sentence in an integrated manner. By doing so, the translation quality can be greatly improved.
Similarly to the second embodiment, for example, a table in which the correspondence between Japanese particles or auxiliary verbs representing Japanese phase information and Chinese phase information is described in the specific pattern DB 201. For example, it is possible to detect and proofread errors in the Chinese translation from the Japanese translation results.

For example, for a Japanese example sentence “I think I'm catching a cold” when assuming a Japanese-Chinese machine translation system, the translation result using only the Chinese statistical model information If an error is detected, there is a high possibility that an erroneous component such as “acquired cold” or “acquired cold” is estimated.
On the other hand, the three candidate words "Chaku", "Ryo" and "Dai" representing the Chinese language corresponding to the Japanese language information "I" are errors in the Chinese translation. As a detection target, when the Chinese translation result of a Japanese sentence is “I am for a common cold”, the error detection targets “Arrival”, “End”, “A” and “I am for a common cold. If the correct answer is the one with the highest sentence generation probability, the correct translation can be obtained.

  As described above, according to the present invention, the translation accuracy of the machine translation system can be greatly improved.

  As mentioned above, although this invention was demonstrated with reference to embodiment and an Example, this invention is not limited to the said embodiment and Example. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

A machine translation apparatus, a translation method, and a program for translating a first language into a second language according to the present invention perform post-editing and creation of a translation rule for a translation result of a conventional machine translation system in a machine translation system or the like It is suitable for a machine translation device, a machine translation method, and a program that can greatly improve the translation quality of a machine translation device while solving the high-cost problem.
According to the present invention, as described above, it is possible to improve the problem that it is difficult to improve the error of the translation result in machine translation, and greatly improve the translation quality. The translation result with high translation accuracy can be correctly provided to the user.

  DESCRIPTION OF SYMBOLS 100 ... Machine translation apparatus, 101 ... 1st language input part, 102 ... Morphological and syntax analysis part, 103 ... Specific pattern DB (only 1st language), 104 ... 2nd language statistical model, 105 ... Specific pattern detection part, 106: second language generation unit, 107: error detection / calibration unit, 108: translation DB, 109: second language output unit, 201: specific pattern DB (corresponding to two languages), 401: translation selection unit, 501: Japanese Input unit 502 ... Chinese statistical model 503 ... Chinese generation unit 504 ... Chinese output unit 601 ... Specific pattern DB (corresponding to daytime) 602 ... Japanese input unit 603 ... Chinese statistical model 604 ... Chinese generation unit 605 ... Chinese output unit.

Claims (12)

A morpheme / syntax analyzer that performs morpheme / syntax analysis on an input sentence expressed in a first language;
A translation database comprising a translation dictionary or translation rules used to translate the first language into the second language;
A second language generation unit that refers to the analysis result of the morpheme / syntax analysis unit and the translation database, and generates a second language translation as a translation result corresponding to the input sentence;
A specific pattern database for storing a specific vocabulary used in the first language as a specific pattern;
A specific pattern stored in the specific pattern database is detected from the analysis result of the morpheme / syntax analysis unit, and the lexical or syntax information of the second language corresponding to the obtained analysis result of the specific pattern is obtained from the translation database. A specific pattern detection unit to be acquired;
A second language statistical model storing statistical co-occurrence information relating to vocabulary co-occurrence used in the second language;
Generated by the second language generator using the vocabulary or syntax information of the second language obtained by the specific pattern detector and the statistical co-occurrence information stored in the second language statistical model A machine translation apparatus comprising: an error detection / calibration unit that detects an error in the second language translation sentence and calibrates the second language translation sentence. The machine translation device according to claim 1,
The specific pattern stored in the specific pattern database includes information on the first language provided as an analysis result of the morpheme / syntax analyzer, and information on the second language corresponding to the information on the first language. Machine translation apparatus characterized by being described in a set of The machine translation device according to claim 1,
The second language statistical model is a co-occurrence having a notation, a prototype, an inflection, a part of speech, a case frame, tense, a state, a phase, a semantic classification, or a dependency relationship regarding the vocabulary or syntax information of the second language. A machine translation device storing at least one piece of statistical co-occurrence information among patterns. The machine translation device according to claim 1,
The error detection / calibration unit uses the second language statistical model to detect unnecessary components, missing components, or word order from the second language translation generated by the second language generation unit. A machine translation apparatus that performs error detection processing of any one of error detection and automatic correction processing of an obtained error. The machine translation device according to claim 1,
The second language is Chinese, the specific pattern stored in the specific pattern database includes a syntactic component of the first language corresponding to a Chinese constant component, and the error detection / calibration unit includes the An error detection process of the word order of the Chinese definite word and the structure particle “target”, and an automatic correction process of the obtained error for the Chinese definite word component of the Chinese translation sentence generated by the second language generation unit A machine translation apparatus characterized by The machine translation device according to claim 1,
The second language is Chinese, the specific pattern stored in the specific pattern database includes a first language syntax component corresponding to a Chinese character component, and the error detection / calibration unit includes the An error detection process of the Chinese word word order and the structure particle “ground” for the Chinese word component of the Chinese translation generated by the second language generation unit, and an automatic correction process of the obtained error A machine translation apparatus characterized by The machine translation device according to claim 1,
The second language is Chinese, the specific pattern stored in the specific pattern database includes a syntax component of the first language corresponding to a Chinese complement component, and the error detection / calibration unit includes the first language Performing error detection processing of the Chinese complement word order and structure particle “Koto” and automatic correction processing of the obtained error for the Chinese complement component of the Chinese translation generated by the bilingual generator Machine translation device characterized by the above. The machine translation device according to claim 1,
The second language is Chinese, the specific pattern stored in the specific pattern database includes a syntax component of the first language corresponding to a Chinese quantile component, and the error detection / proofreading unit includes the first language A machine translation apparatus that performs error detection processing of a Chinese quantile and automatic correction processing of an obtained error on a Chinese quantile component of a Chinese translation sentence generated by a bilingual generation unit. The machine translation device according to claim 1,
The second language is Chinese, and the specific pattern stored in the specific pattern database includes a syntax component of the first language corresponding to vocabulary and syntax information representing Chinese modal information, and the error detection The proofreading unit uses the vocabulary and syntax information component representing the Chinese modal information for the vocabulary and syntax information component representing the Chinese modal information of the Chinese translation generated by the second language generation unit. A machine translation apparatus that performs an error detection process and an automatic correction process for an obtained error. The machine translation device according to claim 1,
The second language is Chinese, and the specific pattern stored in the specific pattern database includes a syntax component of the first language corresponding to an infix that is a Chinese preposition, and the error detection / calibration unit includes: A machine translation, comprising: performing Chinese error detection processing and automatic error correction processing on the Chinese language component of the Chinese translation component generated by the second language generation unit. apparatus. A machine translation method used in a machine translation device that translates a first language into a second language,
A morpheme / syntactic analysis unit that performs a morpheme / syntax analysis on an input sentence expressed in a first language;
The second language generation unit corresponds to the input sentence by referring to the analysis result of the morpheme / syntax analysis unit and a translation database or translation database used for translating the first language into the second language. A second language generation step of generating a second language translation as a translation result to be performed;
A specific pattern detection unit detects a specific pattern stored in a specific pattern database that stores a specific vocabulary used in the first language as a specific pattern from the analysis result of the morpheme / syntax analysis unit, and the obtained specific pattern A specific pattern detection step of acquiring vocabulary or syntax information of a second language corresponding to a pattern analysis result from the translation database;
A second language statistic in which the error detection / proofreading unit stores vocabulary or syntax information of the second language obtained by the specific pattern detection unit and statistical co-occurrence information regarding the co-occurrence of the vocabulary used in the second language Error detection for detecting an error in the second language translation sentence generated by the second language generation unit using statistical co-occurrence information stored in a static model and calibrating the second language translation sentence A machine translation method comprising: a proofreading step. In the computer of the machine translation device that translates the first language into the second language,
A morpheme / syntactic analysis unit that performs a morpheme / syntax analysis on an input sentence expressed in a first language;
The second language generation unit corresponds to the input sentence by referring to the analysis result of the morpheme / syntax analysis unit and a translation database or translation database used for translating the first language into the second language. A second language generation step of generating a second language translation as a translation result to be performed;
A specific pattern detection unit detects a specific pattern stored in a specific pattern database that stores a specific vocabulary used in the first language as a specific pattern from the analysis result of the morpheme / syntax analysis unit, and the obtained specific pattern A specific pattern detection step of acquiring vocabulary or syntax information of a second language corresponding to a pattern analysis result from the translation database;
A second language statistic in which the error detection / proofreading unit stores vocabulary or syntax information of the second language obtained by the specific pattern detection unit and statistical co-occurrence information regarding the co-occurrence of the vocabulary used in the second language Error detection for detecting an error in the second language translation sentence generated by the second language generation unit using statistical co-occurrence information stored in a static model and calibrating the second language translation sentence • A program that executes calibration steps and.

Images