JPH0756919A - Japanese analysis method - Google Patents

Abstract

PURPOSE:To provide a Japanese analysis method by which the modification relation of clauses as against a long Japanese sentence where the number of the clauses is large is efficiently approved and the equivocal generation of modification can be made minimum. CONSTITUTION:A morpheme analysis part 2 refers to a morpheme dictionary 5 and word-divides an input sentence so as to decide the clauses. A modification analysis part 3 refers to a modification dictionary 6 and extracts the collection of more than one clauses constituting a predicate clause. The predicate clause is classified stepwise in accordance with the strength of independency so that it is adjusted to the expression structure of Japanese, and the structure of the input sentence is analyzed based on the strength of the independency of the predicate clause. The modification relation of the respective clauses is approved on the input sentence which is structure-analyzed, and the structure of the Japanese input sentence is analyzed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a Japanese parsing method for parsing an input sentence by recognizing a dependency relation between clauses.

[0002]

2. Description of the Related Art Conventionally, as a syntactic analysis in Japanese sentences, the action of each bunsetsu is classified, and according to this classification, a general rule of dependency based on a so-called phrase structure grammar, that is, (1) Placed before the clause
(2) The related bunsetsu can have only one receiving bunsetsu,
A well-known method is to determine the dependency pattern of the entire sentence by (3) recognizing the dependency between each clause by trial and error by using that the dependency relationship does not intersect.

Further, since the analysis accuracy for a long sentence (10 phrases or more) is extremely low only by the general law, recently, by paying attention to the similarity of the phrase sequence, the parallel structure of the phrases or the clauses is detected. Method for improving accuracy (Kurohashi,
Nagao, "Estimation of Parallel Structures in Long Japanese Sentences (Information Processing Society of Japan Material NL-86-2, November 15, 1991)" and "Parsing of Long Japanese Sentences Based on Detection of Parallel Structures (Information Processing Academic Conference Material NL-88-
1, March 12, 1992) ”) was proposed, and achieved an analysis accuracy of 66% for an average of 14.2 bunsetsu sentences under the severe condition of not making ambiguity in analysis.

[0004]

The above general rule of dependency makes it very difficult to analyze a long sentence, and the above-mentioned method of Kurohashi et al. Is effective when the similarity of bunsetsu sequences cannot be found. There is a basic problem that it cannot exert its effect.

For example, a sentence of 14 clauses "publishing agency / original / profit margin / low /// in addition / publishing /
Demand is slowing / because / suffering / management / is forced / is being done. "(Nikkei Sangyo Shimbun, / means phrase boundary)" contains five predicates ("low", "additionally", "blunted", "painful", "is done"), but There is no parallelism clues to parallel clues, and "low / additional / additional" and "blunting / because" form a predicate phrase, so the bunsetsu is similar to the conventional method. In some cases, the correct answer may not be obtained in order to recognize the dependency relationship in units.

[0006] In addition, a sentence of 20 clauses "(XX company /) / in the water supply business / introducing a billing machine / from the era / to the local government / delivery record / is, but // after / to become an office computer // It is / is being pushed by competitors / is being / pushed / software / systematic / reviewing / project team / strengthening / / in the local government market / recovery from lost land / aiming. (Nikkei Sangyo Shimbun) , 6 predicate phrases are included (“Aru ga,” “Naru”, “Nata,” “Revisit,” “Strengthen,” “Aim.”), But they are also used as clues for parallel clauses. There is no similarity, and a polymorphic explosion is expected in the dependency relationship between predicates.

The present invention has been made in view of the above,
It is an object of the present invention to provide a Japanese analysis method capable of efficiently recognizing a dependency relation between bunsetsu for a Japanese sentence having a large number of bunsetsu and minimizing the occurrence of polysemy of the bunsetsu.

[0008]

In order to achieve the above object, the Japanese analysis method of the present invention recognizes a word included in an input sentence expressed in Japanese, and recognizes it as a group consisting of one or more words. A Japanese parsing method that parses an input sentence by recognizing a dependency relation between bunsetsu using a certain bunsetsu as a basic unit, and extracts a group of one or more bunsetsus that form a predicate phrase, The predecessor words are classified in stages according to the strength of independence so as to match the expression structure of
Analyzing the structure of the input sentence based on the degree of independence of the predescription phrase, further recognizing the dependency relation between each clause in the input sentence subjected to the structure analysis, and parsing the Japanese input sentence. Is the gist.

[0009]

According to the Japanese analysis method of the present invention, a group of one or more clauses forming a predicate phrase is extracted from an input sentence, and the predescription phrase is made to have a high degree of independence so as to match the expression structure of Japanese. According to the degree of independence of the predescription phrase, the structure of the input sentence is analyzed according to the degree of independence, and the dependency relation between the clauses of the input sentence subjected to the structure analysis is further recognized. Parse a word input sentence.

[0010]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a Japanese analysis system for implementing a Japanese analysis method according to an embodiment of the present invention. In the figure, 1 is an input unit for inputting a Japanese sentence, 2 is a morpheme analysis unit that divides an input sentence into words using a morpheme dictionary 5, and 3 is a relation from a morpheme analysis result obtained by the morpheme analysis unit. Dependency analysis units that analyze the dependency relations between phrases using the dependency dictionary 6 and 4 denote output units that output the dependency analysis results.

The outline of the operation of the Japanese analysis system shown in this embodiment is as follows.

The Japanese input sentence input from the input unit 1 is divided into words by the morphological analysis unit 2. Next, in the dependency analysis unit 3, based on the output from the morpheme analysis unit 2,
Dependency analysis between phrases is performed using the dependency dictionary 6.
This result is output from the output unit 4.

FIG. 2 is a block diagram showing a detailed configuration of the dependency analysis unit 3 and the dependency dictionary 6 used in the system shown in FIG. As shown in FIG. 2, the dependency analysis unit 3 is composed of a predicate phrase recognizing unit 31, a predicate phrase classifying unit 32, a predicate phrase relationship recognizing unit 33, and a dependency relationship recognizing unit 34, and the dependency dictionary 6 is a predicate phrase classifying unit. It is composed of a table 61 and a dependency approval table 62.

Predicate phrase classification table 61 of the dependency dictionary 6
Are configured as shown in FIG. In the predicate phrase classification table 61, rule numbers 11, 12, ... And predicate phrase classification S ₀ , S for word configurations of various clauses.
₁ , ... are shown. In the predicate phrase classification table 61 shown in FIG. 3, W ₁ shown in the word structure of the phrase
Represents an arbitrary word (however, the condition of part of speech is specified),
* Represents any zero or more words. Originally, only the part that constitutes the predicate phrase should be described, but in the form of "sahen noun + reading point", it is determined that it is a predicate phrase in order to distinguish it from parallel, such as "research and development". It specifies whether or not the preceding clause is a case element as a condition for doing.

The dependency approval table 62 of the dependency dictionary 6 is configured as shown in FIG. In the dependency approval table 62, a clause as a dependency source and a clause as a dependency destination are shown as a pair for each rule number, and attributes of the dependency are shown.
In addition, {} shall use any one word described in or classification. W ₁ represents an arbitrary word (however, the condition of the part of speech is specified), and * represents 0 or more arbitrary words (may be accompanied by the condition specification of the part of speech). Furthermore, it is assumed that the predicate phrase is a predicate phrase that satisfies the condition of the most recent "related clause".

Next, the operation of the embodiment shown in FIG. 1 will be described with reference to the flow chart shown in FIG. 5 and the analysis results shown in FIGS. 6 and 7.

First, a case will be described in which the input sentence is "Inherently, the profit ratio is low in the publication agency, and the demand for the publication is slowing, so that the management is difficult". This input sentence is input to the morpheme analysis unit 2 via the input unit 1 shown in FIG. 1, and the morpheme analysis unit 2 divides the words while referring to the morpheme dictionary 5, and as shown in FIG. A morphological analysis result is obtained (step 110 in FIG. 5).

The result of this morphological analysis is the dependency analysis unit 3
And the predicate phrase recognizing unit 31 of the dependency analyzing unit 3 refers to the predicate phrase classification table 61 to detect a group of predicate phrases (step 120). In this example, the morpheme analysis result of the clauses 4 to 7 matches the word structure of the clause of the rule 71 of the predicate phrase classification table 61, and the morpheme analysis result of the clauses 9 to 10 is the rule 51 of the predicate phrase classification table 61.
The morpheme analysis result of the bunsetsu 14 matches the word structure of the bunsetsu of the rule 11 of the predicate phrase classification table 61. Therefore, the result is supplied to the predicate phrase classifying unit 32. The predicate phrase classification unit 32 refers to the predicate phrase classification table 61 for this result, determines the classification of the predicate phrase, and assigns the predicate phrase classification S ₅ , S ₄ , and S ₀ as shown in FIG. (Step 130). Note that the predicate phrase recognizing unit 31 detects a group of predicate phrases with reference to the predicate phrase classification table 61 by detecting each predicate phrase of the morphological analysis result as shown in FIG. This is performed by sequentially comparing all the configurations and determining whether the two match.

As described above, the result of classifying the predicate phrases is supplied to the predicate phrase relation recognizing unit 33, and in this predicate phrase relation recognizing unit 33, the relation recognizing table 62 is referred to to determine the relation between the predicate phrases. Receiving is decided (step 14)
0). As shown in the dependency of FIG. 6, the predicate phrase consisting of the clauses 4 to 6 having the predicate phrase classification S ₅ is the dependency authorization table 62 of FIG. The rule 103 determines that the predicate phrase consisting of the clauses 9 and 10 having the predicate phrase classification S ₄ becomes the clause 14 according to the rule 103.
Note that this dependency is determined according to the strength of independence of each predicate phrase. The strength of this independence depends on the relation between each predicate phrase, parallelism, simultaneity, causality, etc. Determined,
For example, it can be said that the independence is high when it means that each predicate phrase is not related at all, and the independence is low when there is a causal relationship such as a cause reason. Further, the determination of this dependency is made by classifying each predicate determined as described above,
For example, for S _5, S ₅ corresponding among the clauses which become relates source dependency certified table 62 to find that in the rule 104, the search result relates destination becomes clause predicates clause It is performed by detecting the classifications S ₀ , S ₁ , ... S ₅ .

Finally, the dependency relationship recognizing unit 34
Refers to the dependency approval table 62 to determine the dependencies of the remaining elements (step 150). This is done, for example, by determining the dependency destination of the clause 1 to be the clauses 9 to 10 according to the rule 201, and changing the dependency destination of the clause 2 to the clause 4 according to the rule 301, and the final dependency analysis of FIG. The result is obtained.

Next, the input sentence is "Since the billing machine was introduced in the water supply business, it has been delivered to local governments, but after that it became an office computer and was pushed by rival companies. We will review and strengthen the project team, aiming to recover the lost land in the local government market. " This input sentence is also word-divided in the morpheme analysis unit 2 with reference to the morpheme dictionary 5 as in the case described above, and as shown in FIG.
A morphological analysis result is obtained.

The result of the morphological analysis is the dependency analysis unit 3
The predicate phrase recognizing unit 31 refers to the predicate phrase classification table 61 to detect a group of six predicate phrases as shown in FIG. 7. Then, as shown in FIG. 7, for these six predicate phrases, clause 6 is S ₁ (rule 21), clause 9 is S ₄ (rule 52), and clause 12 is S ₃ (rule 4).
1), clause 15 is S ₆ (rule 81), clause 17 is S ₃
(Rule 42), the clause 20 is classified as S ₀ (Rule 12), and the dependency between the predicate phrases is determined as shown in FIG. 7, and then the dependency of the remaining elements is determined. Dependency analysis results can be obtained.

Although the above description is based on the premise that the polymorphism of the engagement is not given in the dependency analysis for the sake of simplicity, the case of giving the polysemy of the engagement will be slightly supplemented. For example, since bunsetsu 1 in FIG. 6 has the form of “noun + ha”, the general rule is that it is related to verbs and adjectives, and the four predicates of bunsetsu 4, 6, 9, and 14 are obtained as ambiguous destinations. However, as in the present invention, if the predicate phrase is previously identified and classified, it is limited to the two clauses 9 to 10. Further, regarding the dependency of the predicate phrase itself, the general rule is to consider the five patterns shown in FIG. 8 as the dependency of the four predicates.

According to the present invention, since the number of predicate phrases becomes three, the number of patterns to be considered is reduced to two as shown in FIG. Is considered to be able to significantly reduce the ambiguity of analysis.

Further, if there are six predicates as shown in FIG. 7, 42 patterns must be taken into consideration in the general rule, whereas the present invention is limited to one pattern due to rule restrictions. Furthermore, by determining the dependency of the predicate phrase, the possibility of the dependency of the elements other than the predicate is greatly narrowed down due to the structural constraint, so it is considered that it is easy to find the correct dependency.

[0027]

As described above, according to the present invention,
A group of one or more clauses that compose a predicate phrase is extracted from the input sentence, and the predescription phrases are classified stepwise according to the strength of independence so as to match the expression structure of Japanese. The structure of the input sentence is analyzed based on the strength of independence, and the dependency relation between each phrase in the analyzed input sentence is further recognized, and the Japanese input sentence is parsed. It is possible to efficiently recognize the dependency structure even with many long Japanese sentences, and it is possible to minimize the occurrence of ambiguous ways of involvement.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a Japanese analysis system for implementing a Japanese analysis method according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a detailed configuration of a dependency analysis unit and a dependency dictionary used in the Japanese analysis system shown in FIG.

FIG. 3 is a diagram showing a configuration of a predicate phrase classification table shown in FIG.

FIG. 4 is a diagram showing a configuration of a dependency approval table shown in FIG. 2.

5 is a flowchart showing the operation of the embodiment shown in FIG.

FIG. 6 is a diagram showing an analysis result for an example of an input Japanese sentence.

FIG. 7 is a diagram showing an analysis result for an example of an input Japanese sentence.

FIG. 8 is an explanatory diagram showing a pattern of dependency of a predicate.

FIG. 9 is an explanatory diagram showing a pattern of dependency of a predicate.

[Explanation of symbols]

 1 Input part 2 Morphological analysis part 3 Dependency analysis part 4 Output part 5 Morphological dictionary 6 Dependency dictionary 31 Predicate phrase recognition part 32 Predicate phrase classification part 33 Predicate phrase relation recognition part 34 Dependency relation recognition part 61 Predicate phrase classification table 62 Dependency certification table

Claims (1)

[Claims] 1. A word included in an input sentence expressed in Japanese is identified, and a dependency relation between the phrases is identified by recognizing a dependency relation between the phrases as a basic unit of a phrase consisting of one or more words. A Japanese parsing method that performs syntactic analysis, in which a group of one or more clauses forming a predicate phrase is extracted, and the predescription phrase is graded according to the strength of independence so as to match the Japanese expression structure. And analyze the structure of the input sentence based on the strength of the independence of the predescription phrase, and further certify the dependency relation between each clause in the structure-analyzed input sentence. A Japanese parsing method characterized by parsing.