JPH01113869A - Japanese sentence analyzing system - Google Patents

Abstract

PURPOSE:To obtain the best solution by using matrices showing the connotative relation and the coexistent relation between a meaning and an origin so that matching check is facilitate between predicates and syntax structure elements in terms of meaning against plural results of analysis of Japanese sentences. CONSTITUTION:A solution is taken out of an intermediate expression stack 24 as the best solution BEST and the evaluating point of the BEST is defined as a maximum point MAX. Then the next solution is taken out of the stack 24 is a second solution SECOND with its evaluating point defined as POINT. In case the point POINT of the solution SECOND is higher than the point MAX of the solution BEST, the BEST is disused and the SECOND is defined as the BEST with its evaluating point substitutes the MAX respectively. This processing is repeated until the stack 24 becomes empty. Thus the optimum solution is obtained.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a Japanese sentence analysis method, and is particularly suitable for analyzing the relationship between predicates and predicate elements when a variety of expressions are targeted, such as in a machine translation system. Regarding the method.

[Conventional technology]

As for how to understand the relationship between predicates and predicate elements.

There are two ways to understand it: grammatical roles such as "subject" and "object," and semantic roles such as "actor" and "object." In any case.

The analysis method consists of the syntactic elements required by each predicate and their surface indicators (``ga'', ``wo'').

Particles such as "ni") are patterned, and the pattern number and the pattern itself are written in the dictionary as word information of the predicate.
It is generally known that it is basically effective to perform analysis by referring to it.

In addition, in actual sentences, there may be missing indicators on the surface of syntactic elements, or there may be multiple syntactic elements corresponding to one indicator, so it is not possible to determine a single interpretation based only on the above pattern. Therefore, attempts are being made to eliminate semantically inappropriate interpretations by using semantic information.

For example, in the translation method described in Japanese Patent Application No. 60-4625, the semantic features that each syntactic element should have are described as semantic features in the syntactic pattern of predicates recorded in a dictionary.

This is used as a condition for identifying a noun phrase in an input sentence as a syntactic element of a certain predicate. Also, JP-A-61-8
As described in 2274, there is a method of preparing a "world model" made up of objects corresponding to nouns and verbs, and identifying relationships between concepts by associating words in an input sentence with this model.

[Problem that the invention seeks to solve]

The above conventional technology has the following problems.

(1) Since the predicate uses the semantic features required of syntactic elements as conditions for syntactic element recognition, the more precise the nine conditions are to avoid producing semantically inappropriate solutions, the more expressions will be accepted as input sentences. will be limited to typical ones. This has little practical value when targeting a variety of expressions, such as in machine translation. Not to mention figurative expressions, the semantic co-occurrence relationship between predicates and syntactic elements often deviates from the basics, and the ``basic'' itself is not clear, so it is necessary to treat the meaning with flexibility. It is.

(2) It is difficult or impossible to appropriately check the semantic consistency of predicates and syntactic elements because insufficient consideration is given to subsumption relationships and compatibility between semantic features.

For example, consider parsing the sentence "I also ate the new product." 1 restaurant I ate at.”

Distinguish and correctly analyze "eaten bread". In other words, in order to analyze that r-store represents the place where the action of "eating" occurred, and "bread" represents the object of "eating," the syntactic pattern of "eating" must be changed to "the person responsible for the action is / It is necessary to write the target: food as /ha/mo. on the other hand.

``New product'' can be given the semantic feature ``product'', but since it is not necessarily food, it cannot be given the semantic feature ``food''. Therefore, if you try to parse the sentence ``I also ate the new product'' using the above syntax pattern, you will not be able to give the correct interpretation that ``r new product'' is the object of ``eat.'' In order to do so, it is necessary to describe and use the information that ``Products, Cultivators, and Cultivars, which are incompatible with Food, are compatible'' in the analysis.

SUMMARY OF THE INVENTION An object of the present invention is to provide a Japanese sentence analysis method that can solve the above problems and correctly analyze various expressions by effectively utilizing information regarding the semantic features of individual words.

[Means for solving problems]

The above object is achieved by providing the following means in analyzing the relationship between predicates and syntactic elements.

(1) A means for knowing the relationship between any two semantic features by providing a semantic feature relationship table that expresses the relationships of inclusion, subsumption, compatibility, and incompatibility between semantic features; (2) Predicate and A means for evaluating the semantic appropriateness of the solution based on the relationship between the semantic features of each syntactic element and the semantic features required by the predicate for each syntactic element, based on the analysis result of the relationship between syntactic elements, and giving an evaluation point.

(3) Skill in finding and memorizing all grammatically possible solutions (4) [Means for selecting the solution with the highest evaluation score from among the numerical solutions.

[Effect]

The relationship between arbitrary semantic features X and Y is classified into the following five types (
<> represents a semantic feature).

(1) X and Y are the same (for example, when X is <machine> and Y is <machine>).

(2) X is subsumed by Y (for example, X is a machine).

(3) X subsumes Y (for example, when Y is a product or a machine).

(4) x and Y are not in an inclusive relationship and are compatible (for example,
calyx product>, when Y is <liquid>).

(5) X and Y are not in an inclusive relationship and cannot be compatible (for example, when X is a ``person'' and Y is a ``liquid'').

For every semantic feature used in the analysis, the relationships with other semantic features are memorized according to the five classifications mentioned above, so it is possible to check whether the analysis results of the relationship between grammatical predicates and syntactic elements are semantically appropriate. This can be used to make appropriate decisions when evaluating. For example, when a certain syntactic element has semantic feature X and the semantic feature required by the predicate for that syntactic element is Y, the relationship between
.

If either (2) is true, the analysis result is determined to be appropriate; if either (3) or (4) is true, it is determined to be "may be appropriate," and if (5) is true, it is determined that the analysis result is appropriate. If so, it will be judged as “inappropriate”. This makes it possible to find the optimal solution when there are multiple grammatically possible solutions. Also, X
Even if the relationship between However, this does not have the side effect of increasing the number of unparsable sentences.

[Embodiments of the invention]

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

In this embodiment, a case will be described in which the Japanese sentence analysis method of the present invention is applied to Japanese sentence analysis in a Japanese-English translation system that translates Japanese sentences into English sentences.

FIG. 2 shows the hardware configuration of a Japanese-English translation system showing an embodiment of the present invention, in which a central processing unit 1. Storage device 2. Input device 3. and an output device 4.

The storage device 2 includes a word dictionary 212, a Japanese syntax pattern dictionary 22. An English syntax pattern dictionary 232, an intermediate expression stack 24 for storing the analysis results of Japanese sentences, and a semantic feature relationship table 25 are stored.

FIG. 1 is a diagram showing a part of the semantic feature relationship table 25. As shown in FIG. A semantic feature relationship table is a table that examines the relationship between the semantic features of each syntactic element and the semantic features required by the predicate from the syntactic element based on the analysis results of the relationship between predicates and syntactic elements, and then checks whether the analysis results are semantically appropriate. This is to evaluate whether or not. As shown in FIG. 1, the semantic feature relationship table is a multivalued matrix in which each row and column corresponds to a semantic feature.

The value of the element in row i and column j indicates the relationship between the semantic feature corresponding to row i and the semantic feature corresponding to column j. The semantic feature corresponding to row i is Si, and the semantic feature corresponding to column j is Sj. When (Si, Sj) values 4 to 0 indicate the following relationship.

4:81 and Sj are the same.

3: Si is included in S and j.

2: Si subsumes Sj.

1: Si and Sj are not in an inclusive relationship and are compatible.

0: Si and Sj are not in an inclusive relationship and cannot be compatible.

In the analysis result of the relationship between a predicate and a syntactic element, let Sj be the semantic feature of the syntactic element X, and let Sj be the semantic feature required by the predicate to the syntactic element
As for the syntax element X, it can be said that there is a high probability that the analysis result is appropriate, so the value obtained and summed for all the syntax elements of the predicate is used as the evaluation score of the analysis result.

FIG. 3 is a diagram showing the data structure of the word dictionary 21 in the storage device 2. As shown in FIG. Word dictionary 21 is Japanese headword 2112
Japanese quality 212. Semantic features 213゜Japanese syntactic patterns 214. Syntax element semantic information 215, translation word 216. English parts of speech 217. This is a collection of records containing English syntax patterns 218, and can be searched using Japanese headwords 211 as keys. The syntactic element semantic information 215 represents the semantic feature of a word that is a syntactic element of the predicate. Here, rAJ is the action nominative,
``0'' means the syntactic element ``G'' is the target path, and ``+HUM'' means the person who wears it, and ``+CON'' means the concrete object.
, '+ABST' abstract object>.

'+CNTN' means that it has the semantic feature 'Kuremono'.

FIG. 4 shows the Japanese syntax pattern dictionary 1F2 in the storage device 2.
0 Japanese syntax pattern dictionary 22 which is a diagram showing a part of 2
is the modified pattern number 222 and modified pattern 2 of the Japanese syntax pattern corresponding to the Japanese syntactic pattern number 221.
23 is written, Japanese syntax pattern number 221
You can search using the key. In the modified pattern 223, the case dominated by the predicate is the surface case marker (
It is written with a case particle). The modified pattern J 10 m * J 10 n t J 10 p in Fig. 4 is a modified pattern when the predicate is related to a noun as a relation, and is an example when there are multiple solutions corresponding to one pattern on the surface layer. be.

FIG. 5 is a diagram showing a part of the English syntax pattern dictionary 23 in the storage device 2. As shown in FIG. The English syntax pattern dictionary 23 describes English syntax patterns 232 corresponding to the English syntax pattern number 231.
You can search using the key. The English syntax pattern 232 is used when generating English sentences from intermediate representations.

6(a) and 6(b) are diagrams showing the data structure of the intermediate representation stack 24 in the storage device 2. FIG.

The intermediate representation stack 24 is for storing all the solutions obtained by analysis using the Japanese syntax pattern described above, together with evaluation points. The intermediate expression represents the dependency relationship between concepts contained in a sentence.

The intermediate representation stack 24 contains node number 242 of the node corresponding to the syntactic element in the sentence with solution number 2419 and syntactic element 2.
43. Semantic feature of the word of the syntactic element 244; Dependency light node number 245 indicating the dependent light node of the syntactic element. Case relationship 246 representing the case relationship with the dependent node. Syntactic element semantic information 247 indicating the semantic feature that the predicate requires of the syntactic element when the syntactic element depends on the predicate. A syntactic element semantic evaluation score 248 indicating the consistency between the semantic feature 244 of the syntactic element and the semantic feature required by the predicate for the syntactic element. An evaluation score 249 representing the appropriateness of the entire solution is described by the sum of the syntactic element semantic evaluation points of each syntax element.

Next, the Japanese-English translation process executed by the central processing unit 1 will be explained according to the flowcharts shown in FIGS. 7 and 8.

(1) Overview of Processing FIG. 7 is a flowchart showing an overview of translation processing. First, read Japanese text from the input device 4 (step 1
1) Divide the Japanese sentence into words with reference to the word dictionary 21 (step 12).Next, perform syntactic analysis using the Japanese syntactic pattern of the predicate to find all grammatically possible intermediate expressions, and calculate the meaning. is stored in the intermediate representation stack 24 along with the evaluation score regarding the consistency of the intermediate representation (step 13), and an optimal solution is selected according to the evaluation score (step 14), and an English sentence corresponding to the intermediate representation is generated based on the English syntax pattern. (Step 15) Finally, the obtained English text is output to the device l! Output to 5.

(2) Syntax analysis FIG. 8 shows the above-mentioned "syntax analysis" step 13 in detail. This will be explained below.

Syntactic analysis is performed mainly on predicates, using the syntax-semantic patterns of predicates. First, using the Japanese syntax pattern number 214 of the predicate listed in the word dictionary 21 in FIG. 3 as a key,
The Japanese syntax pattern dictionary 22'' shown in FIG. 4 is searched (step 201).

The Japanese syntactic pattern dictionary 22 provides various deformation patterns for one pattern according to superficial variations such as active sentences, passive sentences, and relational sentences.For example, the deformation patterns in Fig. 4 JIOa, J10m,
J10n respectively.

This is a modified pattern that corresponds to sentences such as "data corresponds to memory,""data to be stored in memory," and "memory to store data." Therefore, the number of modified patterns of the searched Japanese syntax pattern is set to m, the initial value of the counter is set to 1 (step 202), and pattern matching of modified patterns 1 to m and the input sentence is performed one after another. conduct(
Step 203, Step 210). If pattern matching is successful during the process, the solution is evaluated by the following steps and stored together with the evaluation points in the intermediate representation stack 24 shown in FIGS. 6(a) and 6(b) (step 209).

(3) Evaluation of solution The evaluation of solution (steps 204 to 208) will be explained below.

First, let the number of nodes included in the intermediate representation that is the solution be n, the initial value of the node counter i be 1, and the evaluation point P for the solution.
The value of is set to the initial value O (step 204), and the evaluation point P for the solution is determined by the sum of the evaluation points for the consistency between the semantic features of each syntactic element and the semantic features expected for the syntactic element. . Therefore, for nodes 1 to n, the expected semantic feature of the node is set as Sj (Step 205), and if Sj is not empty, the consistency of Sj with the semantic feature Si that the node actually has (Step 206) is determined. is evaluated to obtain an evaluation point for the node, and this is added to the evaluation point P (steps 207 and 208).
). To evaluate the consistency between the semantic feature Si that a node actually has and the semantic feature expected of the node, the semantic feature relationship table shown in FIG. 1 is referred to, and the value of (Si, Sj) is used as the evaluation point.

Through the above steps, all grammatically possible solutions are stored in the intermediate representation stack along with evaluation points, as shown in FIGS. 6(a) and 6(b). FIGS. 6(a) and 6(b) show the analysis results of "data to be stored" and "memory to be stored", respectively. Regarding "data to store", from the syntax pattern, "data" is the operation (A) of "to store",
Target (B).

Three interpretations can be obtained: the end point (G), but “data (
The interpretation ``stores data (object)'' is given a score of 0, and the interpretation ``stores data (object)'' is given a score of 4.
Score 1 for the interpretation “store in data (end point)”
is given.

(4) Selection of optimal solution The process of selecting the optimal solution based on the above evaluation points will be explained according to the flowchart of FIG.

First, one solution is taken out from the intermediate representation stack 24 and set as the optimal solution BEST (step 301), and its evaluation score is set as the highest score MAX (step 302).

Furthermore, the next solution is taken out from the intermediate representation stack 24,
Alternative solution 5 ECOND and step 303), the evaluation point as evaluation point POINT (step 304), Alternative solution 5
If the evaluation point POINT of ECOND is higher than the evaluation point MAX of the optimal solution BEST, the optimal solution BEST is rejected, the alternative solution 5ECOND is set as the optimal solution 11EST, and the evaluation point is substituted for the highest point MAX (step 305). 'Step 303~ until the expression stack is empty.
By repeating the process in step 305, an optimal solution is obtained. Due to this.

As the optimal solutions for "stored data" and "stored memory," intermediate expressions based on the interpretations of "data (object) was stored" and "stored in memory (end point)" are selected, respectively, so 'data' whichwas 5
tored', 'memory where it is
It becomes possible to translate it as tzas 5tored'.

Note that if there is ambiguity in the interpretation, such as in ``stored memory'', a more appropriate interpretation will be chosen, so the interpretation that ``r memory'' is the target of ``stored'' will not be adopted. , when there is no ambiguity such as "memory was stored", the interpretation that "memory" is the target of "storing" is adopted as the optimal solution, regardless of the evaluation score.

〔Effect of the invention〕

As explained above, (1) according to the present invention, the semantic consistency of predicates and syntactic elements can be checked for multiple Japanese sentence analysis results by using a matrix that indicates subsumption relationships (2) compatibility relationships between semantic features. can be easily and accurately performed to find the optimal solution. This has the effect of improving the accuracy of Japanese sentence analysis, since detailed semantic information can be used as a decisive factor in analysis of various expressions as needed.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing part of the contents of a semantic feature relationship table in an embodiment of the present invention, FIG. 2 is a block diagram of a Japanese-to-English translation system in an embodiment of the present invention, and FIG. A diagram showing the data structure of the word dictionary of the system, Figure 4 shows a part of the contents of the Japanese syntax pattern dictionary of the system, and Figure 5 shows part of the contents of the English syntax pattern dictionary of the system. 6 is a diagram showing an example of how the intermediate representation stack of the system is used, and FIGS. 7 to 9 are flowcharts showing the flow of Japanese-English translation processing in the embodiment. 21...Word dictionary, 22...Japanese syntactic pattern. 24...Intermediate expression stack, 25...Semantic feature relation agent Patent attorney Katsuma Ogawa Miko'3rd / (!l 2nd mouth 25 Hiroshi, 7 L of opening brush Qi of Camphor character 3rd I!] Condolence 4 Figure 2” %?l'l'! 2/7-J'ttZ
a 輯222 '14f5! -', 7452/4 Sunne 1 glance Yoshinojo pattern Kiyoshi (No. 2/1? central control scoop 1
Jōhe Scene No. 1 223 Kuren High Turn No. 5 Part 6 Country (α) (G) Tsumugi 7 Figure No. 31!1 Part 9

Claims (1)

[Scope of Claims] 1. In a Japanese sentence analysis system having a means for analyzing Japanese sentences, semantic features representing the semantic features of words and, for predicates, semantic features that should be possessed by words that are syntactic elements of the predicate are used. When describing the predicate in a word dictionary and analyzing the relationship between the predicate and the syntactic element, find all grammatically possible solutions, and evaluate the semantic consistency of the predicate and the syntactic element from the information in the word dictionary, A Japanese sentence analysis method that is characterized by selecting the optimal solution based on the evaluation score. 2. Claim 1, characterized in that, in evaluating the semantic consistency between the predicate and the syntactic element, a semantic feature relationship table is used that expresses subsumption relationships and compatibility relationships between semantic features as matrix values. Japanese sentence analysis method described in section.