JPH03157764A - Machine translation device - Google Patents

Abstract

PURPOSE:To facilitate the collection, correction and management of information for selecting the language by executing the selection of a translated word by utilizing the information in a second language of a co-occurrence relation between the translated word stored in a co-occurrence relation dictionary instead of the information extending over between a first and a second languages. CONSTITUTION:The device is provided with an input part 1 such as a keyboard, etc., an analyzing part 2, a word dictionary 3, a translated sentence generating part 4, and a display part 5 as bases, and also, consists of a co-occurrence relation dictionary 6 connected to the translated sentence generating part 4. In such a state, the selection of a translated word is executed by utilizing the information in a second language of a co-occurrence relation between the translated words stored in the co-occurrence relation dictionary 6 instead of the information extending over between a first and a second languages. In such a way, the collection, correction and management of information for selecting the translated word can be facilitated.

Description

[Detailed description of the invention] Industrial applications The present invention relates to a machine translation device.

Conventional technology Generally speaking, when translating an original text in a machine translation device,
It is necessary to refer to the information in the word dictionary and assign a translation to each expression that makes up the original sentence. At this time, in the word dictionary,
There are cases where multiple second language translations are written for one expression in the first language. Therefore, it is necessary to select one appropriate aI< word from among these plural translation words.

22. Conventionally, when there are multiple translations in the second language for a certain expression (called a headword) in the first language in a word dictionary, it is necessary to I try to choose the appropriate one from among these multiple translations.

In other words, translated words belonging to the second language are selected based on their correspondence with headwords belonging to the first language (see, for example, Japanese Patent Laid-Open No. 60-247787).

Problems to be Solved by the Invention However, such a translation selection method has the following disadvantages.

Generally, translated words are sometimes determined not by their correspondence with headwords in the first language, but by the relationship between the translated words in the translated text in the second language (this is called a co-occurrence relationship). For example, in the case of English-Japanese translation, consider the case where an expression "close eyes" consisting of two headwords in the first language is to be translated. For each headword, “close” and “eyes,” there are multiple translations as follows:
Suppose that it is compatible.

-close: close, shut 'eyes': eyes In this case, ``close'' is more appropriate than ``shut'' as a translation for ``clos 6''. but,
This appropriateness is due to the fact that the lemma in the first language
It does not arise from the syntactic and semantic properties of ``o s e'' and ``eyes,'' but rather the translation in the second language.
and ``Close/Close J'' can be said to ``close your eyes,'' but ``
This arises from the fact that we do not normally say "close your eyes" (this is called a co-occurrence relationship between "eyes" and "close").

In the case of the conventional method of determining the translation word in the second language according to the properties of the entry word in the first language, the following information is required to select the translation word in the above case. .

・If the object is “eyeS”, the translation of “close” is “close”.・If the object is “window”, then the translation is “close”.
The translation of ``shut'' is, in this way, originally a property in the second language (in this case,
co-occurrence relationship) and the property of the first language (in this case, the object is
In order to treat the language as "eyes or not", information that spans the first and second languages is required, and it is difficult to collect, modify, and manage information compared to information that is confined within a single language.

Means for solving the problem The input original text in the first language is analyzed by the analysis unit to extract the structure of the original text, and the structure of the original text and the first language and second language stored in the LJ dictionary are extracted. In a machine translation device in which a translation generation unit generates a translation in a second language using dictionary information regarding words, co-occurrence r4A between translated words
A co-occurrence relationship dictionary is connected to the translated text generating unit and is used for selecting translated words by the translated text generating unit.

Since the selection of action translations is performed using information in the second language, such as co-occurrence relations between translated words stored in the co-occurrence relationship dictionary, rather than information that spans between the first and second languages, the selection of translation words is This makes it easier to collect, modify, and manage information for purposes.

Embodiment An embodiment of the invention recited in claim 1 will be described based on the drawings. FIG. 1 is a block diagram showing a schematic configuration of a machine translation device, which is based on an input section l such as a keyboard, an analysis section 2, a word dictionary 3, a translation generation section 4, a display section 5,
Furthermore, it consists of a co-occurrence relationship dictionary 6 connected to the translation generation unit 4.

Here, as shown in FIG. 2, the input unit 1 inputs the original text in the first language, for example, the original text 2 I C1□s6 my eyes.

Enter. The analysis unit 2 analyzes the original text using the word dictionary 3 and extracts the structure of the original text. In the case of the upper original text, the structure of the parsed original text is as follows. The structure of the original text in this embodiment is such that the dependence relationships between the headwords included in the original text are expressed as tree data. That is, II II II and II eyes II
is dependent on "close", and II myll is dependent on "eyes". The nodes that make up the tree are called nodes. Next, the translation generation unit 4 uses the word dictionary 3 to transform candidate translation words into appropriate word forms according to the type of dependency relationship in the original text for the headwords in the original text that make up the structure of the original text. Give. As a result, the following translation structure is obtained.

One translated word is selected for each node in the structure of the translated text obtained in this way.

For this selection, the following processing is performed in this embodiment.

a. Restore nodes to which multiple translation words are attached in the structure of the translated sentence. Such nodes are called fatty nodes. In this example, only the node "close" is given multiple translations "close" and "close".

b. Examine nodes that have a dependency relationship with the restored multi-fatty node, and find all combinations of translated words assigned to it and translated words of the original node. In this example, four combinations are obtained: (I close) (I close) (I close my eyes) (I close my eyes). These combinations are called co-occurrence candidates.

C9 On the other hand, the co-occurrence relationship dictionary 6 stores correct co-occurrence relationships between words in the second language, as illustrated in FIG. 3(a). In the figure, each row indicates one co-occurrence relationship.

d. Check whether the co-occurrence candidate according to the above is included in the co-occurrence relationship dictionary 6. Such co-occurrence candidates are called active co-occurrence candidates. In this example, (eyes, close) is an active co-occurrence candidate.

e. Count the number of times that multiple translations of a fat node are included in active co-occurrence candidates, and select the translation with the largest count. If there are equal counts, or if all counts are O, set an appropriate default process and narrow it down to one.

In the case of this example, the count of translated words for the fat node ``close'' is 1. Since "Close J" is 0, "Close" is selected.

By the way, in the case of this example, we considered the case where there is only one fat-rich node in the original text, but even if there are multiple fat-rich nodes, the above processes a through to are performed in the same way, and instead of process e, ,
The translation words may be determined so as to create a combination of translation words that maximizes the total count of all fatty nodes in the sentence. In short, when selecting a translation word, the information in the co-occurrence relationship dictionary 6 may be referred to, and the method of reference is not particularly limited.

Further, in this example, a relationship between two words is assumed as the co-occurrence relationship, but a co-occurrence relationship between three or more stories can be handled in the same way.

The contents of the co-occurrence relationship dictionary 6 are shown in FIG.
), words in a co-occurrence relationship may be separated into co-occurrence relationship symbols (in this case, objective case, nominative case, etc.) that generalize and express the type of co-occurrence relationship. and,
In process d, when searching the co-occurrence relationship dictionary 6, it is first checked to see if there is a match between the word pairs, and if there is a match, the co-occurrence relationship symbol matches the co-occurrence relationship of the co-occurrence candidate. whether or not (for example, “wo” is consistent with the objective case)
It is only necessary to determine whether or not it is an active co-occurrence candidate. According to this, co-occurrence relationships can be expressed more efficiently. In short, the method of expressing information in the co-occurrence relationship dictionary 6 is not particularly limited either.

Based on the translated word structure obtained as described above, in which one translated word is assigned to each node, the translated text generation unit 4 generates a translated text, for example, a translated text; I close my eyes.

is generated and output to the display section 5 as necessary.

Next, an embodiment of the invention according to claim 2 will be described.

In the machine translation apparatus of the embodiment described above, the amount of information on co-occurrence relationships required is considerable, so it is desirable to be able to automatically collect the information.

In this regard, attempts have been made to automatically extract such co-occurrence relationships from sentences. For example, according to ``Attempt to create a dictionary of co-occurrence relationships from text'' in the Information Processing Society of Japan's 38th National Conference Paper Collection, Uke uses analysis to automatically extract co-occurrence relationships between words from sentences. It has been shown that However, due to the complexity of natural language, it is difficult to completely exclude incorrect extraction of co-occurrence relationships.

Furthermore, it is difficult to collect and prepare all the necessary information for selecting translations, considering the vastness of linguistic phenomena. It is hoped that this will be done.

In this regard, conventionally, when a user corrects an inappropriate translated word translated by a translation device, the translated word is marked and given priority over other translated words from the next time, so that the translated word is learned. The one that made it possible to
This method has been proposed in Publication No. 3 and Japanese Unexamined Patent Publication No. 61-260366. However, in this method, relationships with other translated words in the translated text are not taken into account. For example, in English-Japanese translation, I
ip+aythe piano” II play I
If the user specifies "play" as a translation of I, If
The translation of "play II" is marked and given priority. Therefore, the expression "play tennis'" results in "playing tennis."

On the other hand, Japanese Unexamined Patent Publication No. 150068/1983 has proposed a system that avoids this inconvenience by storing combinations of translated words in correspondence with combinations of words in the original text. That is, in the upper case, (play, pi
Let's make the pair of translations ``ano'' correspond to the pair of translated words ``play, piano'', and next time we'll use ``(play, piano)''.
) appears in the original text, priority is given to selecting (play, piano) as the translated word. However, as with the conventional technology, this results in learning information that spans the first and second languages, and modifying and managing that learning information is more difficult than information within one language. Become.

In view of these points, this embodiment automatically extracts co-occurrence relations by extracting co-occurrence relations between translated words from the structure of the translated text that has already been obtained for the translated text that has been determined to be correct by the user. Therefore, only correct co-occurrence relationships are automatically collected.

Regarding learning, co-occurrence relationships between translated words selected and modified by the user, which is information in the second language, are extracted from the structure of the translated sentence obtained during translation and stored in the co-occurrence relationship dictionary 6. The system takes into account the relationships between translated words, learns translated words whose information can be easily corrected and managed, and efficiently collects co-occurrence relationships.

Specifically, in the machine translation device of the embodiment, the structure of the translated text is obtained every time it is translated, so the co-occurrence relationship is extracted from the structure of the translated text as follows, It is registered in the co-occurrence relationship dictionary 6.

a. Extract all dependencies from the structure of the translated text that corresponds to the translated text that the user has judged to be correct. Using the specific example of the embodiment described in Section 1, three dependencies are obtained: (I close) (I close my eyes) (My eyes).

b. Search the co-occurrence relationship dictionary 6 for the above dependency relationship, and if it is not registered, each dependency relationship is treated as a co-occurrence relationship and this co-occurrence relationship dictionary i? 6 cj+ Register 4.

Here, even if the translation is incorrect, if the user makes corrections to it to make it a correct translation, the correction information is reflected in the structure of the translation, so that this embodiment can also be used for such translations. Processes a and b can be performed.

Effects of the Invention Since the present invention is configured as described above, according to the invention described in claim 1, the selection of translation words is not information that spans between the first and second languages, but is stored in the co-occurrence relationship dictionary. Since this is done using information in the second language, such as the co-occurrence relationship between translated words, it is possible to easily collect, modify, and manage information for selecting translated words. According to the invention of
Information for selecting translation words can be automatically and efficiently collected in the co-occurrence relationship dictionary.

[Brief explanation of the drawing]

The drawings show one embodiment of the present invention; FIG. 1 is a block diagram, FIG. 2 is a flowchart showing processing, and FIG. 3 is an explanatory diagram showing registered contents of a co-occurrence relationship dictionary. 2...Analysis section, 3...Word dictionary, 4...Translation generation section, 6...Co-occurrence relation dictionary Applicant: Rico Asaichi Atsushi, Ltd. Figure 7 - Ushi 395-

Claims (1)

[Claims] 1. The input original text in the first language is analyzed by the analysis unit to extract the structure of the original text, and the structure of the original text and the first language and second language stored in the word dictionary are extracted. In a machine translation device, a translation generation unit generates a translation in a second language using dictionary information regarding words in the translation unit, which stores a co-occurrence relationship between translation words and uses a co-occurrence relationship between the translation words to be used for selection of translation words by the translation generation unit. A machine translation device characterized in that an origin relation dictionary is connected to the translation generation unit. 2. The machine translation device according to claim 1, wherein the co-occurrence relationship dictionary stores co-occurrence relationships between translated words extracted from the structure of a translated sentence obtained during translation.