JPH0658672B2 - Machine translation system - Google Patents

Description

[Detailed Description of the Invention] [Outline] A machine translation system that performs machine translation using parallel / contrast expressions of an original sentence. A parallel / contrast expression appearing in an original sentence is extracted and used for sentence analysis and translation generation. In order to quickly generate a high-quality translated sentence, the original sentence is decomposed into words, and the word decomposition / phrase generation unit that generates phrases and the phrase generated by this word decomposition / phrase generation unit are parallel to each other. Parallel / contrast expression detecting section for detecting parallel / contrast expressions appearing in the above-described manner, and correspondence / attribute correction for compensating the words and attributes of the words forming the parallel / contrast expression detected by the parallel / contrast expression detecting section. Section, and a translation generation section that generates a translation in a manner that uses the parallel / contrast expression in which the correspondence and attributes of the words are supplemented by the correspondence / attribute correction section, and the translation is generated by the translation generation section. , It is configured to output the generated translated text.

[Industrial application field]

The present invention relates to a machine translation system that performs machine translation using parallel / contrast expressions of original sentences.

[Problems to be solved by conventional technology and invention]

Generally, a machine translation system first decomposes an original sentence into clauses / words and analyzes what is stated in the original sentence using grammar and knowledge. The analysis result is represented by an intermediate representation of each system. Next, the translated sentence according to the target language is generated according to the analysis result of the original sentence.

However, conventionally, in the analysis stage of the original sentence, at each time point, only one part of the original sentence is referred to for the analysis, so that the entire structure of the sentence may not be reflected. In addition, even if there is a so-called parallel / contrast expression in which parts of almost the same structure repeatedly appear, the conventional system repeatedly performs the same processing on these repeated parts,
There was a problem that the translation speed could not be improved. Also,
Since the translation is performed on a sentence-by-sentence basis, it is difficult to determine the content indicated by pronouns used in other related sentences, and it is difficult to create a high-quality translated sentence.

An object of the present invention is to quickly generate a high-quality translated sentence by extracting parallel / contrast expressions appearing in an original sentence and using it for sentence analysis and translated sentence generation.

[Means for solving problems]

Means for solving the problem will be described with reference to FIG.

In FIG. 1, the word decomposition / phrase generation unit 1 decomposes the input original sentence into words and collects the words to generate a phrase.

The parallel / contrast expression detecting unit 2 detects a so-called parallel / contrast expression, which is a block of a phrase repeatedly appearing in the phrase generated by the word decomposition / phrase generating unit 1.

The association / attribute correction unit 3 supplements (or limits) the association of the corresponding word in the parallel / contrast expression detected by the parallel / contrast expression detection unit 2 and the meaning of this word.

The translated sentence generation unit 4 uses the parallel / contrast expression associated and corrected by the association / attribute correction unit 3 to generate a translated sentence in a manner of replacing the sentence appearing after the second time. .

[Action]

The present invention, as shown in FIG.
Is decomposed into original words to generate phrases,
The contrast expression detection unit 2 finds a parallel / contrast expression in which the block of the phrase repeatedly appears, and the correspondence / attribute correction unit 3
Corrects the attributes (translated words, etc.) of the words associated by this parallel / contrast expression. After this correction, the sentence associated by this parallel / contrast expression is
The translated sentence is generated in a manner of replacement while paying attention to the comparison.

Therefore, it is possible to quickly generate a high-quality translated sentence by replacing the translated sentences of the second and subsequent translated sentences corresponding to the parallel / contrast expression in the original sentence and omitting the syntactic analysis. It is possible to easily estimate the instruction content.

〔Example〕

Next, according to the procedure shown in the flowchart of FIG.
The configuration and operation of FIG. 1 will be sequentially described in detail with reference to FIGS.

In FIG. 2, the figure shows a state in which the original text is read. This means, for example, reading the English original sentence shown in FIG.

The figure shows a state in which words are put together to generate a phrase. This is done by morphologically analyzing the original sentence of FIG. 3 (a) into words and applying the following rule (1) to the divided words, for example, FIG. 3 (b) (step 1). And (c) means to be grouped into phrases as shown in (step 2).

This will be described in detail below.

Rule (1): (1) Omit adverbs and interjections.

(2) Articles and nouns are nouns.

(3) If there is a noun immediately after that, consider the adjective as a limited usage, and connect it with the subsequent noun to make it a noun (noun). Otherwise, if there is something similar to the be verb before (become, come, go, etc.), consider it as a narrative usage, and be + adj together with the previous verb.
And

(4) Verbs are divided as follows according to their usage.

(4-1) Participle (past participle, present participle) (4-2) to-infinitive, gerund (4-3) prototype indefinite vision (4-4) auxiliary verb + prototype (including future tense), present tense , Past tense, past perfect tense, present perfect tense (5) Article and possessive case are noun together with later phrases. However, if the latter word is an adjective, mark it as comparative or superlative (noun + est) according to its form.

(6) If the preposition + noun is a noun before, the noun phrase is combined with the preceding noun. Otherwise, if the previous word is similar to the be verb, treat it as an adjective narrative.
If not, treat it as an adverb.

Step 1 shown in Fig. 3 (b) This is the above rule (1)
Is applied to the word separated from the original sentence of FIG. 3 (a) in the first step. Briefly, Fig. 3 (b)
The phrase enclosing the original text in the figure in a rectangle is obtained by applying the above rule (1). For example, the phrase "to se at the beginning of the original sentence
For e ”, apply item (4) of rule (1) and add to-inf
It is generated as (infinitive). Similarly, it is generated as shown in the figure. In addition, () in the figure represents that the word enclosed by this is ignored and applied, and * represents that the rule is not yet applied.

Step 2 shown in FIG. 3 (c): This is for the unapplied words (words indicated by * in the figure) in FIG. 3 (b).
Further, it is generated by applying the rule (1). To briefly explain, a phrase in which the original sentence in FIG. 3 (c) is enclosed in a double rectangle is obtained by further applying the above rule (1). For example, since the original sentence "the fanlt of othefs" is connected (owned) by of, the whole is no as shown in the figure.
set to un. Similarly, the whole of “then of your own” is also designated as noun. The tree structure (list structure) of FIG. 3 (c) generated in this way is shown in FIG. 3 (d). Like

By expressing as shown in FIG. 3 (d) by the above means, it becomes easy to easily extract a so-called parallel / contrast expression, which is a block of a phrase repeatedly appearing in the original sentence of FIG. 3 (a).

The figure shows a state in which repeated contrast expression is detected.
This is a parallel / contrast expression that is a block of a phrase that appears repeatedly from Fig. 3 (d), as shown in (a-1) and (a-2) in the figure.
This means to detect "to-inf → noun → be + adj". In this detection, pay attention to the heading part (top row) of the list in the tree structure in FIG. It is regarded as a one-dimensional list, and two pointers p and q are prepared so that they respectively point to a list of a tree structure. Whether the partial list starting from the pointer p and the partial list starting from the pointer q match or not. At this time, it checks sequentially from the longer partial list and finds the longest matching one.
Detect as parallel / contrast expression. The parallel / contrast expression detected in this manner is to-inf → noun → be + adj ".

The figure shows a state in which the words in the original sentence are associated with each other.
This is the parallel / contrast expression detected in the figure, for example, the parallel / contrast expression in the tree structure in FIG. 3 (d) to-inf → noun → be + ad.
This means that j "is associated as shown in Fig. 4. Specifically, the parallel / contrast expression to-inf →
It is found that to see "and to find" in the original text correspond to the first to-inf "of" noun → be + adj ". Similarly, the second noun "and the third b in Fig. 3
"e + adj" is also associated as shown in FIG.
The underlined characters in FIG. 4 indicate that some words are further represented. For those underlined, the collation is repeated for the child list. In addition, the words enclosed by () should be ignored.

In the figure, indicates the state of supplementing the attribute of the word or limiting the word. This is because the words associated in the figure have different grammatical functions, meanings, and translated words even if they have the same notation, and there are multiple words for the same notation in the word dictionary. In order to generate a high quality translated sentence, the appropriate translation word is selected by utilizing the characteristics of parallel / contrast expression. Specifically, it is shown at the bottom of FIG. In parallel / contrast expression, the words "easy" and "hard" are translated into "easy", "easy", "easy", "don't care ...", etc.
From these translated words, a parallel / contrast expression be-adj "
Easy and hard to translate with the most similar usage
Select. In this way, it is possible to generate a high-quality translated sentence by making use of the characteristics of the parallel / contrast expression to supplement or further limit the attributes (meaning, etc.) of words. As a criterion for selecting the translated word, in addition to selecting the similar ones described above, a collection of opposite words may be provided in advance and selected. Also, othe in the parallel / contrast expression in the middle part of FIG.
For "rs", information such as "possessiveness of a word representing a person" is set.

Next, a translated sentence is generated as shown in, for example, FIG.

The figure shows a state in which the first part is translated. this is,
For example, (a-1) in the parallel / contrast expression extracted from the result of applying rule (1) shown in FIG. 3 (d) is translated, and expressed in a tree structure as shown in the upper part of FIG. I mean.

The figure shows a state in which the comparison expression is translated in a manner that is carefully applied (replaced). As shown by enclosing the ellipse in the lower part of Fig. 5, the parallel / contrast expression appearing after the second time is not parsed and the pair relationship of Fig. 4 parallel / contrast expression is used. It means that the translation is performed in a manner that is simply replaced.

The figure shows the state of output as a translated sentence. This is because, as shown by the dotted line in Fig. 5, the translated sentence represented by the tree structure representation in the upper row from the position "start" in the diagram is translated as "it is easy to see the flaws of others," and then It means that it is quite difficult to find the translated sentence represented in the lower tree structure representation after you translate the conjunction but "existing in the original sentence" into "." .

As described above, it is possible to extract parallel / contrast expressions that appear repeatedly from the original sentence, and to translate them in a simple fitting manner without performing syntax analysis or the like in the second and subsequent translations.

〔The invention's effect〕

As described above, according to the present invention, the parallel / contrast expression that repeatedly appears in the original sentence is extracted, and based on this parallel / contrast expression, it is translated in a mode that is replaced in the second and subsequent translations. Because it is adopted, it appears in the original text 2
For the translated sentences of the parallel / contrast expressions after the first time, it is possible to quickly generate a high-quality translated sentence by omitting the syntactic analysis and replacing the translated sentences. Furthermore, by comparing words in parallel / contrast expressions to each other, it is possible to reflect the meaning content in one sentence in another sentence, for example, to easily reflect the instruction content (person, object, etc.) such as a pronoun. It is possible to generate high quality translated text.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an embodiment of the present invention, FIG. 2 is an operation explanatory flowchart of the present invention, FIG. 3 is an explanatory diagram of a specific example relating to the present invention, and FIG. 4 is a comparative explanatory diagram of parallel / contrast expressions, FIG. 5 shows an example of a tree structure representation of a translated sentence. In the figure, 1 is a word decomposition / phrase generation unit, 2 is a parallel / contrast expression detection unit, 3 is an association / attribute correction unit, and 4 is a translation generation unit.

Claims (1)

[Claims] 1. In a machine translation system for machine translation using parallel / contrast expressions of original sentences, a word decomposition / phrase generation unit that decomposes an original sentence into words and generates phrases.
(1), and a parallel / contrast expression detection unit (2) that detects parallel / contrast expressions that appear in a parallel manner for the phrases generated by the word decomposition / phrase generation unit (1), and this parallel / contrast The correspondence / attribute correction unit (3) that complements the correspondence and the attributes of the words that form the parallel / contrast expression detected by the expression detection unit (2), and the correspondence of the words by the correspondence / attribute correction unit (3). And a translation generation unit (4) for generating a translation in a manner of replacement using a parallel / contrast expression supplemented with attributes and attributes, and the translation generation unit (4) outputs the generated translation. A machine translation system characterized by being configured.