JPS62206670A - Context analysis system for natural language - Google Patents

Abstract

PURPOSE:To accurately extract the context information on an input sentence related to the correspondence by analyzing the unity between an input sentence and preceding sentence. CONSTITUTION:A preceding sentence context data storing means 3 stores a sentence structure tree in the form of th e output obtained from analysis of a sentence structure. While an input sentence to be presently analyzed is supplied to a data input means 1 and analyzed by a unity analyzing means 2. Then a context information generating means 5 produces the information against the preceding sentence and delivers it. In this case, the means 2 refers to the means 3 and a word meaning relation data storing means 4 and at the same time analyzes the connection, i.e., the presence or absence of the meaning relation between the input sentence inputted from the means 1 and the context of the preceding sentence. If the connection is recognized between both sentences, the context processing type is decided to obtain the accurate context information.

Description

[Detailed Description of the Invention] [Summary] When analyzing the context of a natural language sentence after syntactic analysis, it is possible to analyze pronoun anaphora, substitute expression anaphora, and ellipsis anaphora by means of analyzing the cohesion between the input sentence and the preceding sentence context. This is a context analysis method that can extract contextual information from input sentences and obtain good analysis results.

[Industrial Field of Application] The present invention relates to a context analysis method that is equipped with means for analyzing cohesion between an input sentence and a preceding sentence context when analyzing the context of a natural language sentence.

[Prior Art] Attempts are being made to analyze the context of natural language sentences using electronic computers. Usually, the means of analysis is the third
The configuration is shown in the figure. In FIG. 3, morphological analysis is performed on a natural language sentence, for example, a "Japanese sentence written in solid color" in step (2). Morphological analysis is the process of distinguishing between words in a sentence containing a list of words and obtaining combinations of word strings. In the following explanation, a computer performs processing while utilizing a storage device for each step. Next, in step (2), syntax analysis is performed to obtain a syntax tree. What is a syntactic tree? For a word string, find the consecutive parts of speech of each word as (noun + particle), (noun + particle), (verb auxiliary verb), (pause), and the part of speech in parentheses is the noun phrase,
It is understood as a noun phrase and a verb phrase, and when you look at them as a whole, they look like a tree with branches called "sentences", so this term refers to their structure.

Next, in step (2), context analysis is performed. At this time, regardless of the presence or absence of a break in the context, processing is normally performed to the extent of ``anaphora regarding pronouns''. Anaphora for pronouns is “he” and “it” in the current sentence.
This refers to checking which noun a pronoun such as corresponds to in the preceding sentence. The result provides a solution for processing based on linguistics.

[Problems to be Solved by the Invention] In conventional processing, particularly when context analysis is performed, no criteria are set for determining a break in context, and processing is performed as appropriate. In addition, only ``anaphora of pronouns'' was processed as a subject of context processing, and other ``anaphora of substitute expressions'' were not performed. Therefore, the results of context analysis were not good. "anaphora of substitute expressions"
refers to a sentence (word string) that uses a term other than a pronoun as a demonstrative. In the sentence (input sentence), add "man" as in "Man is Hanako's..."
This refers to the case where it is expressed as

The purpose of the present invention is to improve the above-mentioned drawbacks and provide a context analysis method to obtain good results by analyzing the "cohesiveness" of the input sentence and the preceding sentence to accurately obtain the contextual information of the input sentence regarding anaphora. The purpose is to provide a context analysis method that can extract such information.

[Means for solving the problem] Figure 1 is a block diagram showing the principle configuration of the present invention, and shows that the part that performs "context analysis" in a conventional natural language sentence analysis device is improved. . In FIG. 1, 1 is a data input means, 2 is a cohesion analysis means, 3 is a preceding sentence context data storage means, 4 is a word meaning relationship data storage means, and 5 is a data input means.
6 indicates a context information generation means, and 6 indicates a context information analysis result output means. The preceding sentence context data storage means stores a syntax tree as an output when syntax analysis is performed.

An input sentence to be currently analyzed is manually inputted to the data input means 1, analyzed by the cohesion analysis means 2, and then information is generated and outputted by the context information generation means 5 with respect to the preceding sentence.

[Operation] The cohesiveness analysis means 2 in FIG.
It analyzes the presence or absence of a ``connection'', that is, a semantic relationship, with the context of the preceding sentence, and when there is a ``connection,'' it further discovers the type of context processing and obtains accurate context information.

[Example] As specific processing of the present invention, the above-mentioned "semantic relationship analysis" and "context information generation" will be sequentially explained.

A. Regarding the semantic relationships with the preceding sentence context, the semantic relationships include (1) superior-subordinate relationships, (2) synonymous relationships, (3) attribute relationships, and (4) deep case relationships. For each relationship, as illustrated in Figure 2, (1) superior-subordinate relationship refers to ``human'' and ``male''.
``Human beings'' are in a higher rank than ``men.''

(2) A synonymous relationship is a word that has the same upper level relationship (eg, ``human'') and two words that have different lower level relationships (eg, the relationship between ``man'' and ``patient'').

(3) Attribute relationships refer to words that indicate that a person is a patient because they have a disease, such as the relationship between a patient and a disease. There is also an example of the relationship between ``human'' and ``age''.

(4) A deep case relationship refers to, for example, the relationship of ``to be hospitalized'' with ``patient'' or ``hospital.'' This is "-...
By referring to the relationship between verbs such as "to do" and other concepts,
It is said that ``to be hospitalized'' and ``patient'' have a deep hierarchical relationship. Also, rAJ indicates the subject (agent) of the action of the verb. About “to be hospitalized” and “hospital”
0'' indicates the target (object) of the action.

A specific analysis of the semantic relationship between the input sentence and the context is performed as follows. First, the words in the input sentence are looked at one by one from the left, and the meaning of each word with respect to the context is analyzed. In order to find the semantic relationship between the i-th word Wi from the beginning of the input sentence and the context, the words (Wi) in the stored preceding sentences are searched in order of proximity to the input sentence, and the above-mentioned semantic relationship between the two words is searched. To analyze. If any semantic relation R can be discovered,
It is determined that "Wi has a semantic relationship R with Wj in the context." If no semantic relationship can be found with any word in the context, r
"Wt has no relation to context and meaning."

For example, in the relationship diagram of FIG. 2, the word ``sore'' is assumed to have a semantic relationship with ``thing'', and if necessary, the specific relationship is investigated and it is determined that there is a relationship with lower-level words. Regarding the word ``he,'' it is immediately obvious that it means ``male,'' and there is a semantic relationship.

By repeating the above steps, the semantic relationship between each word in the input sentence and the context can be analyzed. If there is a word in the input sentence that has a semantic relationship with the context, this sentence is said to have "cohesion" with the context, and if no word in the input sentence has a semantic relationship with the context, then this sentence is said to have "cohesion" with the context. In other words, this sentence is said to have ``no cohesion'' with respect to the context. As long as there is cohesion, the context has a semantic connection; if there is no cohesion, the context is considered to have no semantic connection.

Types of B0 context processing When it is determined that there is a semantic relationship by the above-mentioned means, the following processing is performed depending on the type.

(1) When the semantic relationship is a superior-subordinate relationship If the word Wi in the input sentence is a noun-like pronoun (such as "it,""these,""he," etc.), "pronoun anaphora" processing is performed.

The word Wi is an adnominal pronoun (“that”, “that”, etc.)
In this case, "pronoun anaphora" processing is performed.

If the word Wi is not a pronoun, "anaphora of substitute expression" processing is performed.

(2) When the semantic relationship is an attribute relationship If the word Wi is an adnominal pronoun, it is a “pronoun anaphora”
Process.

If the word Wi is not a pronoun, "omitted anaphora" processing is performed.

(3) When the semantic relationship is a synonymous relationship If the word Wi is not a pronoun, "anaphora of substitute expression" processing is performed.

(4) When the semantic relationship is a deep case relationship When the word Wi is a verb, "anaphora of omission" processing is performed.

Since the type of processing has been determined as described above, context information is generated.

C0 Context information generation A9 Pronoun anaphora In order to make a data correspondence from the pronoun node in the syntax tree of the input sentence to the node corresponding to the referent in the context of the preceding sentence. Post a pointer.

In the case of anaphora of a substitute expression, a pointer is placed from the node of the word of the substitute expression in the syntax tree of the input sentence to the node corresponding to the referent in the context of the preceding sentence.

C. In the case of elliptical anaphors In the case of deep case relations, among the subtrees containing the verb Wi, the ellipsis phrase (ellipsis) is placed at the same level as the subtree corresponding to words in other deep case relations of the verb Wi. Insert a labeled node and point from that node to the node corresponding to the referent in the context.

After updating the input syntax tree through the above processing, the context information generation means 6 additionally stores this syntax tree at the end of the context stored in the preceding vehicle context data storage means 3.

[Effects of the Invention] Thus, according to the present invention, when an input natural language sentence is analyzed, the cohesion with the preceding context is first analyzed, and the presence or absence of a semantic relationship is determined, thereby eliminating context breaks. Since it is used as a criterion for judgment, the context analysis becomes more accurate. Furthermore, since it deals with the anaphors of substitute expressions in addition to the anaphors of pronouns, the analysis results are even better.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the principle configuration of the present invention, FIG. 2 is a diagram illustrating the semantic relationship with context, and FIG. 3 is a diagram explaining a context analysis method. 2-Cohesion analysis means 3--Antecedent sentence context data storage means 4--Word meaning relationship data storage means 5-Context information generation means Patent applicant: Fujitsu Limited Agent Patent attorney: R-physics invented by Eisuke Suzuki , Picture map 1 ←----Upper lower post correct amount times R interpretation result! Relationship Context Analysis Diagram

Claims (1)

[Claims] In a context analysis method for analyzing the context of syntactic analysis of a natural language sentence, there is provided a data input means (1), a preceding sentence context data storage means (3), and a word meaning relationship data storage means (4). , a cohesion analysis means (2), a context information generation means (5), and a context information analysis result output means (6). A method for analyzing the context of a natural language sentence, characterized in that the cohesion analysis means (2) analyzes the cohesion of an input sentence regarding pronoun anaphor, substitute expression anaphor, and omitted anaphor while referring to the cohesion analysis means (2).