JP5879989B2 - Machine translation system, machine translation method, and machine translation program - Google Patents

Description

  The present invention relates to a machine translation system, a machine translation method, and a machine translation program, and more particularly to a machine translation system, a machine translation method, and a machine translation program that perform statistical machine translation.

  Tree-based statistical machine translation systems that handle the structure of a language include a text-to-text system in which both the source language and the target language learning corpus are text data. An example of such a machine translation system is described in Patent Document 1. The machine translation system of Patent Document 1 converts a source language sentence into a target language (corresponding to a target language in this specification) based on statistical information learned from a bilingual corpus (corresponding to a learning corpus in this specification). It is something to translate. This system generates a variable bilingual corpus that replaces a part of a word registered in the word dictionary as a variable in a source language sentence and a target language sentence in the bilingual corpus, and a translation model using the generated bilingual corpus And learn language models. Also for the input sentence, after the word registered in the word dictionary is replaced with a variable in the input sentence variableizing unit, the statistical translation unit performs translation using the learned model. The variable replacement unit outputs the result of replacing the variable included in the translation result with the word of the target language as the final translation result based on the information when the word in the input sentence is replaced with the variable and the word dictionary.

  However, in the system described in the literature described above, since the learning data is only text data of the parallel corpus of the source language and the target language, grammatical information is not included. For this reason, the translation model learned from here contains a lot of non-grammatical rules, and the translation result created by selecting this non-grammatical rule is likely to be an incorrect translation result. There was a problem of becoming.

  In order to avoid this, Non-Patent Document 1 has a high accuracy by learning a translation model from syntax tree data as a tree-to-tree system using the learning corpus of both the source language and the target language as a syntax tree. It is a translation system.

JP 2010-282453 A

I. Dan Melamed, "Statistical Machine Translation by Parsing", Proceedings of the 42nd Annual Meeting on Association for Computational Linguistics (ACL-04), 2004

  In the technique described in the above-mentioned document, in order to learn the translation model from the syntax tree data, a syntax analysis system for creating the syntax tree data from the text data is required. was there. In particular, in order to construct a system that handles multiple languages, a syntax analysis system for each language is required. Therefore, it is necessary to create the syntax analysis system, and there is a problem that the system development cost is particularly high.

  An object of the present invention is to provide a machine translation system, a machine translation method, and a machine translation program that solve the above-described problem that the system development cost is high.

The machine translation system of the present invention
At least one translation rule obtained by extracting a translation rule indicating the correspondence between the source language and the target language and the correctness of the correspondence from the source language text corpus and the target language text corpus, and the translation rule A translation model generating means for generating a translation model including a probability value indicating the probability of
A syntactic information rule describing a syntactic condition used to change the translation model;
Rule applying means for changing the translation model according to the condition of the syntactic information rule;
Is provided.

The machine translation method of the present invention includes:
Machine translation device
From the source language text corpus and the target language text corpus, a translation rule indicating the correspondence between the two languages and the correctness of the correspondence are extracted, and at least one obtained translation rule and the translation rule Generate a translation model that includes a probability value indicating the probability ,
Describe the syntactic conditions used to change the translation model in the syntactic information rules,
The translation model is changed according to the condition of the syntactic information rule.

The computer program of the present invention is:
From the source language text corpus and the target language text corpus, a translation rule indicating the correspondence between the two languages and the correctness of the correspondence are extracted, and at least one obtained translation rule and the translation rule A procedure for generating a translation model including a probability value indicating the probability ,
A machine translation program for causing a computer to execute a procedure for changing the translation model in accordance with the condition of the syntactic information rule describing a syntactic condition used for changing the translation model.

  It should be noted that any combination of the above-described constituent elements and a conversion of the expression of the present invention between a method, an apparatus, a system, a recording medium, a computer program, etc. are also effective as an aspect of the present invention.

  The various components of the present invention do not necessarily have to be independent of each other. A plurality of components are formed as a single member, and a single component is formed of a plurality of members. It may be that a certain component is a part of another component, a part of a certain component overlaps with a part of another component, or the like.

  Moreover, although the several procedure is described in order in the method and computer program of this invention, the order of the description does not limit the order which performs a several procedure. For this reason, when implementing the method and computer program of this invention, the order of the several procedure can be changed in the range which does not interfere in content.

  Furthermore, the plurality of procedures of the method and the computer program of the present invention are not limited to being executed at different timings. For this reason, another procedure may occur during the execution of a certain procedure, or some or all of the execution timing of a certain procedure and the execution timing of another procedure may overlap.

  According to the present invention, there are provided a machine translation system, a machine translation method, and a machine translation program that can accurately translate at a low cost even when a syntax analysis system is not used.

It is a functional block diagram which shows the structure of the machine translation system which concerns on embodiment of this invention. It is a flowchart which shows an example of operation | movement of the machine translation system which concerns on embodiment of this invention. It is a functional block diagram which shows the structure of the machine translation system which concerns on embodiment of this invention. It is a flowchart which shows an example of operation | movement of the machine translation system which concerns on embodiment of this invention. It is a figure for demonstrating the syntactic information rule of the machine translation system which concerns on embodiment of this invention. It is a figure for demonstrating the machine translation system of the Example of this invention. It is a figure for demonstrating the machine translation system of the Example of this invention. It is a figure for demonstrating the machine translation example of the comparative example of this invention.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In all the drawings, the same reference numerals are given to the same components, and the description will be omitted as appropriate.

(First embodiment)
FIG. 1 is a functional block diagram showing a configuration of a machine translation system 1 according to an embodiment of the present invention.
A machine translation system 1 according to an embodiment of the present invention is a statistical machine translation system, and uses a model created from text-to-text learning data to perform machine translation from a source language to a target language. To do. Note that the statistical method described below is an example, and the present invention is not limited to this. Other statistical techniques of translation can also be used.

  As shown in FIG. 1, the machine translation system 1 according to the embodiment of the present invention generates a translation model 103a that is a correspondence relationship between a source language and a target language from a source language text corpus 122 and a target language text corpus 124. A translation model generation unit 102 that performs, a syntactic information rule 126 that describes syntactic conditions for determining the application of the translation model 103a, and a rule application unit 106 that determines the application of the translation model 103a according to the syntactic information rule 126 .

The machine translation system 1 according to the embodiment of the present invention includes, for example, a CPU (Central Processing Unit), a memory, a hard disk, and a communication device (not shown), an input device such as a keyboard and a mouse, a display device such as a display, The present invention can be realized by any computer connected to an output device such as a printer, or any combination of hardware and software of devices corresponding to them. These elements of the machine translation system 1 are connected to each other via a bus, and the entire machine translation system 1 is controlled together with the elements by the CPU. When the CPU reads the program according to the embodiment of the present invention stored in the hard disk into the memory and executes it, each function of each unit in each drawing of the machine translation system 1 can be realized. It will be understood by those skilled in the art that there are various modifications to the implementation method and apparatus. Each figure described below shows functional unit blocks, not hardware unit configurations.
Moreover, in each figure, about the structure of the part which is not related to the essence of this invention, it has abbreviate | omitted and is not illustrated.

  Further, the machine translation system 1 of the present invention may be constituted by a plurality of computers or may be constituted by a virtual machine. Alternatively, the present invention provides a machine translation function provided by the machine translation system 1 to a user's various terminals such as a personal computer, a portable terminal, or an electronic device with a communication function via a network using a cloud computer or the like. Can also be included.

  The computer program of the present embodiment is a computer for realizing the machine translation apparatus of the machine translation system 1, and a translation model 103a which is a correspondence relationship between the two languages from the source language text corpus 122 and the target language text corpus 124. In accordance with the syntactic information rule 126 describing the syntactic conditions for determining the application of the translation model 103a, and the procedure for determining the application of the translation model 103a.

  The computer program of this embodiment may be recorded on a computer-readable recording medium. The recording medium is not particularly limited, and various forms can be considered. The program may be loaded from a recording medium into a computer memory, or downloaded to a computer through a network and loaded into the memory.

In the present embodiment, the source language text corpus 122 and the target language text corpus 124 are corpuses of text data of two different languages. The original language of the translation is called the “source language”, and the translated language is called the target language. The source language text corpus 122 and the target language text corpus 124 are so-called bilingual learning corpora in which there is a correspondence between the two languages and the correspondence between the sentences can be taken.
The translation model 103b generated by the machine translation system 1 of this embodiment can also be held as a text data file.

  The translation model generation unit 102 generates a translation model 103a with reference to the source language text corpus 122 and the target language text corpus 124. The translation model generation unit 102 extracts, from the source language text corpus 122 and the target language text corpus 124, for example, a translation rule representing the correspondence relationship between the two languages and the correctness of the correspondence relationship using the EM algorithm. Is generated as a translation model 103a. Here, the translation model 103a including at least one translation rule and a probability value indicating the certainty of the translation rule is generated.

  In the present invention, the “syntactic condition” of the syntactic information rule 126 includes a condition related to the relationship of the arrangement of words and word groups in the sentence. The syntactic information rule 126 can describe the relationship of the arrangement of words and word groups in a sentence, or the words and word groups can be described by parts of speech or surface character strings. For example, the syntactic information rule 126 can describe an application condition such as “the surface character string of the first word is“ no ””. The rule application unit 106 determines whether a translation rule that satisfies this condition is included in the translation model 103a. The syntactic information rule 126 can also specify parts of speech such as “noun” and “article” as conditions. Moreover, the range of the probability value of a translation rule, etc. can also be made into conditions. Furthermore, the contents relating to the function language for the source language or the target language may be described as conditions. Here, the “function word” refers to a non-independent word that has almost no lexical meaning and mainly exhibits a syntactic relationship, and can include prepositions, conjunctions, auxiliary verbs, articles, relatives, and the like. Further, the “contents related to function words” can include position information of words such as the first, last, and Nth words from the beginning.

  The syntactic information rule 126 can be held in the storage unit as a data file in a text format (space delimiter, tab delimiter) or CSV (Comma Separated Values) format, for example. These data files are preferably created or edited with a general text editor or the like. That is, the machine translation system 1 according to the present embodiment may include an update unit (not shown) that receives an editing operation of the syntactic information rule 126 and updates the syntactic information rule 126 according to the received editing operation.

The rule application unit 106 refers to the syntactic information rule 126 and determines whether the translation rule is appropriate for each translation rule of the translation model 103 a generated by the storage device 120 according to the syntactic information rule 126 according to the application condition. Here, it is determined that the translation rule that matches the condition is not applied. The translation rule determined not to be applied may be deleted from the translation model 103a and used as the translation model 103b. Alternatively, the translation model 103b may be added with application information indicating that the translation rule is not applied. In that case, a decoder (not shown in FIG. 1) described later performs a translation process with reference to the application information of the translation model 103b.
The decoder can perform machine translation processing using the translation model 103b generated by determining the application of the translation rule in the syntactic information rule 126 in this way.

A machine translation method of the machine translation system 1 of the present embodiment configured as described above will be described below. FIG. 2 is a flowchart for explaining the machine translation method of the machine translation system 1 according to the embodiment of the present invention.
The machine translation method in the machine translation system 1 of the present embodiment generates a translation model 103a that is a correspondence relationship between the two languages from the source language text corpus 122 and the target language text corpus 124 (step S103). A syntactic condition for determining the application of 103a is described in the syntactic information rule 126 (step S101), and the application of the translation model 103a is determined according to the syntactic information rule 126 (step S105).

  The machine translation system 1 of the present embodiment configured as described above is realized by an apparatus that implements the functions of the machine translation system 1 operating according to the flow described in FIG.

First, in the syntactic information rule 126, at least syntactic conditions are described and held (step S101).
Then, the translation model generation unit 102 generates a translation model 103a from the source language text corpus 122 and the target language text corpus 124 (step S103). The translation model 103a includes a plurality of translation rules and their probability values.

  Then, the rule application unit 106 determines application of the translation model 103a in accordance with the conditions stored in the syntactic information rule 126 (step S105). Then, as a result of the determination, the translation rule determined not to be applied is deleted and output as the translation model 103b.

  The procedure for preparing the syntactic information rule 126 in step S101 is preferably performed before the translation process starts. Alternatively, the syntactic information rule 126 can be updated at any time by the user or the system administrator during the translation process.

  In this embodiment, first, the syntactic information rule 126 describes conditions for determining whether or not the translation rule of the translation model 103a is appropriate. However, the present invention is not limited to this. For example, an example of performing an operation such as changing a rule according to a condition will be described later.

  As described above, according to the machine translation system 1 according to the embodiment of the present invention, the translation model is applied according to the syntactic information rule having the syntactic condition, so that the accuracy can be improved even without the parsing system. Can achieve good machine translation.

(Second Embodiment)
FIG. 3 is a functional block diagram showing a configuration of the machine translation system 201 according to the embodiment of the present invention.
The machine translation system 201 of the present embodiment is different from the above embodiment in that the translation model is operated according to the conditions described in the syntactic information rule, and machine translation is performed using the generated translation model. It is different in having.

The machine translation system 201 of this embodiment includes a data processing device 100, a storage device 120, an input device 130, and an output device 140.
The data processing apparatus 100 includes a translation model generation unit 102, a language model generation unit 104, and a rule application unit 106 similar to the machine translation system 1 of the above embodiment of FIG.

  The data processing apparatus 100 is a computer that executes the program for realizing the machine translation system of the present invention described in the above embodiment. Also in the present embodiment, the CPU (not shown) of the data processing apparatus 100 executes the computer program of the machine translation system 201 of the present invention, whereby each function of each block in FIG. 3 can be realized.

  The storage device 120 is, for example, a storage device or storage medium that can be read at least from the data processing device 100, and the source language text corpus 122, the target language text corpus 124, and the syntactic information rule 126 described in the above embodiment. Hold. The storage device 120 is preferably a storage device or storage medium that can be read or written by a computer. In FIG. 3, the storage device 120 is shown as a single device, but is not limited to this, and a plurality of storage devices or storage media can be combined.

  The input device 130 is a device for inputting an original text to be translated. The input device 130 is, for example, for receiving an original text data or data file from another input device such as a keyboard or a mouse that is operated when the user inputs an original text to be translated from another device via a network. It includes an interface device, a storage device that stores the original data file, or a medium access device that accesses the storage medium and reads the original data file. Alternatively, the input device 130 may include an interface that receives text data of an authentication result that has been voice-recognized by the voice recognition processing device.

  The output device 140 is a device that receives and outputs the translation result output by the data processing device 100. The output device 140 is, for example, a display device that displays a translation result on a screen, a printer that prints and outputs the translation result, an interface device for transmitting the translation result to another device via a network, or a storage device that stores the translation result Alternatively, a medium access device that stores data in a storage medium is included. Alternatively, the output device 140 may include an interface that outputs a translation result to speech data that is read out after being converted into speech data.

  Further, for example, the original input data 101 received by the input device 130 may be displayed side by side on the screen of the display device of the output device 140 so that it can be compared with the translation result 107. Alternatively, the original input data 101 received by the input device 130 can be read out by the speech synthesizer from the output device 140 for each phrase together with the translation result 107 in order. Moreover, the input device 130 and the output device 140 can also be comprised by a touch panel etc., for example, and can also implement | achieve the function as a user interface part.

  The language model generation unit 104 refers to the target language text corpus 124 and generates a language model 105. The language model generation unit 104 obtains at least one translation candidate and a probability value representing the linguistic correctness of the translation candidate from the target language text corpus 124, and generates, for example, an N-gram probability language model 105. To do. Here, a language model 105 including at least one translation candidate is generated.

  The syntactic information rule 126 may describe operation information corresponding to the condition for the translation rule of the translation model 103a and its probability value in each entry in association with the condition.

  The operation information includes, for example, the deletion or addition of a translation rule, or the change of the probability value of the translation rule. The probability value itself to be changed or a plus / minus signed probability value (change value of probability value) for adjusting the increase / decrease of the probability value may be designated. You can also specify the translation rule to add and its probability value.

  FIG. 5 shows a description example of the syntactic information rule 126. For example, the condition that the surface character string of the first word is “no” and the operation information “delete” are, as shown in FIG. 5A, the text data composed of three data (character strings). It can be described as one entry. The syntactic information rule 126 can include a plurality of entries.

  As shown in FIG. 5B, the syntactic information rule 126 can describe the condition as a plurality of conditions, for example, “when condition 1 is condition 2”. The syntactic information rule 126 includes an operation character string 222 indicating “deletion” of the operation information, a first condition character string 226 indicating “first” of the condition 1, and a second condition character indicating “no” of the condition 2. Column 228 can be described as one entry. Separation characters 224 such as spaces and tabs are inserted between the operation character string 222 and the condition character strings 226 and 228, respectively. As described above, the syntactic information rule 126 can be described as text data having a simple configuration in which a plurality of data (character strings) are separated by the delimiter 224. In the present embodiment, three data are used for description, but the present invention is not limited to this.

Further, as shown in FIG. 5C, the first data of the entry of the syntactic information rule 126 (the operation character string 222 in FIG. 5B) includes not only “delete” but also “add” and “ Other operation information such as “change of probability value” may be described.
For example, “addition” can be described as a format such as “additional translation rule probability value”. In this case, when the “translation rule” described in the first condition character string 226 (FIG. 5B) is not included in the translation model 103a, the “translation rule” is used as the second condition character string 228. This means that the “probability value” described in FIG. 5B is added to the translation model 103b.

In the case of “change probability value”, it can be described as a format such as “change probability value change translation rule probability value”. In this case, when the condition of the “translation rule” corresponds to the translation model, the probability value of the “translation rule” is changed by the “probability value”.
Further, the “probability value” can describe an unsigned probability value for replacing the probability value and a +/− signed probability value for increasing or decreasing the probability value.

Returning to FIG. 3, the rule application unit 106 operates the corresponding translation model according to the operation information of the syntactic information rule 126 according to the condition of the syntactic information rule 126.
Specifically, the rule application unit 106 determines whether each translation rule of the translation model 103 a generated by the translation model generation unit 102 matches the condition described in the syntactic information rule 126.

  Then, for example, when the rule application unit 106 determines that the translation rule of the translation model 103a corresponds to an entry described in the syntactic information rule 126, the rule application unit 106 follows the operation information described in the entry. Manipulate. That is, in this embodiment, the rule application unit 106 detects a translation rule that matches the condition described in the entry of the syntactic information rule 126, and the operation information described in the entry for the translation rule that matches the condition. Follow the instructions.

  The translation unit 110 converts the input data 101 of the source language input by the input device 130 into the translation model 103b after the rule application unit 106 is operated according to the conditions and the language model 105 generated by the language model generation unit 104. The result is translated into a target language, and the translation result 107 is output to the output device 140.

The operation of the machine translation system 201 of this embodiment configured as described above will be described below.
FIG. 4 is a flowchart showing an example of the operation of the machine translation system 201 of this embodiment. Hereinafter, a description will be given with reference to FIGS. 3 and 4.

  First, syntactic conditions are described and prepared in the syntactic information rule 126, and the syntactic information rule 126 is stored in the storage device 120 (step S201). Note that the procedure for preparing the syntactic information rule 126 in step S201 is preferably performed in advance before the translation process is started, as in the above embodiment. Alternatively, the syntactic information rule 126 can be updated at any time by the user or the system administrator during the translation process.

  Then, the translation model generation unit 102 of the data processing apparatus 100 accesses the storage device 120, refers to the source language text corpus 122 and the target language text corpus 124, and generates a translation model 103a (step S203). Here, a plurality of translation rules and the probability values of the rules are generated as the translation model 103a.

  Further, the language model generation unit 104 of the data processing device 100 accesses the storage device 120, refers to the target language text corpus 124, and generates the language model 105 (step S205). Here, at least one translation candidate is generated as the language model 105. In the present embodiment, the order in which step S203 and step S205 are executed is not particularly limited. Further, the source language text corpus 122, the target language text corpus 124, or the syntactic information rule 126 can be arbitrarily updated.

  Then, the rule application unit 106 accesses the storage device 120, refers to the syntactic information rule 126 prepared in step S205, and sets each translation rule of the translation model 103a generated in step S201 in the entry of the syntactic information rule 126. It is determined whether or not the described conditions are met (step S207). For the translation rule that the rule application unit 106 determines to match the condition (YES in step S207), the rule application unit 106 operates the translation rule according to the operation information described in the entry that matches the condition (step S209). The operated translation rule is included in the translation model 103b and delivered to the translation unit 110. On the other hand, the translation rule that is determined not to match the condition (NO in step S207) is still included in the translation model 103b that is delivered to the translation unit 110.

  In this figure, the procedure for repeating the determination in step S207 for each translation rule is omitted. Steps S207 and S209 are repeatedly executed for the translation rules included in the translation model 103a, and when the processing for all the translation rules is completed, the translation model 103b including the translation rules operated according to the conditions is converted into the translation unit 110. And the process proceeds to step S211.

  Then, based on the determination result by the rule application unit 106 in step S207, the translation unit 110 includes the translation model 103b including the translation rule operated according to the condition in step S209, and the language model generated by the language model generation unit 104 in step S203. Based on 105, a translation result 107 for the input data 101 is generated and output to the output device 140 (step S211). The output device 140 displays the translation result 107 on the screen of a display device (not shown). Note that the language model 105 and the translation model 103b used by the translation unit 110 for decoding in step S211 can also be stored in the storage device 120.

  As described above, according to the machine translation system 201 according to the embodiment of the present invention, the same effect as the above embodiment can be obtained, and even when the parsing system is not used, the text data (source language text corpus) The translation model 103a created by using 122 and the target language text corpus 124) as learning corpora can be translated with high accuracy by using syntactic information.

  That is, according to the machine translation system 201 according to the embodiment of the present invention, even when the source language text corpus 122 or the target language text corpus 124 does not include syntactic information (syntax information rule 126), the generated translation model 103a. Can be enhanced as the translation model 103b, so that a highly accurate translation result can be obtained.

  Further, as compared with the case where a grammar rule necessary for a syntax analysis system for creating syntax tree data as a learning corpus is created, the translation model 103a that is a problem in the translation result is changed as in the present invention. When the syntactic information (syntax information rule 126) is described, the amount of work of the developer or programmer and the memory capacity necessary for storing the rule are reduced. Therefore, a machine translation system can be constructed at a low cost.

As mentioned above, although embodiment of this invention was described with reference to drawings, these are the illustrations of this invention, Various structures other than the above are also employable.
For example, the condition of the translation rule of the second and third data (condition 1, condition 2) of the syntactic information rule entry given as an example above is not only “start” but also “end” or “start from In addition to location information such as “Nth” and the designation of the surface layer of words such as “no”, part of speech such as “noun” can be described.

  Moreover, in the said embodiment, although the condition of the translation rule regarding the target language which the translation model production | generation part 102 produced | generated was described in the syntactic information rule 126, it is not limited to this. For example, the syntactic information rule 126 can include a condition for the source language, or a condition for the target language, or both a condition for the source language and a condition for the target language.

  In this case, the translation model generation unit 102 can determine whether or not the translation rule of the generated translation model 103a is appropriate according to the syntactic information rule 126 describing the conditions for the source language, or can operate the translation rule according to the operation information. That is, at the stage where the translation model generation unit 102 generates the translation model 103a, it is possible to determine whether or not the translation rule is appropriate or to operate according to the syntactic information rule 126. Further, the rule application unit 106 determines whether or not the translation rule of the translation model 103a is appropriate according to the syntactic information rule 126 in which the condition for the target language is described, or operates the translation rule according to the operation information. May be output.

  The machine translation system of the present invention can be applied to, for example, translation selection.

(Example)
Next, embodiments of the present invention will be described with reference to the drawings. This example shows an example in the case of applying to translation selection in the machine translation system 201 of the above embodiment of FIG.
When an English sentence “Board the train for Misakiguchi at Shinagawa” is input from the keyboard of the input device 130, an example of translating the English sentence into Japanese is taken as an example. The operation will be described.
The translation model generation unit 102 generates a translation model 103a from the source language text corpus 122 and the target language text corpus 124 using the EM algorithm (step S203 in FIG. 4).
Then, the language model generation unit 104 generates a language model 105 with an N-gram probability from the target language text corpus 124 (step S205 in FIG. 4).

  As the translation model 103a related to this input sentence, for example, as shown in FIG. 6, "Board X → Managing X", "Board X → Get on X", "the train for → train", "the Train for X → Train to X ”,“ Misakiguchi → Misakiguchi ”,“ Xat → X for X ”,“ X at Y → Y for X ”,“ Shinagawa → Shinagawa ”and each rule Assume that a probability value (not shown) has been generated.

  Here, “X” and “Y” indicate that an arbitrary word string is included, and “Board X → Managing X” as the first translation rule is a sequence of “Board” and “arbitrary word string (X)”. Indicates that translation is performed in the order of “management” and “translation of an arbitrary input word string”.

Next, the rule application unit 106 operates the translation model 103a generated by the translation model generation unit 102 in step S203 based on the syntactic information rule 126 (steps S207 and S209 in FIG. 4).
For example, as shown in FIG. 5C, it is assumed that the syntactic information rule 126 includes an entry having three pieces of data “deletion head“ no ””.

As described above, the second and third data of this entry represent condition 1 and condition 2 of the translation rule, and the first data represents the processing operation for the translation rule corresponding to condition 1 and condition 2. Represents.
In other words, the second “first” and third ““ of ”” data represents the condition when the surface character string of the “first” word of the translation rule is “no”, and this condition is met. When there is a translation rule to be performed, this means that the translation rule 210 is “deleted”.
At this time, since this entry corresponds to “the train for → train” (FIG. 6) of the translation rule 210 of the translation model 103a, the rule application unit 106 deletes the translation rule from the translation model 103a.

  Then, the translation unit 110 uses the translation model 103b from which the translation rule determined not to be applied according to the condition by the rule application unit 106 is used, and is analyzed by, for example, a CYK parser. Then, an analysis result as shown in FIG. 7 is obtained, and only one translation result “get on a train bound for Misakiguchi in Shinagawa” is obtained. Then, the translation unit 110 outputs the obtained translation result from the output device 140 and displays it on the display.

(Comparative example)
Next, as a comparative example of the machine translation system of the present invention, a text-to-text tree-based statistical translation system that does not include the syntactic information rule 126 and the rule application unit 106 of the present invention will be described below.
In this system, the translation unit translates input data and outputs a translation result using the translation model created by the translation model generation unit and the language model generated by the language model generation unit from the target language text corpus.

At this time, in the translation unit, when the input sentence is analyzed by the CYK parser using these translation rules, an analysis result as shown in FIG. 8 is obtained. When this is translated into Japanese using the translation rules used, two translation results are obtained: “Make a Misakiguchi Shinagawa on the managing train” and “Get on the train bound for Misakiguchi in Shinagawa”.
When multiple translation results are obtained, the translation unit's probability value of the translation model used to obtain each translation result and the target language's language model's probability value are multiplied by the translation unit. Suppose that the one with the larger value is selected.

  At this time, when a syntactically unnecessary translation rule is included, the value of the translation result using this unnecessary rule (in this case, “Misakiguchi Shinagawa on the managing train”) may increase. Thus, in the comparative example, the accuracy may be deteriorated.

  In the present invention, accurate translation is realized by changing the translation rule by the syntactic information rule 126 in advance in order to deal with a case like the comparative example. For example, in the present invention, there is always a word before the particle “no”, and “no” is related to that word, so there is no need for a rule that has nothing before “no”. From this point of view, an entry of “deletion head“ no ”” is created in the syntactic information rule 126. The rule application unit 106 changes the translation model 103a corresponding to this entry, so that the translation can be performed with high accuracy.

  While the present invention has been described with reference to the embodiments and examples, the present invention is not limited to the above embodiments and examples. 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.

In addition, this invention can also include the following aspects.
(Appendix 1)
Machine translation device
From the source language text corpus and the target language text corpus, generate a translation model that is the correspondence between the two languages.
Describe syntactic conditions for determining the application of the translation model in the syntactic information rule,
A machine translation method for determining application of the translation model according to the syntactic information rule.
(Appendix 2)
In the machine translation method described in (Appendix 1),
The syntactic information rule is described by associating operation information for the translation model corresponding to the condition with the condition,
A machine translation method in which the machine translation device changes a corresponding translation model according to the operation information of the syntactic information rule according to the syntactic information rule.
(Appendix 3)
In the machine translation method described in (Appendix 1) or (Appendix 2),
The syntactic information rule is described by associating operation information for deleting a translation model corresponding to the condition with the condition,
A machine translation method in which the machine translation device deletes a corresponding translation model according to the syntactic information rule.
(Appendix 4)
In the machine translation method according to any one of (Appendix 1) to (Appendix 3),
The syntactic information rule includes a condition for a source language or a condition for a target language.
(Appendix 5)
In the machine translation method according to any one of (Appendix 1) to (Appendix 4),
The syntactic information rule is a machine translation method in which a part of speech or a surface character string for a source language or a target language can be described as the condition.
(Appendix 6)
In the machine translation method according to any one of (Appendix 1) to (Appendix 5),
The syntactic information rule is a machine translation method in which contents relating to a function word for a source language or a target language are described as the condition.
(Appendix 7)
In the machine translation method according to any one of (Appendix 1) to (Appendix 6),
The machine translation device is
Generating a language model indicating the linguistic correctness of the language from the target language text corpus;
Translation means for translating an input text using the translation model and the language model;
A machine translation method further comprising:
(Appendix 8)
In the machine translation method according to any one of (Appendix 1) to (Appendix 7),
The machine translation method, wherein the machine translation device further receives an editing operation of the syntactic information rule and updates the syntactic information rule according to the received editing operation.
(Appendix 9)
A procedure for generating a translation model corresponding to the correspondence between the two languages from the source language text corpus and the target language text corpus;
A machine translation program for causing a computer to execute a procedure for judging application of the translation model according to a syntactic information rule describing a syntactic condition for judging application of the translation model.
(Appendix 10)
In the machine translation program described in (Appendix 9),
The syntactic information rule is described by associating operation information for the translation model corresponding to the condition with the condition,
A machine translation program for causing a computer to execute a procedure for changing a corresponding translation model according to the operation information of the syntactic information rule according to the syntactic information rule.
(Appendix 11)
In the machine translation program described in (Appendix 9) or (Appendix 10),
The syntactic information rule is described by associating operation information for deleting a translation model corresponding to the condition with the condition,
A machine translation program for causing a computer to execute a procedure for deleting a corresponding translation model in accordance with the syntactic information rule.
(Appendix 12)
In the machine translation program according to any one of (Appendix 9) to (Appendix 11),
The syntactic information rule is a machine translation program including a condition for a source language or a condition for a target language.
(Appendix 13)
In the machine translation program according to any one of (Appendix 9) to (Appendix 12),
The syntactic information rule is a machine translation program in which a part of speech or a surface character string for a source language or a target language can be described as the condition.
(Appendix 14)
In the machine translation program according to any one of (Appendix 9) to (Appendix 13),
The syntactic information rule is a machine translation program that describes, as the condition, contents related to a function word for a source language or a target language.
(Appendix 15)
In the machine translation program according to any one of (Appendix 9) to (Appendix 14),
Generating a language model indicating the linguistic correctness of the language from the target language text corpus;
A machine translation program for causing a computer to further execute a procedure for translating an input text using the translation model and the language model.
(Appendix 16)
In the machine translation program according to any one of (Appendix 9) to (Appendix 15),
A machine translation program for receiving a syntactic information rule editing operation and causing a computer to further execute a procedure for updating the syntactic information rule according to the received editing operation.
(Appendix 17)
A translation model generating means for generating a translation model that is a correspondence relationship between the source language and the target language from the source language text corpus and the target language text corpus;
A syntactic information rule describing a syntactic condition for determining application of the translation model;
Rule application means for determining application of the translation model according to the syntactic information rule;
A machine translation system comprising:
(Appendix 18)
In the machine translation system described in (Appendix 17),
The syntactic information rule describes operation information for a translation model corresponding to the condition in association with the condition,
The machine application system, wherein the rule applying unit changes a corresponding translation model according to the operation information of the syntactic information rule according to the syntactic information rule.
(Appendix 19)
In the machine translation system described in (Appendix 17) or (Appendix 18),
The syntactic information rule describes operation information for deleting a translation model corresponding to the condition in association with the condition,
The said rule application means is a machine translation system which deletes a corresponding translation model according to the said syntactic information rule.
(Appendix 20)
In the machine translation system according to any one of (Appendix 17) to (Appendix 19),
The syntactic information rule is a machine translation system including a condition for a source language or a condition for a target language.
(Appendix 21)
In the machine translation system according to any one of (Appendix 17) to (Appendix 20),
The syntactic information rule is a machine translation system in which a part of speech or a surface character string for a source language or a target language can be described as the condition.
(Appendix 22)
In the machine translation system according to any one of (Appendix 17) to (Appendix 21),
The syntactic information rule is a machine translation system in which contents relating to a function word for a source language or a target language are described as the condition.
(Appendix 23)
In the machine translation system according to any one of (Appendix 17) to (Appendix 22),
Language model generation means for generating a language model indicating the linguistic correctness of the language from the target language text corpus;
Translation means for translating an input text using the translation model and the language model;
A machine translation system further comprising:
(Appendix 24)
In the machine translation system according to any one of (Appendix 17) to (Appendix 23),
A machine translation system further comprising an updating unit that receives an editing operation of the syntactic information rule and updates the syntactic information rule according to the received editing operation.

DESCRIPTION OF SYMBOLS 1 Machine translation system 100 Data processing apparatus 101 Input data 102 Translation model production | generation part 103a Translation model 103b Translation model 104 Language model production | generation part 105 Language model 106 Rule application part 107 Translation result 110 Translation part 120 Storage device 122 Source language text corpus 124 Purpose Language text corpus 126 Syntactic information rule 130 Input device 140 Output device 201 Machine translation system 210 Translation rule 222 Operation character string 224 Separator character 226 Condition character string 228 Condition character string

Claims (10)

At least one translation rule obtained by extracting a translation rule indicating the correspondence between the source language and the target language and the correctness of the correspondence from the source language text corpus and the target language text corpus, and the translation rule A translation model generating means for generating a translation model including a probability value indicating the probability of
A syntactic information rule describing a syntactic condition used to change the translation model;
Rule applying means for changing the translation model according to the condition of the syntactic information rule;
A machine translation system comprising: The machine translation system according to claim 1,
The syntactic information rule describes operation information for a translation model corresponding to the condition in association with the condition,
The said rule application means is a machine translation system which changes the said translation rule which matches the said conditions of the said syntactic information rule according to the said operation information of the said syntactic information rule. The machine translation system according to claim 1 or 2,
The syntactic information rule describes operation information for deleting a translation rule that matches the condition, in association with the condition,
The rule application unit is a machine translation system that deletes a translation rule that matches the condition of the syntactic information rule. The machine translation system according to any one of claims 1 to 3,
The syntactic information rule is a machine translation system including a condition for a source language or a condition for a target language. The machine translation system according to any one of claims 1 to 4,
The syntactic information rule is a machine translation system in which a part of speech or a surface character string for a source language or a target language can be described as the condition. The machine translation system according to any one of claims 1 to 5,
The syntactic information rule is a machine translation system in which contents relating to a function word for a source language or a target language are described as the condition. The machine translation system according to any one of claims 1 to 6,
Language model generation means for generating a language model indicating the linguistic correctness of the language from the target language text corpus;
Translation means for translating an input text using the translation model and the language model;
A machine translation system further comprising: The machine translation system according to any one of claims 1 to 7,
A machine translation system further comprising an updating unit that receives an editing operation of the syntactic information rule and updates the syntactic information rule according to the received editing operation. Machine translation device
From the source language text corpus and the target language text corpus, a translation rule indicating the correspondence between the two languages and the correctness of the correspondence are extracted, and at least one obtained translation rule and the translation rule Generate a translation model that includes a probability value indicating the probability ,
Describe the syntactic conditions used to change the translation model in the syntactic information rules,
A machine translation method for changing the translation model according to the condition of the syntactic information rule. From the source language text corpus and the target language text corpus, a translation rule indicating the correspondence between the two languages and the correctness of the correspondence are extracted, and at least one obtained translation rule and the translation rule A procedure for generating a translation model including a probability value indicating the probability ,
A machine translation program for causing a computer to execute a procedure of changing the translation model according to the condition of the syntactic information rule describing a syntactic condition used for changing the translation model.

Images