JPH04112366A - Natural language sentence analysis device - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] The present invention relates to a natural language sentence analysis device, and more specifically, to a natural language sentence analysis device.
The present invention relates to a natural language sentence parsing device used in realizing a system that takes natural language sentences as input, such as a machine translation device.

In conventional natural language sentence analysis devices, the morphological analysis processing section and the syntactic analysis processing section are usually constructed as separate modules, and then combined to form the whole. This is because the morphological analysis processing unit can be used not only in a syntactic analysis device but also in a kana-kanji conversion device, etc., so the software has high modularity and high reusability. Therefore, conventional morphological analysis processing units select and output the most likely solution according to several evaluation criteria such as the minimum number of clauses method and the longest match method. The problem with such conventional techniques is how to deal with analysis errors in the morphological analysis processing section. For example, (,) he carried out the repairs.

Considering a sentence like , if the morphological analysis processing unit splits ``gone'' into morphemes ``line <'' (online form) + ``ta'', no matter how high-performance the syntactic analysis processing unit is, it will not be able to parse it. You will fail. In this example, it is essentially ambiguous at the morphological analysis level, so two solutions are produced, ``go + ta'' and ``do + ta'', and either one is selected by the syntactic analysis processing section or the analysis result evaluation section. It should be done.

A conventional technique for solving the above problems is a method using a pack trunk. According to the hack track method, when the parsing process fails, the parsing unit requests the morphological analysis unit to output the second solution, that is, the next most likely solution. In the case of example sentence (a) above, at the time of syntactic analysis processing, "go" does not take a case, so the parsing fails, and the syntactic analysis processing section requests the output of the second solution from the morphological analysis processing section. do.

As a result, the second answer, "Ido+da" is output, and the correct answer can finally be obtained.

However, since the backtrack method is a depth-first analysis method, if an answer that is not truly correct but not incorrect is found first, there is a high possibility that the correct answer will be dropped. For example, (b) The construction work was done in April last year.

There was a problem with sentences such as ``I went'' which was split into ``Go + ta'', leaving a strange solution where ``Architecture'' was the subject of ``Go''.

Conventional techniques that attempt to solve the above problems include, for example, "On a Method of Dialogue Translation, J.
1990). This technology attempts to eliminate ambiguity in the morphological analysis processing unit through dialogue with the user. However, according to this prior art, for example: (e) I will wait until he comes.

For sentences like ``rurujuku de'' (to come) or ``kurumajude'' (by car), the user is asked. This ambiguity in the morphological analysis processing unit can be determined arbitrarily by using the knowledge in the syntactic analysis processing unit, ie, the knowledge that the subject of ``wait'' would be two if it were ``r car''. The above-mentioned conventional technology has the problem that inquiries are made to the user even in such cases, and the inquiries are often made in vain.

1-1 Loss The present invention was made in view of the above-mentioned circumstances.
The system is configured to obtain the most likely syntactic analysis result for the input sentence by not narrowing down the multiple meanings to one during the morphological analysis process, but narrowing them down to one during the syntactic analysis process and the subsequent solution selection process. The purpose was to provide a natural language sentence analysis device.

(ii) In order to achieve the above object, the present invention includes an input section for inputting a natural language sentence, a morphological analysis processing section for dividing the input sentence from the input section into morphemes, and an output from the morphological analysis processing section. the morphological analysis processing section comprises a parsing processing section that constructs a parsing tree using the parsing process, and an analysis result evaluation section that calculates the suitability of the parsing tree and selects the most likely parsing tree for the input sentence. It is characterized by outputting multiple analysis results. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a block diagram for explaining an embodiment of a natural language sentence analysis device according to the present invention. In the figure, 1 is an input section, 2 is a display section, 3 is a processing section, and 4 is a dictionary section.

5 is a morphological analysis processing section, 6 is a syntactic analysis processing section, 7 is an analysis result evaluation section, 8 is a word/dictionary section, 9 is a syntax rule section, and 10 is a constraint section.

The processing section 3 includes a morphological analysis processing section 5, a syntactic analysis processing section 6, and an analysis result evaluation section 7. Further, the dictionary section 4 includes a common word dictionary section 8, a syntax rule section 9. It consists of a constraint section 10. An example of a word dictionary section is shown as a word dictionary D, an example of a syntax rule section is shown as a syntax rule R1, and an example of a constraint section is shown as a constraint C.

An input sentence from an input unit 1 that inputs a natural language sentence is divided into morphemes by a morphological analysis processing unit 5, and a syntactic analysis tree is constructed by a syntactic analysis processing unit 6 using the output of the morphological analysis processing unit 5. The analysis result evaluation unit 7 calculates the suitability of the syntactic analysis tree and selects the most likely syntactic analysis tree for the input sentence, and the morphological analysis processing unit 5 outputs a plurality of analysis results.

Figure 2 (a) shows a 1-word dictionary.Each element of the common word dictionary is expressed by a headword, a word name (part of speech), and a feature. A feature is a list of the form (feature name, feature value). Here, features are extracted properties of various quantities. n indicates a noun, p indicates a particle, ■ indicates a verb, and sj indicates a conjunctive particle.

FIG. 2(b) shows the syntax rule R.
is written using a phrase structure grammar with labels, which is an extension of the notation of known phrase structure grammars. Each element on the right-hand side is a labeled or unlabeled nonterminal. In the case of (N P ; topic) in R1, NP is a non-terminal symbol and the label topic is added. Also, the lowercase alphabetic letters in the non-terminal symbols on the right side are pre-terminal symbols (various quantity range 11i)
, uppercase alphabetic characters are nonterminal symbols other than preterminal symbols.

Also, the labels in the syntax rules represent function names. Here, the term "function name" is used in the same manner as in the known word hall functional grammar. That is, the description of R1 is the same as the following RIO in the function grammar.

RIOVP −> NP VP↓=↑to
pic ↓:↑ Figure 2 (c) shows constraint C, and each element of constraint C is a production rule with a function name and a penalty. The notation is (Function name: Penalty) in the form of a constraint rule. Here, the penalty is a numerical value representing the grammatical strength of each constraint, and the larger the value, the stronger the constraint is grammatically, meaning that there are fewer exceptions.

Each constraint rule refers to information in the functional structure, especially features. Here, the term "functional structure" is used in the same manner as in the known functional grammar, and is a recursive matrix in which the functional name is the attribute name and the functional structure is the attribute value. FIG. 5 shows an example of the functional structure.

FIG. 3 is a flowchart of the processing section of the natural language sentence analysis device according to the present invention. Below, each step will be explained in order.

13 The user inputs sentences from the human resources department.

The input unit is a keyboard device, a voice input device, or the like. Assume that the input sentence from the input section is the above-mentioned (e).

(e) I will wait until he comes.

The morphological analysis processing unit that receives the input sentence divides the input sentence into morphemes. Here, it is assumed that the analysis result of the morphological analysis processing unit may have ambiguous meanings.

Such a configuration of the morphological analysis processing section can be realized by a known morphological analysis technique that outputs multiple solutions. For example, the literature is "Proposal of morphological analysis using minimum connection cost method and evaluation of computational complexity" (Toru Hisamitsu, Yoshihiko Nitta, Institute of Electronics, Information and Communication Engineers Technical Research Report, NLC90-08, 1990
).

For example sentence (e), (el) (I: n) (Ha: p) (He: n) (Ga:
P) (Come: V) (Until: 5j) (Wait: V) (e2)
Two solutions are output: (I: n) (Ha: p) (He: n) (Ga: p) (Car 20) (De: P) (Wait: v). Here, each morpheme is expressed in the form (headword bilingual category III). A language category is a so-called part of speech, and in syntactic rule R, it is expressed by lowercase alphanumeric characters.

扛虻I: It is determined whether the morphological analysis sequence remains.

If there are none left, go to step 4, which will be described later.

%; If a morphological analysis sequence remains in step 2, a syntactic analysis tree is created. That is, next, for the string of morphemes that is the output of the morphological analysis processing section, the syntactic analysis processing section constructs a syntactic analysis tree using the information of the syntax rule section. Here, the parse tree has a structure similar to the constituent structure in the known quantitative functional grammar, and is a labeled tree structure as shown in FIG. Labels attached to the tree structure represent function names, and can be obtained by referring to the labels described in the syntax rules when applying the syntax rules.

Regarding the process of creating a constituent structure from a sequence of morphemes,
Many methods have been proposed and are known, but here we will use CKY (Cocke-Kasan+i-Y) from the bottom up.
Constructed using the ``Ounger'' method.

When creating a constituent structure from a single morpheme sequence, the constituent structures that reach the final state when there are no more morphemes to process are all possible solutions. Since there can be multiple morpheme sequences as input for , the output of 5tep 3 generally has multiple solutions. here.

To determine whether the final state has been reached, use the syntax rule R:
It is determined whether the root note of the constituent structure created for the entire sentence is a non-terminal symbol S. One of the constituent structures (C3) created from the morpheme sequence of (el) is
As shown in the figure.

Next, it is determined whether or not a parse tree remains. If no parse tree remains, proceed to step 5, which will be described later. Note that the process described below is an example of the process of the analysis result evaluation section.

A functional structure is created from each of the plurality of component structures obtained in Step 3 above. This process may follow the process of creating a functional structure from a constituent structure in a well-known quantitative functional grammar, or if a child node in a constituent structure has no label, it is written as a main word in the entry section and labeled. For child notes, use the label as the attribute name,
The functional structure may be created recursively so that the functional structure of the child node is used as the attribute value. FIG. 5 shows a functional structure (C4) created from the component structure (C3) shown in FIG. 4. Further, in step 5, information from the dictionary section is also used, but this is the feature section in the various quantity dictionary.

Since there is a one-to-one correspondence between constituent structure and functional structure, the functional structure in the output of step 5 generally has multiple solutions. Constraints are applied to each functional structure using information from the constraint section. Here, since each constraint can be referenced by function name (attribute name of the functional structure), it is possible to recursively apply the recursive structure of the functional structure. For example, the functional structure (C4) shown in FIG. 5 is as follows. Clause ``he''
The function name corresponding to is 5ubj function, so 5ubj
Extract all the constraints labeled by the function and apply them in order. In the case of constraint C, Co, C1, and C2 are applied. In this case, the value of one's case marker feature is "ga", and the parent's functional structure has a main word of "kuru/ma J, 5ubj".
Functions are subcategorized by “kuru” and 5
Since the ubj function is the only one in the parent functional structure, both are satisfied. Next is the renyo function, but since there is no labeled constraint in C, it does nothing. Finally, topi
c function, but the value of its case marker feature is "wa", the main word of the parent's functional structure is "wait", and the empty 5ub
j function. In other words, "waiting" is a subcategory of the 5ubj function, but there is no 5ubj function in this functional structure. Therefore it is fulfilled. Also, the topic function is the only one in the parent functional structure.

If any of the above constraints are not met, the corresponding penalty will be added.

The above processing in steps 5 and 5 is performed for each component structure. As shown in FIG. 6, for the functional structure (C5) corresponding to the morpheme sequence (C2), the penalty is 30 because the constraint C2 is not satisfied.

1. Based on the penalty calculated for each constituent structure in steps 5 and 5 as described above, a constituent structure corresponding to the functional structure with the minimum penalty is selected. In this case, select (C3) and use it as the output of the parser.

In this way, the most likely syntactic analysis result for the input sentence can be obtained by narrowing down the multiple meanings to one at the time of the syntactic analysis process and the solution selection process that is the subsequent process, instead of narrowing down the multiple meanings to one during the morphological analysis process. be able to.

As is clear from the above explanation, according to the present invention, multiple meanings at the time of morpheme division can be selected after performing syntactic analysis processing without narrowing down to one at the time of morpheme division. can. Therefore, it is possible to solve the problem that the syntactic analysis processing section is more likely to fail in analysis due to an analysis failure in the morphological analysis processing section. In addition, if the ambiguity is resolved during the morphological analysis process but can be resolved during the syntactic analysis process, consider a system that resolves the ambiguity based on information from other information sources such as user inquiries and discourse analysis. Even above, you can prevent waste. In other words, instead of considering morphological analysis and syntactic analysis as separate things as in the past, there is an effect that polysemy can be considered as a whole.

[Brief explanation of drawings]

FIG. 1 is a block diagram for explaining an embodiment of a natural language sentence analysis device according to the present invention, FIG. 2 is a diagram showing a word dictionary, syntax rules, and constraints, and FIG. 3 is a diagram for explaining an embodiment of a natural language sentence analysis device according to the present invention. FIG. 4 is a flowchart of the processing unit of the natural language sentence analysis device, FIG. 4 is a diagram showing the constituent structure, and FIGS. 5 and 6 are diagrams showing the functional structure. 1... Input section, 2... Display section, 3... Processing section, 4
... Dictionary part, 5... Morphological analysis processing part, 6... Syntactic analysis processing part, 7... Analysis result evaluation part, 8... Various quantity dictionary part, 9... Syntactic rule part, 10. ...Restriction part. Figures (a) (b) (c)

Claims (1)

[Claims] 1. An input section that inputs a natural language sentence, a morphological analysis processing section that divides the input sentence from the input section into morphemes, and a syntax analysis processing section that constructs a parse tree using the output of the morphological analysis processing section. , an analysis result evaluation unit that calculates the suitability of the parse tree and selects the most likely parse tree for the input sentence, and the morphological analysis processing unit outputs a plurality of analysis results. Language sentence analysis device.