JPS59103171A - Mechanical translator - Google Patents

Abstract

PURPOSE:To attain selection of translated multivocal words by comparing a subject field vector with the field vector of each translated word and selecting the translated word having the most coincident translated word field vector. CONSTITUTION:A field vector synthesizing part 11 adds successively the field vector V of each word prepared for a buffer 22 by a vector adder 23 and stores the result vector S to a subject field vector register 24. At the same time, a translated word selecting part 7 actuates an inner product arithmetic part 12, reads a translated word field vector out of the buffer 22 with a switching process 25 to give it to an inner product arithmetic part 26 and carries out an inner product operation between the vector S of the register 24 and the translated word field vector. A comparing part 13 compares the vector S with the translated word field vector by a maximum detector 27 to detect the maximum value. A translated word deciding part 14 sets the translated word code to an output buffer 29 with extraction 28 of translated word for the translated word corresponding to the maximum translated word field vector value. The part 7 repeats the same processing for every word of the original sentence and extracts all translated words to deliver them through an output part 8.

Description

Detailed Description of the Invention (5) Technical Field of the Invention The present invention relates to a machine translation device, particularly a device that uses an electronic computer to translate and output an input original text, which translates and outputs a plurality of translated words for one word appearing in the input original text. The present invention relates to a machine translation device that can relatively easily select the most appropriate translation word when there are many translated words.

+BL Technology background and problems FIG. 1 shows an example of the configuration of a general machine translation device.

In a machine translation device, natural language is input as digital information, and the result is automatically translated, for example, from English to Japanese, and from Japanese to English. In FIG. 1, a control unit 1 fetches and executes instructions to control the entire machine translation apparatus. The original text to be translated inputted from the input unit 2 is divided into words by the division unit 3. For each divided word, the dictionary search unit 4 calculates the number of words.

The word dictionary ■0 prepared in advance in the magnetic disk device 9 or the like is searched. In the single-convex dictionary 10, attributes, translations, etc. are registered for each word of the dictionary to be translated. The word order determining unit 6 analyzes the syntax of parts, etc. Based on the analysis result, the word order determining unit 6 determines the order in which the translated words are arranged when the words are converted into translated words.The translated word selecting unit 7 When there are m translations corresponding to , one of them is selected.The output unit 8 arranges the translations selected by the translation word selection unit 7 according to the word order determined by the word order determination unit 6, Create and output the translation.

The present invention particularly relates to the above-mentioned translation word selection part. For example, when translating the words + I frt and nction J into Japanese, there are some words that have to be translated as "functions," and other words that have to be translated as "functions." In conventional machine translation devices, when a plurality of translations for a certain word can be considered, the translation is selected using a first-order method.

The first method is to output a plurality of translated words corresponding to the word to be translated and nine people select them. However, this method requires one person's effort, which goes against the purpose of machine translation, which is automatic translation.

The second method is 2For example, the field of mathematics, the field of sports, etc.
Separate word dictionaries are prepared for each field in which the original text to be translated is used. According to this method, it is necessary to prepare multiple word dictionaries, which is not only complicated, but also difficult to use because, for example, when translating newspaper articles, the field to be translated is often not known in advance. The range is limited. As another method, a selection method based on advanced semantic processing using so-called knowledge pace has been studied. However, this method requires an extremely large amount of computer processing, so there is a problem that it cannot be used in a large-scale system.

tel OBJECTS AND STRUCTURE OF THE INVENTION The present invention aims to solve the above-mentioned problems, and aims to provide a machine translation device that quickly selects appropriate translated words without using advanced semantic processing. Therefore, 1 the present invention provides a word field vector or a translated word field that indicates, as strength, the compatibility between a word registered in a word dictionary and each translated word corresponding to the word, and each field in which the word or translated word is used. The target field vector obtained by assigning vectors and composing the word field vectors of words that appear in the original text and each translated word field vector are compared to select the translated word with the most matching target field vector. It is.

In other words, the machine translation device of the present invention inputs an original text,
The input original text is divided into words, and translated words corresponding to the original text words registered in the word dictionary are extracted.

In a machine translation device that composes and outputs a translated sentence, the word dictionary has a word field whose dimension is the number of fields predetermined for each registered word, and whose components are the degree of compatibility between the word and each of the above fields. In addition to holding vector information, for each translated word of the registered word, translation word field vector information, which is a vector corresponding to the word field vector and whose component is the degree of compatibility between the translated word and each of the above fields, is stored and translated. a synthesis unit that generates a target domain vector by synthesizing the plurality of word domain vectors obtained from the target word group;

a translation word selection unit that compares the target field vector synthesized by the synthesis unit and the plurality of translation word field vectors for one word, and determines the translation having the most similar translation word field vector as the translation of the word; It is characterized by This will be explained below with reference to the drawings.

+D+ Embodiment of the Invention FIG. 2 is an example of the configuration of a word dictionary according to the present invention, and FIG. 3 and 4 are explanatory diagrams of field vectors.

Diagram N15 shows the configuration of an embodiment of the present invention.

A word dictionary contains each word in the source language to be translated.

The corresponding translation words are registered and stored in advance, and one word has a field structure as shown in FIG. 2, for example. The key 10-1 has key information such as character code strings for words in the source language. The attribute field 10-2 stores general word information such as part-of-speech type ring of the word. The word field vector field 10-3 stores vector information indicating the degree of compatibility with each field in which the word is used, as will be described in detail later. The translation word field 10-4 stores translation words corresponding to the key l0-1 as many as the appropriate number of translation words for the word. Further, in the translation word field vector field 10-5, for each translated word, vector information indicating the degree of compatibility with each field in which the translated word is used is set at the time of registration.

The word field vector, the translation word field vector, and the target field vector, which will be described later, have the elements shown in FIG. The dimension of the vector is determined by classifying the fields in which the word is widely used, and the number of possible fields is determined as the dimension. And each element of the vector is.

For example, the first field at the beginning is the field of "mathematics", and the second field is "
The content of the field is commonly defined for each vector, such as the field of "sports".

For example, when translating the English word [τeαctorJ] into Japanese, in some cases it is appropriate to translate it as ``nuclear reactor j,'' and in other cases it is necessary to translate it as ``reactor.'' In general, in the field of atomic science, it is appropriate to translate it as [reactor j], and in the field of chemistry, it is often more appropriate to translate it as "reactor J".In such cases, the word + rraactorJ is used. The translation field vector for the translation word "nuclear reactor" is determined such that the component in the field of atomic science has a large value.On the other hand, the translation field vector for the translation word "reactor j" is determined such that the component in the field of chemistry has a large value. It is made to have a large value.

Similarly, for the word field vector, the degree of compatibility with each field is determined as the value of each component. For example,
``That J,'' ``this J, etc. are words that are commonly used in all non-distinct fields.

Every vector component is assigned a small value.

Assuming that the fields are classified into n fields, for example, as shown in FIG. 4 (in an n-dimensional vector space.

A field vector V will be given. That is, 1
Each field vector V has a direction and strength, and can be considered to indicate the individuality of a word or a translated word in relation to the field in which it is used.

Next, referring to FIG. 5, a configuration of an embodiment of the present invention will be described in which translated words are determined based on these field vectors.

In FIG. 5, numerals 2 to 10 correspond to tree 1,
Reference numeral 11 denotes a field vector synthesis unit that generates a target field vector by synthesizing word field vectors associated with words that appear in the original text to be translated, and 12 an inner product that calculates the internal value of the target field vector and each translated word field vector. Arithmetic unit, 1
3 is a comparison unit that compares the inner product calculation results; 14 is a translation determination unit that determines and extracts a translated word based on the comparison result; 20 is an input buffer; 21 is word data of the original text; 22 is a buffer; 23 is a vector adder. , 24 is a target field vector register, 25 is a group of switching processing instructions, 26 is a group of internal calculation instructions, 27 is a group of maximum value detection instructions, 28 is a group of translated word extraction instructions, 2
9 represents the cover sofa.

The input unit 2 inputs the original text to be translated into the input buffer 20 . The word dividing unit 3 divides the input original text into 9 spaces, for example.
Divide into words according to data.

The dictionary search unit 4 searches the word dictionary 1 for each word divided into nine parts.
0 is searched, and word information as shown in FIG. Based on this word information,
The syntactic analysis unit 5 analyzes the syntax, and the one-word order determining unit 6 determines the word order of the translated word when the translated word is selected. The above processing may be the same as the conventional one.

The field vector synthesis unit 11 sequentially adds the word field vectors of each word prepared in the buffer 22 using a vector adder 23, and stores the result in the target field vector register 24. In addition.

An appropriate coefficient is multiplied to prevent overflow during addition. For example, if the target field vector is 3. Let the number of words in the original text be thunder, and let the O(th word field vector be V(O
Then, the subject field vector S stored in the subject field vector register 24 is as follows.

This target field vector S may be considered to indicate the characteristics of the field in which the original text to be translated is currently used.

The translation word selection section 7 activates the inner product calculation section 12 . When there are a plurality of translations for one word, the internal bond calculation unit 12 uses the switching processing instruction group 25 to read the translation field vector of each translation from the buffer 22 and provides it to the inner product calculation instruction group 26. By the internal debt operation instruction group 26,
Target field vector 3 of target field vector register 24
An inner product operation is performed between and each target word field vector. For example.

The Tj′-th translation field vector of the i-th word is
3(i), the inner product of vectors S−V j(S)
However, j=1.2. ... after each k (k is the number of translated words).

Calculated. Needless to say, the size of this internal value indicates the degree of coincidence between each translated word field vector V 3 (*) and the target field vector S.

The comparison unit 13 uses the maximum value detection command group 27.

The above inner products 3-V3(s) are compared and the one having the maximum value is detected. The translation determination unit 14 uses a translation extraction command group 28
Therefore, the translation word Tjo(1) corresponding to the translation word field vector V10(S) that has the maximum internal bond with the target field vector
is set as the translation of the word, and the code of the translation is set in the output buffer 29. The translation selection section 7 is.

Repeat the process for all words in the original text,
If all translated words could be extracted.

Call the output section 8. The output unit 8 prints out the translation result on the output buffer 29 to, for example, a line printer.

When the field vector synthesis unit 11 synthesizes word field vectors, it may be synthesized for each sentence of the original sentence. Also 1 For example, target field vector register 2
Instead of clearing the contents of 4 for each sentence, the range of 9 synthesis may be a series of sentences or a predetermined number of word groups.

Note that the word field vectors registered in the word dictionary 10 and the translation field vectors of each translated word are determined and set to appropriate values based on experience, for example, at the time of creating the word dictionary IO, but they are roughly updated to appropriate values through learning. It is also possible to do so.

(El) As described in detail, according to the present invention, a translation of a polysemous word can be selected without performing sophisticated semantic processing and by using relatively little computer processing time and memory.

[Brief explanation of drawings]

Figure 1 is an example of the configuration of a general machine translation device; Figure 2 is an example of the configuration of a word dictionary according to the present invention; Figures 3 and 4 are explanatory diagrams of field vectors; Figure 5 is an example of the invention. An embodiment of the configuration is shown below. In the figure, 7 is a translation selection section, 10 is a word dictionary, 11 is a field vector synthesis section, 12 is a domestic debt calculation section, 13 is a comparison section, 14
represents the translation word determination section. Patent applicant Fujitsu Limited

Claims (1)

[Claims] An original text is input, and the input original text is divided into words. In a machine translation device that extracts translated words corresponding to original words registered in a word dictionary to compose and output a translated sentence, the word dictionary has a dimension of the number of fields determined in advance for each registered word, and In addition to holding word field vector information whose components are the degree of compatibility between a word and each of the above fields, one for each translation of the registered word, a vector corresponding to the above word field vector that contains the corresponding translation and each of the above fields. It holds translation field vector information whose component is the degree of compatibility with the target word. a synthesis unit that generates a target domain vector by synthesizing the plurality of word domain vectors obtained from a group of words to be translated;
a translation word selection unit that compares the target field vector synthesized by the synthesis unit and the plurality of translation word field vectors for one word, and determines the translation having the most similar translation word field vector as the translation of the word; A machine translation device characterized by: