JP2007080221A - Apparatus, method and program for machine translation - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an apparatus, a method and a program for machine translation, which may easily translate a document by using dictionary data for the machine translation. <P>SOLUTION: The method and the program for the machine translation, comprising the steps of: generating a featured vector by using a featured vector generating unit 10 for a machine translation object representing a feature for the document to be translated; generating a specified field vector by using a featured vector generating unit 20 for the machine translation object representing a feature for a specified field of the document; estimating a degree of belonging to each specified field for the documents to be translated base on the said vectors by using a category sorting unit 50; and selecting a specified dictionary data, which is fittest to translate the document, based on a result obtained through the above step for estimation. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a machine translation apparatus and a machine translation method for inputting a translation target sentence and performing translation using dictionary data.

  In recent years, machine translation devices that input a translation target sentence and perform translation using dictionary data have become widespread. Some of these machine translation apparatuses are provided with dictionary data (specialized dictionary data) for each specialized field such as politics and economy so as to enable more appropriate translation. In this case, it is important how to select one suitable for the sentence to be translated from among a plurality of specialized dictionary data.

For example, Patent Document 1 proposes a technique in which a dictionary data selection priority is set by a user, and dictionary data is selected and translated based on the priority order. Further, in Patent Document 2, a context vector that indicates in which specialized dictionary data the basic word is registered is assigned to each basic word included in the translation target sentence, and the translation target sentence is divided into a basic word and a specialized field. There has been proposed a technique for converting into a table as an attribute and selecting specialized dictionary data based on this table and a specialized dictionary correspondence table.
JP-A-7-141375 JP-A-6-332946

  However, in the technique of Patent Document 1 described above, the user manually sets the priority order for selecting dictionary data, which is complicated. In the technique disclosed in Patent Document 2, a certain specialized knowledge is required for setting the specialized dictionary correspondence table.

  The present invention has been made in view of the above points, and an object thereof is to provide a machine translation device, a machine translation method, and a machine translation program capable of easily translating using dictionary data suitable for a sentence to be translated. Yes.

  The machine translation device according to the present invention includes a first holding unit that holds a special field feature vector representing a feature of a sentence belonging to each of a plurality of specialized fields, and the translation target sentence based on the translation target sentence. A first creation means for creating a translation target feature vector representing a feature; a translation target feature vector created by the first creation means; and each specialized field feature vector held in the first holding means; Based on the above, the estimation means for estimating the possibility that the translation target sentence belongs to each of the specialized fields, and the specialized dictionary data corresponding to the specialized field to which the translation target sentence is likely to belong are preferentially selected. And a translation unit that translates the sentence to be translated using the specialized dictionary data selected by the selection unit.

  Thereby, specialized dictionary data corresponding to a specialized field to which the translation target sentence is likely to belong is automatically selected, and appropriate translation using the specialized dictionary data becomes possible.

  Preferably, the machine translation apparatus includes a second creation unit that creates the specialized field feature vector based on a sentence corresponding to each of the specialized dictionary data.

  Preferably, in the machine translation device, the first creation unit further creates the translation target feature vector based on a sentence around the translation target sentence.

  Preferably, the machine translation device includes providing means for providing the possibility of belonging to each of the specialized fields of the translation target sentence so as to be visible.

  As a result, the user can grasp which specialized field the translation target sentence may belong to.

  Preferably, in the machine translation apparatus, the providing unit displays a graph representing the possibility of the translation object sentence belonging to each of the specialized fields.

  The machine translation method according to the present invention includes a first creation step for creating a translation target feature vector representing a feature of the translation target sentence based on the translation target sentence, and a translation created by the first creation step. An estimation step for estimating a possibility that the translation target sentence belongs to each of the specialized fields based on the target feature vector and each specialized field feature vector representing a feature of the sentence belonging to each of the plurality of specialized fields; The translation step is translated using a selection step that preferentially selects specialized dictionary data corresponding to a specialized field to which the translation subject sentence is likely to belong, and the specialized dictionary data selected in the selection step A translation step.

  Preferably, the machine translation method includes a second creation step of creating the specialized field feature vector based on a sentence corresponding to each of the specialized dictionary data.

  Preferably, in the machine translation method, the first creation step further creates the translation target feature vector based on a sentence around the translation target sentence.

  Preferably, the machine translation method includes a providing step of providing the possibility of belonging to each of the specialized fields of the translation target sentence so as to be visible.

  Preferably, in the machine translation method, the providing step displays a graph representing the possibility of the translation target sentence belonging to each of the specialized fields.

  A machine translation program according to the present invention includes a first creation step for creating a translation target feature vector representing a feature of a translation target sentence based on the translation target sentence, and a translation created by the first creation step. An estimation step for estimating a possibility that the translation target sentence belongs to each of the specialized fields based on the target feature vector and each specialized field feature vector representing a feature of the sentence belonging to each of the plurality of specialized fields; The translation step is translated using a selection step that preferentially selects specialized dictionary data corresponding to a specialized field to which the translation subject sentence is likely to belong, and the specialized dictionary data selected in the selection step And causing the computer to execute a translation step.

  Preferably, the machine translation program causes the computer to execute a second creation step of creating the specialized field feature vector based on a sentence corresponding to each of the specialized dictionary data.

  Preferably, in the machine translation program, the first creation step further creates the translation target feature vector based on a sentence around the translation target sentence.

  Preferably, the machine translation program causes the computer to execute a providing step of providing the possibility of belonging to each of the specialized fields of the translation target sentence so as to be visible.

  Preferably, in the machine translation program, the providing step displays a graph representing the possibility of the translation target sentence belonging to each of the specialized fields.

  According to the present invention, specialized dictionary data suitable for translation of a translation target sentence is automatically selected, so that appropriate translation using specialized dictionary data can be easily performed.

  The best mode for carrying out the present invention will be described below with reference to the drawings.

  FIG. 1 is a diagram showing a configuration of a machine translation apparatus according to an embodiment of the present invention. A machine translation apparatus 100 shown in FIG. 1 includes a feature vector creation unit 10, a feature vector creation unit 20, a specialized field labeled corpus database (DB) 30, a specialized field representative vector group database (DB) 40, a category classifier 50, and a display. The unit 52, the specialized dictionary selection unit 60, and the translation module 70 are configured.

  The feature vector creation unit 10 inputs a translation target sentence. Next, the feature vector creation unit 10 performs morphological analysis, which is a well-known document analysis technique, on the input translation target sentence, and divides the translation target sentence into words. Furthermore, the feature vector creation unit 10 counts the appearance frequency (TF) for each word. Then, the feature vector creation unit 10 creates a vector (translation target feature vector) representing each feature of the translation target sentence by including each numerical value of the appearance frequency for each word as a one-dimensional element corresponding to the word. To do. Here, the number of dimensions of the translation target feature vector and the word corresponding to each dimension are determined in advance and coincide with a specialized field single feature vector and a specialized field representative feature vector, which will be described later.

  FIG. 2 is a diagram illustrating an example of creating a translation target feature vector. When the feature vector creation unit 10 inputs a sentence to be translated “US President Bush and a US Army reorganization problem at a hotel in New York City”, the morphological analysis is performed on the sentence to be translated and the word is divided into words. To do. Further, here, the feature vector creation unit 10 extracts only nouns from words obtained by morphological analysis. That is, the noun words “Bush”, “US”, “President”, “New York”, “city”, “hotel”, “resident in Japan”, “US”, “military”, “reorganization”, “problem” Is extracted.

  Next, the feature vector creation unit 10 counts the appearance frequency for the extracted words. Here, the words “Bush”, “President”, “New York”, “city”, “hotel”, “resident in Japan”, “army”, “reorganization”, and “problem” appear once each. Therefore, these words are counted as “1”, and since the word “rice” appears twice, this word is counted as “2”.

  Further, the feature vector creation unit 10 corresponds to the words “bush”, “president”, “New York”, “city”, “hotel”, “resident in Japan”, “military”, “reorganization”, “problem”. For the dimension, the element is “1”, for the dimension corresponding to the word “rice”, the element is “2”, and for the dimension corresponding to words other than these, that is, words that do not appear in the sentence to be translated A feature vector to be translated is created with “0” being set.

  Note that the feature vector creation unit 10 does not use the numerical value of the appearance frequency as an element of the feature vector to be translated as it is for a predetermined word such as a word considered important, for example. Weighting may be performed by, for example, using a value obtained by multiplying the weighting coefficient as an element of the translation target feature vector.

  When a sentence around the translation target sentence (peripheral sentence) is input together with the translation target sentence, the feature vector creation unit 10 describes the sentence set including both the translation target sentence and the peripheral sentence as described above. A translation target feature vector may be created by a similar method. The translation target feature vector created by the feature vector creation unit 10 is sent to the category classifier 50.

  On the other hand, the feature vector creation unit 20 reads a corpus with specialized field labels from the corpus DB 30 with specialized fields label. A corpus with a specialized field label is obtained by adding a label, which is identification information of a specialized field, to a sentence of a specialized field such as politics or economy (specialized field sentence). The specialized field labeled corpus DB 30 does not necessarily have to be configured inside the machine translation apparatus 10, and the feature vector creation unit 20 uses a specialized field labeled corpus acquired using an information providing service provided by a vendor. It is also possible to collect a specialized field sentence from an Internet site and add a label to the specialized field sentence as a corpus with a specialized field label. In addition, when the dictionary data described later includes a commentary of a word, the feature vector creation unit 20 sets the commentary sentence as a specialized field sentence and adds a label to the corpus with a specialized field label. May be used as

  Next, the feature vector creating unit 20 is a vector that represents the feature of the specialized field sentence from the specialized field sentence in the read corpus with the specialized field label (similar to the creation of the translation target feature vector by the feature vector creating unit 10 described above). Special field single feature vector). That is, the feature vector creation unit 20 performs morphological analysis on the specialized field sentence and divides it into words, and counts the appearance frequency for each word. Furthermore, the feature vector creation unit 10 creates a specialized field single feature vector that includes each of the numerical values of the appearance frequencies for each word as a one-dimensional element corresponding to the word. The feature vector creation unit 20 adds a label in the specialized field-labeled corpus used for the creation to the created specialized field single feature vector, and stores it in a built-in storage unit (not shown).

  Further, the feature vector creating unit 20 uses a specialized field single feature vector in the built-in storage unit to obtain a specialized field representative feature vector representing a specialized field single feature vector corresponding to the specialized field for each specialized field. It is created and registered in the specialized field representative feature vector group DB 40.

  Specifically, the feature vector creation unit 20 classifies the specialized field single feature vector in the built-in storage unit for each specialized field based on the label to which the specialized field single feature vector is added. Further, when there is only one specialized field single feature vector corresponding to a predetermined specialized field, the feature vector creating unit 20 directly uses the single specialized field single feature vector as the specialized field representative feature vector. Along with the added label, it is registered in the specialized field representative feature vector group DB 40.

  On the other hand, when there are a plurality of specialized field single feature vectors corresponding to a predetermined specialized field, the feature vector creating unit 20 creates specialized field representative feature vectors representing the plurality of specialized field single feature vectors. . For example, the feature vector creation unit 20 calculates an average value for elements of the same dimension of a plurality of specialized field single feature vectors, and creates a specialized field representative feature vector using these average values as elements. Further, the feature vector creation unit 20 registers the created specialized field representative feature vector in the specialized field representative feature vector group DB 40 together with the label added to the specialized field single feature vector used for the creation.

  The category classifier 50 receives the feature vector to be translated from the feature vector creation unit 10, and stores the specialized field representative feature vector for each specialized field registered in the specialized field representative feature vector group DB 40 and the specialized field representative. Read the label corresponding to the feature vector.

  Next, the category classifier 50 estimates the possibility that the translation target sentence belongs to each of the specialized fields based on the translation target feature vector and each specialized field representative feature vector.

  FIG. 3 is a diagram illustrating a specific example of the processing of the category classifier 50. For each specialized field, the category classifier 50 has a specialized function representative feature vector corresponding to the specialized field as a positive example, and a specialized function representative feature vector corresponding to a specialized field other than the specialized field as a negative example. Create This distribution function calculates the degree of approximation between the translation target feature vector and the specialized field representative feature vector used as a positive example, in other words, the possibility that the translation target sentence belongs to the specialized field corresponding to the classification function. Is a function for In FIG. 3, a classification function is created for each specialized field of “sports”, “computer”, “economy”, “politics”, “history”, and “science”.

  Next, the category classifier 50 substitutes the element of the translation target feature vector for each classification function. As a result of this substitution, the category classifier 50 has a positive classification function value. The larger the value, the higher the possibility that the translation target sentence belongs to the specialized field corresponding to the classification function (assignment possibility). It is determined that the value of the classification function is negative, and the smaller the value is, the smaller the possibility that the translation target sentence belongs to the specialized field corresponding to the classification function.

  Next, the category classifier 50 creates a category classification result by adding a label corresponding to the specialized field representative feature vector used as a positive example when creating the classification function to the value of the classification function. Further, the category classifier 50 causes the display unit 52 to display a graph indicating the possibility of belonging to each specialized field of the translation target sentence based on the category classification result. FIG. 3 shows a graph indicating that the translation target sentence is more likely to belong in the order of specialized fields “economics”, “politics”, “history”, “science”, “computer”, and “sports”. Further, the category classifier 50 outputs the category classification result to the specialized dictionary selection unit 60.

  When the specialized dictionary selection unit 60 receives the category classification result from the category classifier 50, the specialized dictionary selection unit 60 determines the possibility of belonging to each specialized field of the translation target sentence based on the value of the dispersion function and the label included in the category classification result. recognize. Further, the specialized dictionary selection unit 60 selects a specialized field in which the translation target sentence is highly likely to belong to among specialized dictionary data (professional dictionary data) stored in a built-in storage unit (not shown). Preferentially select corresponding specialized dictionary data.

  In the example of FIG. 3, the selection priority is the order of specialized dictionary data corresponding to the specialized field to which the translation target sentence is highly likely to belong, that is, specialized dictionary data of the specialized field “economics”, specialized dictionary data of “politics”, The order is "history" specialized dictionary data, "science" specialized dictionary data, "computer" specialized dictionary data, and "sports" specialized dictionary data. Therefore, the specialized dictionary selection unit 60 selects the specialized dictionary data of the specialized field “Economy” with the highest priority.

  The specialized dictionary selection unit 60 may select a plurality of specialized dictionary data. For example, the specialized dictionary selection unit 60 provides a threshold value for the dispersion function value, identifies a value of the dispersion function included in the category classification result that is equal to or greater than the threshold value, and adds the value to the identified dispersion function value. Select specialized dictionary data in the specialized field corresponding to the label.

  The specialized dictionary data selected by the specialized dictionary selection unit 60 is sent to the translation module 70. The translation module 70 translates the translation target sentence using the input specialized dictionary data.

  Thus, in the machine translation apparatus 100, the feature vector creation unit 10 creates a translation target feature vector that represents the feature of the translation target sentence, and the feature vector creation unit 20 uses the specialized field feature that represents the feature of the specialized field sentence. Create a vector. Then, based on these, the category classifier 50 estimates the possibility of belonging to each specialized field of the translation target sentence according to the value of the classification function, and the specialized dictionary selection unit 60, based on the estimation result. The specialized dictionary data suitable for the translation of the translation target sentence is selected, and the translation using the dictionary data suitable for the translation target sentence becomes possible. In addition, since the possibility of belonging to each specialized field of the translation target sentence is displayed in a graph, the user can grasp which specialized field the translation target sentence may belong to. When the specialized dictionary data selected by the specialized dictionary selection unit 60 cannot be properly translated, the specialized dictionary data corresponding to another specialized field to which the translation target sentence is highly likely to belong is selected again. Is also possible.

  As described above, the machine translation device, the machine translation method, and the machine translation program according to the present invention can be easily translated appropriately using specialized dictionary data, and are useful as machine translation devices and the like.

It is a figure which shows the structure of a machine translation apparatus. It is a figure which shows an example of translation object feature vector preparation. It is a figure which shows the specific example of a process of a category classifier.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Machine translation apparatus 10, 20 Feature vector creation part 30 Corpus database with special field label 40 Special field representative vector group database 50 Category classifier 52 Display part 60 Special dictionary selection part 70 Translation module

Claims (15)

First holding means for holding special field feature vectors representing features of sentences belonging to each of a plurality of special fields;
First creation means for creating a translation target feature vector representing a feature of the translation target sentence based on the translation target sentence;
The translation target sentence belongs to each of the specialized fields based on the translation target feature vector created by the first creating unit and each specialized field feature vector held in the first holding unit. An estimation means for estimating the possibility;
Selection means for preferentially selecting specialized dictionary data corresponding to a specialized field to which the translation target sentence is likely to belong;
A machine translation apparatus comprising translation means for translating the sentence to be translated using the specialized dictionary data selected by the selection means. The machine translation apparatus according to claim 1, further comprising a second creation unit that creates the specialized field feature vector based on a sentence corresponding to each of the specialized dictionary data. The machine translation apparatus according to claim 1, wherein the first creation unit further creates the translation target feature vector based on sentences around the translation target sentence. The machine translation device according to claim 1, further comprising a providing unit that provides visibility of the translation target sentence to each of the specialized fields. 5. The machine translation apparatus according to claim 4, wherein the providing means displays a graph representing the possibility of belonging to the specialized field of the translation target sentence. A first creation step of creating a translation target feature vector representing a feature of the translation target sentence based on the translation target sentence;
Based on the translation target feature vector created in the first creation step and each specialized field feature vector representing the feature of the sentence belonging to each of a plurality of specialized fields, the translation target sentence is each of the specialized fields. An estimation step for estimating the likelihood of belonging to,
A selection step of preferentially selecting specialized dictionary data corresponding to a specialized field to which the translation target sentence is likely to belong;
A machine translation method comprising: a translation step of translating the sentence to be translated using the specialized dictionary data selected in the selection step. The machine translation method according to claim 6, further comprising a second creation step of creating the specialized field feature vector based on a sentence corresponding to each of the specialized dictionary data. The machine translation method according to claim 6 or 7, wherein the first creation step further creates the translation target feature vector based on a sentence around the translation target sentence. 9. The machine translation method according to claim 6, further comprising a providing step of providing the possibility of belonging to each of the specialized fields of the translation target sentence so as to be visible. The machine translation method according to claim 9, wherein the providing step displays a graph indicating the possibility of belonging to each of the specialized fields of the translation target sentence. A first creation step of creating a translation target feature vector representing a feature of the translation target sentence based on the translation target sentence;
Based on the translation target feature vector created in the first creation step and each specialized field feature vector representing the feature of the sentence belonging to each of a plurality of specialized fields, the translation target sentence is each of the specialized fields. An estimation step for estimating the likelihood of belonging to,
A selection step of preferentially selecting specialized dictionary data corresponding to a specialized field to which the translation target sentence is likely to belong;
A machine translation program that causes a computer to execute a translation step of translating the sentence to be translated using the specialized dictionary data selected in the selection step. The machine translation program according to claim 11, wherein the computer executes a second creation step of creating the specialized field feature vector based on a sentence corresponding to each of the specialized dictionary data. The machine translation program according to claim 11 or 12, wherein the first creation step further creates the translation target feature vector based on a sentence around the translation target sentence. The machine translation program according to any one of claims 11 to 13, wherein the computer executes a providing step of providing the possibility of belonging to each of the specialized fields of the translation target sentence in a visible manner. The machine translation program according to claim 14, wherein the providing step displays a graph representing the possibility of belonging to each of the specialized fields of the translation target sentence.

Images