JP2000148755A - Method and device for semantic analysis and record medium where semantic analysis program is recorded - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and device for semantic analysis which shorten the execution time of analysis by specifying a deep-structure case needed for later key word extraction and generating a semantic structure having a minimum necessary size without analyzing the semantics in other unnecessary deep- structure cases and lighten the load of processing system maintenance by showing the flow of analysis explicitly and to provide recording medium storing the semantic analysis program. SOLUTION: It is decided (S1) whether the object of analysis is a verbal phrase having a verb as a subject, a non-verbal phrase having mainly a noun as a subject or a part of speech by referring to analysis object section classification knowledge included in information which is appended to the analyzed object and is a sub-classification of a phrase. When the analyzed object is a verbal phrase, the relation between the verb and the deep-structure case is analyzed (S2) and when the analyzed object is a non-verbal phrase, the modification relation including a nominal phrase and an adjectival phrase is analyzed (S3). When the analyzed object is a part of speech, the semantic analysis is stopped.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for semantic analysis, and a recording medium on which a semantic analysis program is recorded. Performs analysis, syntax analysis, and semantic analysis, and as a result, extracts a keyword to be used as a search key from the obtained semantic structure, and uses the extracted keyword to search for the category name of the search target corresponding to the natural language sentence of the query And a storage medium storing a semantic analysis program in an information search system capable of outputting a set of search targets belonging to the category name.

[0002]

2. Description of the Related Art As a conventional semantic analysis method, an extension for generating a semantic structure by going backward from the part-of-speech node of a syntax tree as an input to the node of a sentence (S) as a top node from the bottom up. There is a method such as LINGOL. In the extended LINGOL, a semantic processing function is described in each grammar rule used for syntactic analysis, and after the syntactic analysis is completed, these semantic processing functions are activated to perform semantic analysis. This method can be referred to in "Natural Language Processing" by Shun Ishizaki (Shokodo).

[0003]

However, in the information guidance service, when an inquiry from a user is a natural language sentence, morphological analysis, syntax analysis, and semantic analysis are performed, and a search is performed from the semantic structure obtained as a result. In an information retrieval system capable of extracting a keyword used as a key, a keyword to be extracted often has a specific deep case of a verb. For example, "I want
To buy flowers in Tokyo. ”, the keywords to be extracted are the verb“ buy ”, the product“ flowers ”, and the address“ Tokyo ”. It is the head of the infinitive phrase "to buy flowers in Tokyo" that comes in the target case of "want", and its target case and place.

[0004] As described above, the deep case necessary for keyword extraction is determined for each information guide service, and other meaning structures inside the deep case are unnecessary for keyword extraction. In the conventional bottom-up search, since the semantic analysis inside the deep case that does not require keyword extraction after the semantic analysis is performed, there is a problem that efficiency is deteriorated.

In the method of describing a semantic processing function in each grammar rule used for syntactic analysis, since a semantic processing function is created individually for each grammar rule, it is created while grasping the flow of the entire semantic analysis. There is a problem that is difficult. The present invention has been made in view of the above points, and designates a deep case necessary for keyword extraction later, without performing semantic analysis inside other unnecessary deep cases, and in a necessary minimum range. A semantic analysis method and apparatus and a semantic analysis program that reduce the execution time of an analysis by generating a semantic structure and reduce the burden of maintenance of a processing method by making the entire analysis flow explicit An object of the present invention is to provide a recording medium on which is recorded.

[0006]

To achieve the above object, the present invention is configured as follows. The present invention provides a semantic analysis method for analyzing a semantic structure of a natural language sentence, wherein a step of inputting a syntax tree of the natural language sentence and a subclassification of a phrase attached to a node to be analyzed of the syntax tree are performed. From the information of a certain type, referring to the analysis target classification knowledge, it is determined whether the analysis target is a verb phrase having a verb as a head or a non-verb phrase mainly having a noun as a head. If the analysis target classification determination process, the analysis target is the verb phrase, a verb and a verb phrase analysis process for analyzing the relationship between the deep case of the verb, and the analysis target is the non-verb phrase If so, a non-verb phrase analysis process of analyzing a modification relation including a noun phrase and an adjective phrase, a semantic frame corresponding to a certain node of the syntax tree, a head relation connecting the semantic frames, and a deep case relation The natural language, consisting of a set of links in a qualifying relationship Comprising a step of outputting the semantic structure of the sentence.

In the above arrangement, the verb / phrase analysis step obtains a verb node to be the subject to be analyzed with reference to the positional knowledge on the parse tree, and sets a head / verb frame corresponding to the head / verb node. The subject verb frame extraction process of linking a subject relation link from the analysis target frame corresponding to the analysis target to the head verb frame, and referring to information on a state attached to the node including the analysis target to perform the analysis. A state determining step of determining an object state, an extraction target deep case determining step of determining a deep case to be extracted from the deep case described in the verb semantic information attached to the head verb node, A parse tree node corresponding to the deep case determined in the deep case determination process is obtained, a deep case frame corresponding to the parse tree node is set, and the deep case frame is determined from the main verb frame. As a result of the deep case frame extracting step of linking the deep case frame to the stratified case frame and the deep case frame extracting step, if the deep case frame is set, the semantic analysis is recursively performed using the deep case frame as an input. A deep case frame analysis process to be called, and when the deep case frame analysis process is completed, the extraction target deep case determination process is performed again. As a result of the processing in the extraction target deep case determination process, the deep case to be extracted If there is no case,
The semantic analysis including the verb phrase analysis process is completed.

In the deep case frame extracting step, the state determined in the state determining step and the verb meaning information attached to the head verb node are referred to.
A surface case information obtaining step of obtaining a surface case corresponding to the deep case determined in the extraction target deep case determination step, and obtaining a parse tree node corresponding to the surface case based on the surface case and the parse tree position knowledge. A syntax tree node obtaining step, a constraint condition matching step of matching information including phrase, type, and semantic information associated with the syntax tree node with the deep case constraint condition knowledge described in the verb semantic information. If the result of the constraint matching process is true, a deep case frame generating step of setting a deep case frame corresponding to the parse tree node and linking a deep case relation from the head verb frame to the deep case frame Consists of

In the non-verb phrase analysis process, a phrase or a part of speech node which is a subject to be analyzed is obtained with reference to the positional knowledge on the parse tree, and a subject frame corresponding to the subject node is set. A subject frame extraction process of linking a subject relationship from the analysis target frame corresponding to the analysis subject to the subject frame, a subject frame analysis process of recursively calling the semantic analysis by using the subject frame as an input, With reference to the position knowledge, a modifier node for modifying the head node is obtained, a modifier frame corresponding to the modifier node is set, and a modifier link that links a modifier relationship from the head frame to the modifier frame is provided. A frame extracting step, a qualifying frame extraction determining step of determining whether a qualifying phrase frame is set as a result of the qualifying frame extracting step, As a result of the processing of the frame extraction determination step, if the modifier frame is set, the modifier frame analysis step of recursively calling the semantic analysis with the modifier frame as an input, wherein the modifier frame analysis step is performed. When the processing is completed, the processing of the qualifying phrase frame extraction process is performed again. If the qualifying phrase frame is not determined as a result of the qualifying phrase frame extraction determining step, the semantic analysis including the non-verb phrase class analysis step is ended. .

[0010] The present invention is also an information retrieval method using the above-described semantic analysis method. The information retrieval method is executed on a computer having a storage device and an interface unit, and converts a natural language sentence through the interface unit. Inputting, performing a morphological analysis and syntax analysis on the natural language sentence, generating a syntax tree of the natural language sentence, performing the semantic analysis using the syntax tree as input, Generating a semantic structure; extracting a keyword representing a user's intention from the semantic structure; and using the keyword to refer to at least one knowledge database stored in the storage device, Acquiring the category name and the address information corresponding to, and using the category name and the address information, classifying by the category name and the address information Obtaining the search target information by referring to the information database including the search target information, and transmitting the obtained information to another computer via the interface unit, or displaying the obtained information on a display connectable to the computer Process.

In order to achieve the above object, the present invention can be configured as follows. The present invention is a semantic analysis device for analyzing a semantic structure of a natural language sentence, and a means for inputting a syntax tree of a natural language sentence, and an analysis target classification for determining whether the verb phrase or the non-verb phrase is With reference to the analysis target category classification knowledge, from the information of the type that is the sub-classification of the phrase attached to the analysis target category classification knowledge storage device in which the classification knowledge is stored and the node to be analyzed of the syntax tree, Analysis target classification determining means for determining whether the analysis target is a verb phrase class having a verb as a head or a non-verb phrase mainly having a noun as a head, and a syntax for obtaining a node of a syntax tree A syntactic tree position knowledge storage device in which tree position knowledge is stored, and a verb phrase class analyzing means for analyzing a relationship between the verb and the deep case of the verb if the analysis target is the verb phrase class. ,
If the object to be analyzed is the non-verb phrase, a non-verb phrase analysis means for analyzing a modification relation including a noun phrase and an adjective phrase; a semantic frame corresponding to a certain node of the syntax tree; Means for outputting a semantic structure of the natural language sentence, comprising a set of links of a head relation, a deep case relation, and a modification relation that connect between them.

Further, the verb phrase class analyzing means obtains a verb node to be the subject to be analyzed with reference to the position knowledge on the parse tree, and sets a main verb frame corresponding to the main verb node, A subject verb frame extracting means for linking a subject relation from the analysis target frame corresponding to the analysis target to the head verb frame, and information on a state attached to the node including the analysis target by referring to information of the analysis target; State determination means for determining a state, extraction target deep case determination means for determining a deep case to be extracted from the deep case described in the verb semantic information associated with the head verb node, and the extraction target deep case A parse tree node corresponding to the deep case determined by the determining means is obtained, a deep case frame corresponding to the parse tree node is set, and the deep case frame is determined from the head verb frame. A deep case frame extraction means to link the deep case relations to arm, the result of the processing of the deep case frame extracting means, if set the deep case frame,
Deep case frame analysis means for recursively calling the semantic analysis with the deep case frame as an input, and when the processing of the deep case frame analysis means ends, performs the processing of the extraction target deep case determination means again, If there is no deep case to be extracted as a result of the processing in the extraction target deep case determination means, the semantic analysis including the processing of the verb phrase class analysis means is terminated.

Further, the deep case frame extracting means refers to the state determined in the processing of the state determining means and the verb meaning information attached to the head verb node, thereby obtaining the deep case determining object to be extracted. Surface case information obtaining means for obtaining a surface case corresponding to the deep case determined in the processing of the above, and a syntax tree node obtaining means for obtaining a syntax tree node corresponding to the surface case based on the surface case and the position knowledge on the syntax tree Constraint condition matching means for matching information including a phrase, a type, and semantic information attached to the syntax tree node with constraint knowledge of the deep case described in the verb semantic information; If the result of the processing by the means is true, a deep case frame corresponding to the parse tree node is set, and a deep case relation link is established from the head verb frame to the deep case frame. Consisting of a deep case frame generating means.

The non-verb phrase analysis means refers to the position knowledge on the parse tree to find a phrase or a part of speech node to be the subject of the analysis, and sets a subject frame corresponding to the subject node. A subject frame extracting means for linking a subject relation from the analysis target frame corresponding to the analysis subject to the subject frame, and a subject frame analyzing means for recursively calling the semantic analysis with the subject frame as an input;
By referring to the position knowledge on the parse tree, a modifier node for modifying the head node is obtained, a modifier frame corresponding to the modifier node is set, and a modification relation link from the head frame to the modifier frame is provided. Qualifier frame extraction means, and a result of the processing of the qualifier frame extraction means, a qualifier frame extraction determination means for determining whether a qualification phrase frame is set, and a result of the processing of the qualification phrase frame extraction determination means, If the modifier frame is set, the modifier frame analyzing means for recursively calling the semantic analysis by inputting the modifier frame is provided. The processing of the qualifying phrase frame extracting means is performed, and as a result of the processing of the qualifying phrase frame extracting / determining means, if the qualifying phrase frame is not set, the To end the semantic analysis, including the processing of phrase such analysis means.

Further, the present invention is an information retrieval apparatus using the above-mentioned semantic analysis apparatus, comprising: a storage device and an interface unit; means for inputting a natural language sentence via the interface unit; Means for performing morphological analysis and syntax analysis to generate a syntax tree of the natural language sentence, means for performing the semantic analysis using the syntax tree as input, and generating a semantic structure of the natural language sentence, Means for extracting a keyword representing the intention of the user, and means for using the keyword and referring to at least one knowledge database stored in the storage device to obtain category name and address information corresponding to the natural language sentence Using the category name and the address information, and referring to an information database including search target information classified by the category name and the address information. Means for obtaining an elephant information, or transmitted to another computer acquired information via the interface unit, or, and means for displaying a connectable display to the computer.

In order to achieve the above object, the present invention can be further configured as follows. The present invention provides a method for inputting a syntax tree of a natural language sentence in a recording medium storing a semantic analysis program for analyzing a semantic structure of a natural language sentence, and a procedure attached to a node to be analyzed in the syntax tree. From the information of the type that is the sub-classification of the phrase, referring to the analysis target classification classification knowledge, the analysis target is a verb phrase having a verb as a head or a non-verb phrase having a noun as a head. An analysis target category determination procedure for determining whether there is a verb phrase class if the analysis target is the verb phrase class, and a verb phrase class analysis procedure for analyzing the relationship between the verb and the deep case of the verb; If it is a non-verb phrase, a non-verb phrase analysis procedure for analyzing a modification relation including a noun phrase and an adjective phrase; a semantic frame corresponding to a certain node of the syntax tree; Links for relationships, deep case relationships, and qualification relationships Consisting case, and a procedure for outputting the semantic structure of the natural language text.

Further, the verb phrase class analysis procedure obtains a verb node to be the subject to be analyzed with reference to the position knowledge on the parse tree, sets a head verb frame corresponding to the head verb node, and sets By referring to a subject verb frame extraction procedure for linking a subject relation verb from the analysis target frame corresponding to the analysis subject to the subject verb frame, and information on a state attached to the node including the analysis subject, the state of the analysis target is determined. A state determination procedure for determining, from the deep case described in the verb semantic information associated with the head verb node, an extraction target deep case determination procedure for determining a deep case to be extracted,
A parse tree node corresponding to the deep case determined in the extraction target deep case determination procedure is obtained, a deep case frame corresponding to the parse tree node is set, and a deep case relationship from the main verb frame to the deep case frame is set. A deep case frame extracting step of establishing a link, and a deep case frame analyzing step of recursively calling the semantic analysis with the deep case frame as an input if the deep case frame is set as a result of the deep case frame extracting procedure. When the deep case frame analysis procedure is completed, the processing of the extraction target deep case determination procedure is performed again.As a result of the processing in the extraction target deep case determination procedure, if there is no deep case to be extracted, The semantic analysis including the verb phrase analysis procedure is ended.

Further, the deep case frame extraction procedure refers to the state determined in the state determination step and verb semantic information attached to the head verb node.
A surface case information obtaining step of obtaining a surface case corresponding to the deep case determined in the extraction target deep case determination step, and a syntax tree node corresponding to the surface case based on the surface case and the parse tree position knowledge. A syntax tree node acquisition procedure, a constraint condition matching procedure for matching information including phrase, type, and semantic information associated with the syntax tree node, and the deep case constraint knowledge described in the verb semantic information; If the result of the constraint matching procedure is true, a deep case frame generating step of setting a deep case frame corresponding to the parse tree node and linking a deep case relation from the head verb frame to the deep case frame Consists of

Further, the non-verb phrase class analysis procedure refers to the positional knowledge on the parse tree to find a phrase or a part of speech node to be the subject to be analyzed, and sets a subject frame corresponding to the subject node, A subject frame extraction procedure for linking a subject relationship from the analysis target frame corresponding to the analysis subject to the subject frame, a subject frame analysis procedure for recursively calling the semantic analysis with the subject frame as an input, With reference to the position knowledge, a modifier node for modifying the head node is obtained, a modifier frame corresponding to the modifier node is set, and a modifier link that links a modifier relationship from the head frame to the modifier frame is provided. A frame extraction procedure, a qualifier frame extraction determination procedure for determining whether a qualifier frame is set as a result of the processing of the qualifier frame extraction procedure, As a result of the processing of the decoration phrase frame extraction determination procedure, if the decoration phrase frame is set, the decoration phrase frame analysis step of recursively calling the semantic analysis with the decoration phrase frame as an input, When the processing of the procedure is completed, the processing of the qualifying phrase frame extraction procedure is performed again. As a result of the processing of the qualifying phrase frame extraction determining procedure, if the qualifying phrase frame is not set, the non-verb phrase class analysis procedure is performed. Terminates the semantic analysis including.

Further, the present invention is a recording medium recording an information search program having the above-mentioned semantic analysis program, and the information search program is executed on a computer having a storage device and an interface unit, and is executed via the interface unit. Inputting a natural language sentence, performing a morphological analysis and syntax analysis on the natural language sentence, generating a syntax tree of the natural language sentence, performing the semantic analysis using the syntax tree as input, Generating a semantic structure of the natural language sentence, extracting a keyword representing a user's intention from the semantic structure, and using the keyword to refer to at least one knowledge database stored in the storage device. Acquiring a category name and address information corresponding to the natural language sentence, and using the category name and the address information to acquire the category. A step of acquiring the search target information by referring to the information database containing the search target information classified by Lee name and 該住 office information, or transmits the acquired information to another computer via the interface unit,
Or, a procedure for displaying on a display connectable to the computer.

As described above, in the semantic analysis according to the present invention, a semantic structure is generated by going back up the syntax tree from the node of the sentence (S), which is the top node of the input syntax tree, to the part-of-speech node. . When going down the parse tree from the top down, for each phrase type, the knowledge that defines the head of the phrase, the surface case for the head, and the position of the qualifier that qualifies the head in the parse tree is retained and referenced. Therefore, a semantic processing function is not described for each grammar rule as in the related art.

As described above, in order to proceed the analysis from the top down, only the phrase of the specified deep case that is necessary for keyword extraction, starting from the head of the sentence, is analyzed.
Unnecessary analysis inside the deep case is eliminated.
In addition, since the knowledge that prescribes the position of the head, the surface case for the head, and the position of the qualifier that modifies the head in the parse tree is held for each phrase type, the flow of the entire semantic analysis is visually It is easy to understand, and maintenance such as correction of the processing method becomes easy.

Further, according to the information retrieval method of the present invention, since the above-described semantic analysis method is used, the flow of the entire semantic analysis can be easily understood visually, and the information retrieval that is easy to maintain such as correction of the processing method can be performed. It can be realized. Other features and advantages of the present invention will be described in the embodiments of the present invention with reference to the drawings.

[0024]

FIG. 1 is a diagram for explaining the principle of a semantic analysis method according to the present invention. As shown in the figure, first, a node to be analyzed is input as a syntax tree, and information on a type of a phrase or part of speech and a sub-classification of a phrase attached to the analysis target is obtained. To determine whether the analysis target is a verb phrase class having a verb as a head, a non-verb phrase mainly having a noun as a head, or a part of speech, and outputting the determination result. An analysis target classification determination process (step 1) is executed.

Subsequently, if the object to be analyzed is a verb phrase, the frame to be analyzed corresponding to the object to be analyzed is input, and the relationship between the verb and the deep case of the verb in the frame to be analyzed is analyzed. Then, a verb / phrase analysis process (step 2) for outputting the semantic structure in the analysis target frame is executed. If the analysis target is a non-verb phrase, the analysis target frame corresponding to the analysis target is input, and the modification relationship including the noun phrase and the adjective phrase in the analysis target frame is analyzed. The non-verb phrase analysis process (step 3) which outputs the semantic structure of is executed. If the analysis target is a part of speech, the semantic analysis ends.

FIG. 2 is a flowchart showing an example of an information retrieval method using the above-mentioned semantic analysis method. As shown in the figure, in the natural language input type information search method, first, a natural language sentence is input, and a set of a set of attribute information such as a constituent word of the natural language sentence and a part of speech of the constituent word is output. The morphological analysis process (step 11) is executed. Next, the output result of the morphological analysis process (step 11) is used as an input, and the syntax tree of the natural language sentence, in which each node is accompanied by information such as the type of a phrase or part of speech, the type of sub-classification of the phrase, and semantic information. Is executed (step 12).

Then, the syntax tree of the natural language sentence is input and
A semantic frame corresponding to a certain node of the syntactic tree, a head relation (a relation between a phrase and a phrase / part of speech that determines many syntactic properties in the phrase), a deep case relation ( It is composed of a set of links of a verb and a noun phrase corresponding to the target case, place case, etc. of the verb, and a modification relationship (a relation between the noun phrase and an adjective phrase modifying the noun phrase). Executing a semantic analysis process (step 13) for outputting the semantic structure of the natural language sentence, inputting the semantic structure of the natural language sentence,
A keyword representing the intention of the searcher in the natural language sentence is extracted, and a keyword extracting step (step 14) of outputting the keyword is executed.

Subsequently, a keyword extraction process (step 1)
4) With the keyword output as an input, execute a category name inference process (Step 15) of outputting a category name of a search target corresponding to a natural language sentence, and input an output result of a morphological analysis process (Step 11), A word having attribute information such as a designated part of speech is extracted, a word extraction step (step 16) for outputting the word is executed, and a keyword corresponding to an address output in the keyword extraction step (step 14) is input, and the keyword An address analysis step (step 17) of outputting address information corresponding to the address is executed.

Finally, the category name output by the category name inference step (step 15) and the address information output by the address analysis step (step 17) are input, and the address information is searched for and belongs to the category name. A search is performed for a set of items to be searched, and a word having the word output as a keyword in the word extraction process (step 16) is assigned a higher priority, and a search process (step 18) for outputting the set of search targets is executed.

FIG. 3 is a diagram showing the configuration of the information retrieval apparatus of the present invention. The information search device 2 shown in FIG. 1 includes an overall control device 3, a natural language processing device 4, and a search device 26. The general control device 3 receives the natural language sentence input from the input device 1 and sends it to the natural language processing device 4. Next, the natural language processing device 4 receives the search target category name, address information, and extracted words corresponding to the natural language sentence. This category name, address information, extracted words,
Send to search device 26. Next, the search device 26 outputs
A priority-ordered set of search targets corresponding to the category name and the address information is received and sent to the display device 29. In addition,
Each search target is accompanied by detailed attribute information. When a specific search target is specified from the input device 1, the general control device 3 displays the detailed attribute information of the search target on the display device 2.
It is also possible to send to 9.

The natural language processing device 4 comprises a morphological analysis device 5, a syntax analysis device 8, a semantic analysis device 11, a keyword extraction device 14, a category name inference device 17, a word extraction device 20, and an address analysis device 23. Control device 3
It outputs the corresponding search target category name, address information, and extracted words from the natural language sentence sent from. In the morphological analysis device 5, a natural language sentence is input, and the morphological analysis unit 6 refers to the word dictionary 7 to generate a set of a set of attribute information such as a constituent word of the natural language sentence and a part of speech of the constituent word. Output.

In the syntax analysis device 8, the output result of the morphological analysis device 5 is input, and the syntax analysis unit 9 refers to the phrase structure rule table 10 and classifies each node with the type of phrase and part of speech, and the fine classification of the phrase. And outputs a syntax tree of a natural language sentence accompanied by information such as type and semantic information. Semantic analysis device 11
, A syntax tree of a natural language sentence is input, and the semantic analysis unit 12 refers to the various knowledge storage devices 13 and a semantic frame corresponding to a certain node of the syntax tree and a headword that connects the semantic frames. Relationships (relationships between phrases and phrases and parts of speech that determine many of the syntactic properties in the phrases), deep case relationships (relationships between verbs and noun phrases that correspond to verb target cases, place cases, etc.) And outputs a semantic structure of a natural language sentence composed of a set of links of a modification relation (a relation between a noun phrase and an adjective phrase modifying the noun phrase).

The keyword extracting device 14 receives the semantic structure of a natural language sentence as an input,
Refers to the extraction rule table 16 and extracts a keyword indicating the intention of the searcher in the natural language sentence, and outputs the keyword. In the category name inference device 17,
With the keyword output from the keyword extraction device 14 as an input, the category name inference unit 18 refers to various knowledge databases 19 such as a thesaurus and outputs a category name of a search target corresponding to a natural language sentence.

In the word extraction device 20, the output result of the morphological analysis device 5 is input, and the word extraction unit 21 refers to the extraction target word table 22 to extract a word having attribute information such as a designated part of speech. , And outputs the word. In the address analysis device 23, the keyword corresponding to the address output by the keyword extraction device 14 is input, and the address analysis unit 2
4 outputs the address information corresponding to the keyword with reference to the address table 25.

The search device 26 includes a search unit 27 and a guidance information database 28. The guidance information database 28 classifies a set of search targets by category name and address information. Further, each search target is accompanied by detailed attribute information, which includes the keyword of the search target. In the search device 26, the category name, the address information, and the extracted words sent from the overall control device 3 are input, and the search unit 27 refers to the guidance information database 28 to search for the objects belonging to the category name. A set of those having the address information is searched, and a set having the extracted word as a keyword is prioritized as a higher order, and the set of search targets is output.

Note that, in addition to displaying the search target information on the display device 29, it is also possible to provide an interface unit for communicating with another device to transmit the search target information to another device. FIG. 4 shows the configuration of the semantic analysis device of the present invention. This semantic analysis device corresponds to the semantic analysis device 11 in the information retrieval device 2 described above.

The semantic analysis apparatus 1000 shown in FIG. 3 includes an analysis target category determination unit 100, a verb phrase analysis unit 200, a non-verb phrase analysis unit 300, and the type of analysis target (corresponding phrase and its sub-classification type). For each of the combinations), an analysis target category classification knowledge storage device 400 storing knowledge that classifies the analysis target into one of a verb phrase, a non-verb phrase, and a part of speech, and
For each type of analysis target, the knowledge that defines the head of the analysis target, the surface case for the head, and the position on the parse tree of the qualifier that qualifies the head, and the node corresponding to the surface case for the head are input and It is configured from a syntax tree position knowledge storage device 500 that stores knowledge defining the position of the destination of the surface case when it is not on the syntax tree.

The verb phrase class analysis unit 200 includes a head verb frame extraction unit 210 and a target head verb frame determination unit 22 to be processed.
0, a state determination unit 230, an extraction target deep case determination unit 240, a deep case frame extraction unit 250, a processing target deep case frame determination unit 260, and a deep case frame analysis unit 270. The head verb frame extraction unit 210 receives the analysis object as input, refers to the syntax tree position knowledge stored in the syntax tree position knowledge storage device 500, obtains a verb node to be the head of the analysis object, and By setting a head verb frame corresponding to the verb node and linking a head relation from the analysis target frame corresponding to the analysis target to the head verb frame, a part of the semantic structure in the analysis target frame is output.

Processing target head / verb frame determination unit 220
Inputs a head verb frame, determines which head verb frame of the head verb frame is to be analyzed in the subsequent processing, and outputs the head verb frame.
The state determination unit 230 receives the analysis target and the subject verb node as inputs, determines the state to be analyzed by referring to information on the state attached to the node including the analysis target, and outputs the state.

The extraction target deep case determination unit 240 receives the subject verb node as input, determines the deep case to be extracted from the deep case described in the verb meaning information associated with the subject verb node, and Output case. The deep case frame extraction unit 250 receives the analysis target, the head verb node, the deep case, and the state as inputs, obtains a syntax tree node corresponding to the deep case, and has not set a deep case frame corresponding to the syntax tree node. Then, a part of the semantic structure in the analysis target frame is output by linking a deep case relationship from the head verb frame to the deep case frame. Details will be described later.

The processing target deep case frame determination unit 260
The deep case frame newly set by the deep case frame extraction unit 250 is input, and it is determined which of the deep case frames is to be analyzed, and the deep case frame is output. The deep case frame analysis unit 270 receives the deep case frame as an input, and recursively calls a semantic analysis to output the semantic structure in the deep case frame.

The non-verb phrase class analyzer 300 includes a subject frame extractor 310, a subject subject frame determiner 320, a subject frame analyzer 330, and a modifier phrase frame extractor 34.
0, a processing target modifier phrase frame determiner 350 and a modifier phrase frame analyzer 360. The head frame extraction unit 310 receives the analysis target as input, refers to the position knowledge on the parse tree, and obtains a phrase or part of speech node that is the head of the analysis target,
By setting a head frame corresponding to the head node and linking a head relation from the analysis target frame corresponding to the analysis target to the head frame, a part of the semantic structure in the analysis target frame is output.

The processing subject head frame determination unit 320 receives the head frame, determines which head frame of the head frames is to be analyzed in the subsequent processing, and outputs the head frame. The head frame analysis unit 330
The semantic analysis in the subject frame is output by recursively calling the semantic analysis with the subject frame as an input.

The modifier phrase frame extractor 340 receives the analysis target and the head node as inputs, finds a modifier node that modifies the head node with reference to the position knowledge on the syntax tree, and obtains a modifier corresponding to the modifier node. If the phrase frame has not been set, it is set, and a part of the semantic structure in the analysis target frame is output by linking a modification relationship from the head frame to the modifier phrase frame.

The processing target modifier frame determining unit 350
The qualifier phrase frame newly set by the qualifier phrase frame extraction unit 340 is input, and it is determined which of the qualifier phrase frames is to be analyzed, and the qualifier phrase frame is output. The processing target modifier frame determination unit 3
Since 50 can also determine whether or not there is a qualifying phrase frame to be processed, it may be referred to as a qualifying phrase frame extraction determination unit. Further, also in the case of the process, the process of determining the processing target modifier phrase frame may be referred to as a modifier phrase frame extraction determination process.

The modifier phrase frame analyzer 360 receives the modifier frame as an input, and recursively invokes semantic analysis to output the semantic structure in the modifier frame.
In the configuration of the semantic analysis device shown in FIG. 4, an input unit and a display unit may be provided to input a syntax tree from the input unit and output a semantic structure to the display unit. It is also possible to adopt a configuration in which the process is displayed on the display unit so that the flow of the semantic analysis can be visually understood. Further, a configuration in which a semantic structure is stored in the storage device may be adopted.

FIG. 5 shows the configuration of the deep case frame extraction unit of the present invention. The deep case frame extraction unit 250 includes a surface case information acquisition unit 251, a syntax tree node acquisition unit 1 (252), a constraint condition matching unit 1 (253), and a deep case frame generation unit 1.
(254), surface case destination information acquisition unit 255, syntax tree node acquisition unit 2 (256), constraint condition matching unit 2 (25)
7), composed of a deep case frame generation unit 2 (258).

The surface case information acquisition unit 251 inputs the deep case determined by the extraction target deep case determination unit 240, the state determined by the state determination unit 230, and other prerequisites such as the presence or absence of a parse tree node. As the search key of, the surface case information including the set of the surface case and the constraint corresponding to the search key, which is described in the verb meaning information attached to the head verb node, is output.

The syntax tree node acquisition unit 1 (252) receives the surface case in the surface case information as input and outputs a syntax tree node corresponding to the surface case with reference to the position knowledge on the syntax tree. The constraint condition matching unit 1 (253) receives the syntax tree node and the constraint condition in the surface case information as input, and includes information including types, types, and semantic information of phrases and parts of speech attached to the syntax tree node. And outputs the matching result.

The deep case frame generation unit 1 (254) receives a syntax tree node whose processing result of the constraint condition matching unit 1 (253) is true, and the deep case frame corresponding to the syntax tree node has not been set. Then, a part of the semantic structure in the analysis target frame is output by linking a deep case relationship from the head verb frame to the deep case frame.

The surface case movement destination information acquisition unit 255 is configured to execute the processing when the surface case information cannot be acquired as a result of the processing of the surface case information acquisition unit 251 or the syntax tree node acquisition unit 1 (252). If a tree node cannot be obtained, or if a syntax tree node for which the result of the processing of the constraint condition matching unit 1 (253) is true cannot be obtained, the deep case, state, and the presence / absence of a syntax tree node are determined. Using the other prerequisites as the input search key, output the surface case move destination information, which is described in the verb semantic information associated with the main verb node, and is composed of the set of the surface case move destination and the constraint condition corresponding to the search key. I do.

The syntax tree node acquisition unit 2 (256) receives the surface case movement destination in the surface case movement destination information and refers to the syntax tree position knowledge to find a syntax tree node corresponding to the surface case movement destination. Output. The constraint condition matching unit 2 (257) receives the syntax tree node and the constraint condition in the surface case destination information, and inputs information including the type, type, and semantic information of the phrase and part of speech attached to the syntax tree node, Match the constraint,
The matching result is output.

The deep case frame generation unit 2 (258) receives a syntax tree node whose processing result of the constraint condition matching unit 2 (257) is true, and the deep case frame corresponding to the syntax tree node has not been set. Then, a part of the semantic structure in the analysis target frame is output by linking a deep case relationship from the head verb frame to the deep case frame.

The operation of the above configuration will be described below. FIG. 6 is a flowchart showing the operation of the semantic analysis method of the present invention. In the semantic analysis flow shown below,
A semantic frame corresponding to the node of the sentence (S) that is the top node of the input syntax tree is set, and the semantic frame is used as an input for semantic analysis.

Step 100) Analysis target section judging section 10
At 0, the node corresponding to the input frame is analyzed and stored in the analysis target category classification knowledge storage device 400 from the information on the type of the phrase or part of speech and the subclassification of the phrase attached to the analysis target. With reference to the analysis target classification knowledge, it is determined whether the analysis target is a verb phrase having a verb as a head, a non-verb phrase mainly having a noun as a head, or a part of speech. As a result of the analysis target classification determination, if the analysis target is a verb phrase, the process proceeds to step 200 for performing a verb phrase analysis, and if the analysis target is a non-verb phrase, the process proceeds to a step 300 for performing a non-verb phrase analysis. If it is a part of speech, the semantic analysis ends.

Step 200) A verb phrase analysis is performed. Step 210) The subject verb frame extraction unit 210 receives the analysis target as an input, and stores
The reference verb node to be analyzed is obtained by referring to the position knowledge on the parse tree stored in 00, a head verb frame corresponding to the head verb node is set, and a head verb frame is set from the analysis target frame corresponding to the analysis target. By linking the head relation to the verb frame, a part of the semantic structure in the analysis target frame is output.

Step 220) The subject head verb frame determining unit 220 receives the head verb frame as an input, and determines which head verb frame of the head verb frame to be analyzed in the subsequent processing. Output the head verb frame. Step 230) State determination unit 2
At 30, an analysis target and a subject verb node are input, and the state of the analysis target is determined by referring to information of a state attached to the node including the analysis target, and the state is output.

Step 240) Extraction target deep case determination unit 2
At 40, a head case to be extracted is determined from the head cases described in the verb meaning information attached to the head verb node, and the head case is output. Step 250) In the deep case frame extraction unit 250, the analysis target, the head verb node, the deep case, and the state are input, and a syntax tree node corresponding to the deep case is obtained.
If the deep case frame corresponding to the syntax tree node has not been set, it is set, and a part of the semantic structure in the analysis target frame is output by linking the deep case relation from the head verb frame to the deep case frame. Details will be described later.

Step 260) The deep case frame determining unit 260 receives the deep case frame newly set in the deep case frame extraction process (step 250) and analyzes which deep case frame among the deep case frames. Is performed, and the deep case frame is output. Step 270) The deep case frame analysis unit 270 receives the deep case frame as an input and recursively calls a semantic analysis to output the semantic structure in the deep case frame.

Deep Case Frame Analysis Process (Step 27)
When 0) is completed, the process of the deep case frame to be processed (step 260) is performed again, and as a result of the process of determining the deep case frame to be processed (step 260), if there is no deep case frame to be processed, The extraction target deep case determination process (step 240) is performed again. As a result of the extraction target deep case determination process (step 240), if there is no deep case to be extracted, the processing target main verb frame determination process is performed again. (Step 220), and if there is no subject verb frame to be processed as a result of the subject head verb frame determining process (Step 220), the semantic analysis including the verb phrase analysis process (Step 200) is performed. finish.

Next, the non-verb phrase class analysis process (step 30)
0) will be described. Step 310) The subject frame extraction unit 310 receives the analysis object as input, refers to the positional knowledge on the parse tree, finds the phrase or part of speech node to be the subject of analysis, sets a subject frame corresponding to the subject node, By linking a head relation from the analysis target frame corresponding to the analysis target to the head frame, a part of the semantic structure in the analysis target frame is output.

Step 320) The subject frame determining section 320 receives the subject frame as input, determines which of the subject frames is to be analyzed in the subsequent processing, and outputs the subject frame. Step 330) The subject frame analysis unit 330 receives the subject frame as input, and recursively invokes semantic analysis to output the semantic structure in the subject frame.

Step 340) Modifier phrase frame extractor 3
At 40, a qualification node that qualifies the head node is obtained by referring to the analysis target and the head node as input and referring to the position knowledge on the syntax tree, and if a qualification frame corresponding to the qualification node is not set, it is set. A part of the semantic structure in the analysis target frame is output by linking a modification relation from the head frame to the decoration phrase frame.

Step 350) The processing target modifier phrase frame determining unit 350 receives the modifier phrase frame newly set in the modifier phrase frame extracting step (step 340), and selects any of the modifier phrase frames among the modifier phrase frames. Determine whether to perform the analysis and output the modifier frame.
Step 360) The modifier frame analyzing unit 360 receives the modifier frame as input, and recursively invokes semantic analysis to output the semantic structure in the modifier frame.

The modifier frame analysis process (step 36)
When 0) is completed, the processing of the processing target modifier phrase frame determining step (step 350) is performed again, and as a result of the processing of the processing target modifier phrase frame determining step (step 350), if there is no modifier phrase frame to be processed, The processing of the subject head frame determination process (step 320) is performed again,
If there is no subject frame to be processed as a result of the processing of the subject frame determining process (step 320), the semantic analysis including the non-verb phrase analysis process (step 300) is ended.

Next, the operation of the deep case frame extraction process (step 250) in the deep case frame extraction unit 250 will be described. FIG. 7 is a flowchart of a deep case frame extraction process according to the present invention. Step 251) In the surface case information acquisition unit 251, the deep case determined in the extraction target deep case determination step (step 240), the state determined in the state determination step (step 230), the presence / absence of a syntax tree node, and the like. Using the other prerequisites as an input search key, surface case information composed of a set of a surface case and a constraint corresponding to the search key, which is described in the verb meaning information attached to the head verb node, is output. If the acquisition is successful, the process proceeds to the syntax tree node acquiring process 1 (step 252), and if the acquisition is unsuccessful, the process proceeds to the surface case destination information acquiring process (step 255).

Step 252) Syntax tree node acquisition section 1
At (252), the surface case in the surface case information is input, and the syntax tree node corresponding to the surface case is output with reference to the positional knowledge on the syntax tree. If the acquisition is successful, the process proceeds to the constraint condition matching process 1 (step 253). If the acquisition is unsuccessful, the process proceeds to the surface case destination information acquisition process (step 255).
Step 253) In the constraint condition matching unit 1 (253), the syntax tree node and the constraint condition in the surface case information are input, and the information including the type, type, and semantic information of the phrase and part of speech attached to the syntax tree node, It collates with the constraint condition and outputs the collation result. If there is a syntax tree node for which the processing result is true, the process proceeds to the deep case frame generation process 1 (step 254); otherwise, the process proceeds to the surface case movement destination information acquisition process (step 255).

Step 254) Deep case frame generator 1
In (254), constraint condition matching process 1 (step 2)
53) Input a parse tree node for which the result of the processing of true) is true, set if a deep case frame corresponding to the parse tree node is not set, and link a deep case relation from the head verb frame to the deep case frame. Outputs a part of the semantic structure in the analysis target frame. When step 254 is completed, the deep case frame extraction process (step 2)
50) is ended.

Step 255) The surface case movement destination information acquisition unit 255 uses the deep case, the state, and other prerequisites such as the presence or absence of a parse tree node as search keys for input, and enters the verb semantic information attached to the main verb node. It outputs surface case movement destination information that is described and includes a set of a case movement destination and a constraint condition corresponding to the search key. If the acquisition is successful, the process proceeds to the syntax tree node acquisition process 2 (step 256). If the acquisition is unsuccessful, the deep case frame extraction process (step 250) is terminated.

Step 256) Syntax tree node acquisition unit 2
In (256), the surface case destination in the surface case destination information is input, and the syntax tree node corresponding to the surface case destination is output with reference to the syntax tree position knowledge. If the acquisition is successful, the process proceeds to the constraint condition matching process 2 (step 257), and if the acquisition is unsuccessful, the deep case frame extraction process (step 250) is terminated.

Step 257) Constraint condition collating unit 2 (25)
In 7), the syntax tree node and the constraints in the surface case destination information are input, and information including the type, type, and semantic information of the phrase and part of speech attached to the syntax tree node is checked against the constraints. The matching result is output. If there is a syntax tree node for which the processing result is true, the process proceeds to the deep case frame generation process 2 (step 258), and if not, the deep case frame extraction process (step 250) is terminated.

Step 258) Deep case frame generator 2
In (258), the constraint matching process 2 (step 2)
57) Input a syntax tree node for which the result of the process of true is true, set if the deep case frame corresponding to the syntax tree node is not set, and link a deep case relationship from the head verb frame to the deep case frame. Outputs a part of the semantic structure in the analysis target frame. When step 258 is completed, the deep case frame extraction process (step 25)
0) is ended.

(Embodiment) Hereinafter, the first embodiment of the present invention will be described with reference to the drawings.
An example will be described. The operation of the semantic analysis method will be described below using an example of a natural language sentence. FIG. 8 is a diagram for explaining the types and symbols of phrases and parts of speech according to the first embodiment of the present invention.

Here, as an example of a natural language sentence, "I want
FIG. 9 shows a syntax tree of a natural language sentence according to the first embodiment of the present invention. Different nodes that become symbols of the same phrase or part of speech in the syntax tree. Are distinguished from each other by subscripts.

Each syntax tree node is accompanied by information on a set of combinations of attributes such as phrases, parts of speech, types, which are subclassifications of phrases, semantic information, states, and their values. FIG. 10 shows a semantic structure output by the semantic analysis processing of the first embodiment of the present invention. The semantic structure is composed of a semantic frame corresponding to a certain node of the syntax tree, and a set of links of a head relation, a deep case relation, and a modification relation connecting the semantic frames. Each frame line enclosed by a square in FIG. 10 represents a meaning frame, and arrows connecting the meaning frames are links. The meaning frame corresponding to the node X having the syntax tree in FIG.
It is represented by "X ^* ".

A link whose subscript is a head is a link of a head relation, and a link having a deep case name such as a nominative case, a target case, a place case is a link of a deep case relation and is a modifier. The link in question is a link having a qualifying relationship. FIG. 11 is a diagram visually illustrating a semantic structure output by the semantic analysis processing when the syntax tree according to the first embodiment of the present invention is input. The meaning frame which is the head of a certain meaning frame X ^* is represented by a dark frame line, is placed inside X ^* , and the head relation link is omitted.

FIG. 12 is an example of the analysis target category classification knowledge stored in the analysis target category classification knowledge storage device according to the first embodiment of the present invention. For each type of analysis target (combination of a corresponding phrase and a type that is a fine classification thereof), it indicates whether the analysis target is classified into a verb phrase or a non-verb phrase. FIG. 13 shows an example of syntax tree position knowledge stored in the syntax tree position knowledge storage device according to the first embodiment of this invention. For each type of analysis target, the head of the analysis target, the surface case for the head, and the position of the qualifier that modifies the head on the syntax tree are defined.

FIG. 9A shows each position when the analysis target is S and the type is VP immediately below. FIG. 6B shows each position when the analysis target is NP and the type is an infinitive phrase. FIG. 3C shows each position when the analysis target is NP and the type is general or pronoun. FIG. 14 shows an example of the value of the semantic information of the verb node of the syntax tree according to the first embodiment of this invention.
One verb semantic information is a set of a combination of a deep case, a constraint condition defining an attribute of a phrase corresponding to the deep case, and a surface case for each state corresponding to the deep case. The type of deep case to be analyzed is described in the verb semantic information.

FIG. 9A shows a verb node of the syntax tree shown in FIG.
This is semantic information of V5 (want). Phrases whose deep case is the nominative case are noun phrases and semantic information is human, and the surface case is the nominative case when the subject to be analyzed whose subject is the verb is active is the nominative case. by ". FIG. 9B shows the semantic information of the verb node V10 (stay) in the syntax tree shown in FIG.

The operation of semantic analysis when the syntax tree shown in FIG. 9 is input will be described below. The operation will be described with reference to the flowchart shown in FIG. 6, but in order to make the description easier to understand, only the operation of the main part in the flowchart will be described in the first embodiment. Therefore, some of the processes that require case division are omitted, and in FIG.
Assume that there are only 254. A second embodiment will be described in more detail.

A semantic frame S1 * corresponding to the top node S1 of the syntax tree is set, and S1 * is used as an input for semantic analysis. First, from the information attached to the analysis target S1, the type of the analysis target S1 has a VP immediately below. Therefore, it is determined that S1 is a verb phrase based on the analysis target classification knowledge in FIG. 12 (step 100—analysis target classification). Judgment process).

According to FIG. 13A, the head verb node of the analysis target S1 is V5 (want) in FIG. A semantic frame V5 * corresponding to V5 is set, and a head relation link is established from "S1 *" to "V5 *" (step 200-verb phrase analysis process, step 210-head verb frame extraction process).
From the information attached to the analysis target S1, the state of S1 is determined to be active (step 230-state determination process).

From the information attached to the head verb node V5, V5
Is the content shown in FIG. FIG.
From the deep case described in 4 (A), the deep case to be extracted is determined as the main case (step 240-extraction target deep case determination process). Since the state of the analysis target S1 is active, FIG. 14 (A)
Accordingly, the surface case becomes the main case (step 250-deep case frame extraction process, step 251-surface case information acquisition process).

From FIG. 13A, the nominative case as the surface case in the analysis target S1 is “NP2” in FIG. 9 (step 25).
2-Syntax tree node acquisition process). From FIG. 14A, the constraint of the nominative case as the deep case is “noun phrase” and the semantic information is obtained as “person”. From the information attached to "NP2", "NP2" is a noun phrase and the semantic information is "person".
The collation result is “true” (step 253-constraint condition collation process).

The meaning frame “NP2 *” corresponding to “NP2”
Is set, and a link of the nominative relation is established from "V5 *" to "NP2 *" (step 254-deep case frame generation process).
The semantic analysis is recursively called using "NP2 *" as an input (step 270-deep case frame analysis process). Analysis target "NP
Since the type of the analysis target "NP2" is a pronoun from the information attached to "2", "NP2" is obtained from the analysis target classification classification knowledge in FIG.
2] are determined to be non-verb phrases (step 100-
Analysis target classification determination process).

From FIG. 13C, the head node of the analysis target "NP2" is "PN3 (I)" in FIG. A meaning frame "PN3 *" corresponding to "PN3" is set, and a head relation link is provided from "NP2 *" to "PN3 *" (step 300-).
Non-verb phrase analysis process, step 310-subject frame extraction process). Since "PN3" is a part of speech, nothing is performed (step 330-head frame analysis process).

From FIG. 13C, it can be seen that there is no modifier in FIG. 9 that modifies the head node "PN3" of the analysis target "NP2" (step 340-modifier phrase extraction process). As a result of the decoration phrase frame extraction process (step 340), it is determined that the decoration phrase frame has not been set, and the semantic analysis for the analysis target frame “NP2 *” is terminated (step 350—processing target decoration phrase frame determination process).

An extraction target is selected from the deep cases that have never been extracted in FIG. 14A, which are the semantic information of the head verb node “V5”, and the deep case to be extracted is selected as the target case. (Step 240-extraction target deep case determination process). Since the state of the analysis target “S1” is active, FIG.
From (A), the superficial case becomes the objective case (step 250-
Deep case frame extraction process, step 251-surface case information acquisition process).

As shown in FIG. 13A, the target in the analysis target "S1" is "NP6" in FIG. 9 (step 252-syntax tree node acquisition process). From FIG. 14A, the target case constraint condition is that the noun phrase has semantic information “thing”.
From the information attached to "NP6", "NP6" is a noun phrase and the meaning information is "thing", so that the matching result is "true" (step 253-constraint condition matching process).

The meaning frame "NP6 *" corresponding to "NP6"
Is set, and a link of the target case relation is created from "V5 *" to "NP6 *" (step 254-deep case frame generation process).
The semantic analysis is recursively called using "NP6 *" as an input (step 270-deep case frame analysis process). Analysis target "NP
Since the type of the analysis target "NP6" is an infinitive phrase based on the information attached to "6", from the analysis target classification classification knowledge in FIG.
It is determined that "NP6" is a verb phrase (step 100).
-Analysis target classification determination process).

From FIG. 13B, the head verb node of the analysis target "NP6" is "V10 (stay)" in FIG. 『V1
Set the meaning frame “V10 *” corresponding to “
A link of subject relations is provided from "*" to "V10 *" (step 200-verb phrase analysis process, step 210-head verb frame extraction process). The state of “NP6” is determined to be active based on the information accompanying the analysis target “NP6” (step 230—state determination process).

From the information attached to the head verb node V10,
The semantic information of V10 is the content shown in FIG.
The deep case to be extracted is determined to be the main case from the deep case described in FIG. 14B (step 240-extraction target deep case determination process). Since the state of the analysis target "NP6" is active, the superficial case becomes the nominative case from FIG. 14B (step 250).
-Deep case frame extraction process, step 251-surface case information acquisition process).

FIG. 13B shows that there is no nominative case as a surface case in the analysis target "NP6" (step 25).
2-Syntax tree node acquisition process). As a result of the syntax tree node acquisition process (step 252), it is found that the head verb node V10 does not have a nominative case as a deep case, so nothing is performed (step 253-constraint condition matching process).

Syntax Tree Node Acquisition Process (Step 252)
As a result, it is found that there is no nominative case as the deep case of the head verb node V10, so that nothing is done (step 254-deep case frame generation process). As a result of the deep case frame extraction process (step 250), there is no nominative case of the head verb node V10 as a deep case, so nothing is performed (step 270-).
Deep case frame analysis process).

FIG. 1 showing the semantic information of the head verb node V10
An extraction target is selected from the deep cases described in 4 (B) which have never been an extraction target, and the deep case to be extracted is determined as a location case (step 240-extraction target deep case determination process). Since the state of the analysis target “NP6” is active,
From (B), the surface case is the object of the preposition "at" (step 250-deep case frame extraction process, step 251-surface case information acquisition process).

The position of the object of the preposition designated in FIG. 13B and whose preposition is "at" is searched in "NP6" in FIG. 9 to obtain "NP13" (step 252).
-Syntax tree node acquisition process). From FIG. 14 (B), the place condition constraint condition is that the semantic information is occupation in a noun phrase. From the information attached to "NP13", since "NP13" is a noun phrase and the semantic information is occupation, the collation result is "true" (step 253-constraint condition collation process). "NP13"
Set the meaning frame "NP13 *" corresponding to "V10 *"
(Step 254-deep case frame generation process).

The semantic analysis is recursively called using "NP13 *" as an input (step 270-deep case frame analysis process). Since the type of the analysis target "NP13" is generally based on the information attached to the analysis target "NP13", it is determined that "NP13" is a non-verb phrase based on the analysis target classification knowledge in FIG. Object classification determination process).

From FIG. 13C, the head node of the analysis target “NP13” is “NP14” in FIG. A semantic frame "NP14 *" corresponding to "NP14" is set, and a head relation link is provided from "NP13 *" to "NP14 *" (step 300).
-Non-verb phrase analysis process, step 310-subject frame extraction process). The semantic analysis is recursively called using "NP14 *" as an input (step 330--head frame analysis process).

Since the type of the analysis target "NP14" is generally based on the information attached to the analysis target "NP14", it is determined that "NP14" is a non-verb phrase based on the analysis target classification classification knowledge of FIG. Step 100-analysis target classification determination process). 13C, the head node of the analysis target “NP14” is “NP16” in FIG. A semantic frame "NP16 *" corresponding to "NP16" is set, and a head relation link is established from "NP14 *" to "NP16 *" (Step 300-Non-verb phrase analysis process, Step 310-Head frame extraction process) ).

The semantic analysis is recursively called using "NP16 *" as an input (step 330--head frame analysis process).
From the information attached to the analysis target "NP16", the analysis target "NP1
Since the type of “6” is “general”, it is determined that “NP16” is a non-verb phrase class from the analysis target classification knowledge of FIG. 12 (step 100—analysis target classification determination process).

From FIG. 13C, the head node of the analysis target "NP16" is "N17" (hotel) in FIG. 『N1
7) and set the meaning frame "N17 *" to "NP1
6 * ”to“ N17 * ”is linked with a head relation (step 300-non-verb phrase analysis process, step 310-head frame extraction process). Since "N17" is a part of speech, nothing is performed (step 330-head frame analysis process).

From FIG. 13C, it can be seen that there is no modifier phrase for modifying the head node "N17" of the analysis object "NP16" in FIG. 9 (step 340--modifier phrase frame extraction process). As a result of the modification phrase frame extraction process (step 340), it is determined that the modification phrase frame has not been set, and the semantic analysis for the analysis target frame "NP16 *" is terminated (step 350-processing target modification phrase frame determination process).

From FIG. 13C, it can be seen that there is no modifier phrase for modifying the head node "NP16" of the analysis target "NP14" in FIG. 9 (step 340--modifier phrase frame extraction process). As a result of the decoration phrase frame extraction process (step 340), it is determined that no decoration phrase frame has been set, and the semantic analysis for the analysis target frame "NP14 *" is terminated (step 350-processing target decoration phrase frame determination process).

FIG. 13 (C) shows that the main data of the analysis target “NP13”
The modifier to modify the word node “NP14” is “NP20” in FIG.
Becomes The meaning frame "NP20 *" corresponding to "NP20"
Set and modify link from "NP14 *" to "NP20 *"
(With) (Step 340-Modifier phrase frame extraction)
process). Modifier phrase frame extraction process (step 340)
As a result, the modifier phrase frame “NP20 ^*Is set
(Step 350—Determining the processing target modifier frame)
About).

The semantic analysis is recursively called with "NP20 *" as an input (step 360--modifier phrase frame analysis process). Since the type of the analysis target “NP20” is “general” from the information attached to the analysis target “NP20”, it is determined that “NP20” is a non-verb phrase from the analysis target classification knowledge in FIG. 12 (step 100). -Analysis target classification determination process).

From FIG. 13C, the head node of the analysis target "NP20" is "N21" (steeple) in FIG. 『N21
], Set the meaning frame “N21 *” corresponding to “NP20 *
Is linked to "N21 *" (step 300-non-verb phrase analysis process, step 310-subject frame extraction process). Since "N21" is a part of speech, nothing is performed (step 330--head frame analysis process).

From FIG. 13C, it can be seen that there is no modifier phrase for modifying the head node "N21" of the analysis object "NP20" in FIG. 9 (step 340--modifier phrase frame extraction process). As a result of the decoration phrase frame extraction process (step 340), it is determined that the decoration phrase frame has not been set, and the semantic analysis for the analysis target frame "NP20 *" is terminated (step 350-processing target decoration phrase frame determination process).

FIG. 13 (C) shows that the main data of the analysis target “NP13”
The modifier to modify the word node “NP14” is “NP24” in FIG.
Becomes NP24 * meaning frame corresponding to NP24
Set and modify links from "NP14 *" to "NP24 *"
(In) (step 340—modifier phrase frame extraction
About). Result of Modifier Phrase Frame Extraction Process (Step 340)
As a result, the modifier frame “NP24 ^*Is set.
(Step 350-Processing target modifier frame determination
About).

The semantic analysis processing is recursively called using "NP24 *" as an input (step 360--modifier phrase frame analysis process). Since the type of the analysis target “NP24” is “general” from the information attached to the analysis target “NP24”, it is determined that “NP24” is a non-verb phrase from the analysis target classification knowledge of FIG. 12 (step 100-). Analysis target classification determination process).

As shown in FIG. 13C, the head node of the analysis target "NP24" is "N25" (Tokyo) in FIG. 『N25
Set the meaning frame "N25 *" corresponding to "
) To "N25 *" (step 300-non-verb phrase analysis process, step 310-subject frame extraction process). Since "N25" is a part of speech, nothing is performed (step 330--head frame analysis process). FIG.
From FIG. 9 (C), it can be seen that there is no modifier in FIG.
40-Modifier phrase frame extraction process).

Modifier phrase frame extraction process (step 34)
As a result of 0), it is determined that the qualifying phrase frame has not been set, and the semantic analysis for the analysis target frame "NP24 *" is terminated (step 350-processing target qualifying phrase frame determination process). From FIG. 13C, it can be seen that there are no more modifier phrases in FIG. 9 to modify the head node "NP14" of the analysis target "NP13" (step 340-modifier phrase frame extraction process).

Modifier phrase frame extraction process (step 34)
As a result of 0), it is determined that the modifier frame has not been set, and the semantic analysis for the analysis target frame "NP13 *" is terminated (step 350-processing target modifier phrase frame determination process). Figure 1 which is the semantic information of the head verb node "V10"
Since all the deep cases described in FIG. 4 (B) are to be extracted, it is determined that there is no deep case to be extracted, and the semantic analysis for the analysis target frame “NP6 *” is completed (step 2).
40—Determination of deep case to be extracted).

Since all the deep cases described in FIG. 14A, which are the semantic information of the head verb node "V5", are to be extracted, it is determined that there is no deep case to be extracted, and the analysis target frame " The semantic analysis for "S1 *" is terminated (step 240--the process of determining the deep case to be extracted). As described above, as an output result when the syntax tree of FIG. 9 is input, FIG.
Alternatively, the semantic structure of FIG. 11 is obtained.

Next, a second embodiment of the present invention will be described. FIG. 15 is a diagram for explaining types and symbols of phrases and parts of speech according to the second embodiment of the present invention. Here, as an example of a natural language sentence, "what fish and what shellfish do t
hey buy and sell at the market with roofs in Tokyo
? ". The notational meaning in each figure is
This is the same as the corresponding figure in the embodiment.

FIG. 16 shows a syntax tree of a natural language sentence according to the second embodiment of the present invention. FIG. 17 shows a semantic structure output by the semantic analysis processing of the second embodiment of the present invention. FIG.
FIG. 8 is a diagram visually illustrating the semantic structure output by the semantic analysis processing when the syntax tree according to the second embodiment of the present invention is input.

FIG. 19 shows an example of analysis target category classification knowledge stored in the analysis target category classification knowledge storage device according to the second embodiment of the present invention. FIG. 20 shows an example of knowledge regarding the normal position of the surface case, of the position knowledge on the syntax tree stored in the position knowledge storage on the syntax tree according to the second embodiment of the present invention. In the figures (A), (B), (C) and (D), the analysis target is S
Indicates the position of the subject node, the subject case, the subject case, and the object of the preposition when the type is a question sentence question sentence. A plurality of verbs, which are heads, are connected in parallel by a conjunction.

FIG. 17E shows an example of the position of a head node when the analysis target is NP and the type is a question phrase. (F) and (G) show another example of the position of the head node and the position of the modifier when the analysis target is NP and the type is a question phrase. In the figures (H) and (I), the analysis target is NP,
When the type is general, the position of the head node and the position of the modifier are shown.

FIG. 21 shows a syntax tree in which a node corresponding to a surface case corresponding to a subject is an input from syntax tree location knowledge stored in the syntax tree location knowledge storage device according to the second embodiment of the present invention. The following is an example of knowledge defining the position of the destination of the surface case when it is not above. FIG. 7A shows an example of the position of the target case after the movement. FIG. 7B shows another example of the position of the target case after the movement. "Beer to drink" is an example of an NP that corresponds to this example. NP "beer" is ADJP
The purpose of the verb "drink" in "to drink" has been moved.

FIG. 22 shows an example of the value of the meaning information of the verb node of the syntax tree according to the second embodiment of the present invention. One verb semantic information includes a search key (a set of other preconditions for a deep case, a state, the presence or absence of a parse tree node, and the like) and surface case information corresponding to the search key (a surface case, a constraint that defines attributes of phrases). A set of conditions), a search key (a set of other prerequisites for a deep case, a state, the presence or absence of a parse tree node, etc.) and surface case destination information (surface case move destination, phrase) corresponding to the search key (A set of constraints) defining the attributes of
In FIG. 22, the verb node V18 of the syntax tree node of FIG.
(Buy) and V20 (sell) are shown. Phrases whose deep case is the nominative case are noun phrases and semantic information is human,
The superficial case is the nominative case when the state of the analysis subject whose verb is the head is active, and the preposition "by"
It indicates that the object is Also, if the node corresponding to the nominative case of the superficial case is not on the input parse tree at the time of active voice, the superficial case will be the position of the nominative case after movement, and the nominative case of the deep case will be If not obtained, it indicates that the surface case is the position after the movement of the object of the preposition "by".

Hereinafter, the semantic analysis operation when the syntax tree of FIG. 16 is input will be described in detail with reference to the flowchart of FIG. A semantic frame S1 * corresponding to the top node S1 of the syntax tree is set, and S1 * is used as an input for semantic analysis. First, from the information attached to the analysis target S1, the type of the analysis target S1 is a question verb question sentence. Therefore, it is determined that S1 is a verb phrase based on the analysis target classification classification knowledge of FIG. Judgment process).

According to FIG. 20A, the head verb nodes of the analysis target S1 are V18 (buy) and V20 (sell) in FIG. A semantic frame V18 * corresponding to V18 is set, and a head relation link is created from S1 * to V18 *. In addition, a semantic frame V20 * corresponding to V20 is set, and a head relation link is provided from S1 * to V20 * (step 200-verb phrase analysis process,
Step 210-head verb frame extraction process).

Among the head verb frames set in step 210, the head verb frame to be processed thereafter is determined to be V18 * (step 220-process for determining the head verb frame to be processed). VP that is an upper node of the head verb node V18
From the information attached to 17, it is determined that the state of the analysis target S1 is active (step 230-state determination process). From the information attached to the head verb node V18, the semantic information of V18 has the contents shown in FIG. From the deep cases described in FIG. 22, the deep case to be extracted is determined as the target case (step 2).
40—Determination of deep case to be extracted).

As shown in FIG. 22, surface case information (surface case: object case, constraint condition: noun phrase and semantic information: deep case: object case, state: active, precondition: none) is used as a search key.
(Step 250-deep case frame extraction process, step 251-surface case information obtaining process). FIG.
From (B), the target case as the surface case in the analysis target S1 is:
It can be seen that it is not on the syntax tree of FIG. 16 (step 252).
-Syntax tree node acquisition process 1). Next, referring to FIG. 22, the deep case: target case, state: active, prerequisite: no case node is used as a search key, and the surface case movement destination information (surface case:
Objective after movement, constraints: Noun phrase and semantic information: thing)
(Step 255-surface layer destination information acquisition process).

As shown in FIG. 21A, the target case after the movement is NP2 in FIG. 16 (step 256-syntax tree node acquisition process 2). The constraint condition in the surface case destination information is a noun phrase and semantic information: a thing, and since the NP2 is a noun phrase and a semantic information is a thing than the information attached to NP2, the collation result is true (step 257-). Constraint condition matching process 2). Since the semantic frame corresponding to NP2 has not been set, the semantic frame NP2 * is set, and a link of the target case relation is created from V18 * to NP2 * (step 258-deep case frame generation process 2).

Of the target case frames newly set in step 250, the target case frame to be processed is determined to be NP2 * (step 260-process for determining deep case frame to be processed). Then, the semantic analysis is recursively called using NP2 * as an input (step 270-deep case frame analysis process).
Since the type of the analysis target NP2 is an interrogative phrase from the information attached to the analysis target NP2, it is determined that NP2 is a non-verb phrase from the analysis target classification classification knowledge of FIG. 19 (step 10).
0-analysis target classification determination process).

According to FIG. 20 (E), since the head nodes of the analysis target NP2 are NP3 and NP8 in FIG. 16, a semantic frame NP3 * corresponding to NP3 is set, and the head relation link from NP2 * to NP3 * is set. Put up. In addition, the meaning frame corresponding to NP8
NP8 * is set, and a head relation link is provided from NP2 * to NP8 * (step 300-non-verb phrase analysis process, step 31)
0-head frame extraction process).

Of the head frames set in step 310, the head frame to be processed thereafter is determined to be NP3 * (step 320-process for determining head frame to be processed).
Then, the semantic analysis is recursively called using NP3 * as an input (step 330-head frame analysis process). Analysis object
Since the type of the analysis target NP3 is an interrogative phrase based on the information attached to the NP3, it is determined that the NP3 is a non-verb phrase based on the analysis target classification knowledge in FIG. 19 (step 100-analysis target classification determination process).

According to FIG. 20 (F), since the head node of the analysis target NP3 is NP5 in FIG. 16, a semantic frame NP5 * corresponding to NP5 is set, and a head relation link is provided from NP3 * to NP5 *. (Step 300-Non-verb phrase analysis process, Step 310-Subject frame extraction process). Step 31
Of the head frames set at 0, the head frame to be processed thereafter is determined as NP5 * (step 320—processing head frame determination process), and the semantic analysis is recursively called using NP5 * as input (step 330- Head frame analysis process). From the information attached to the analysis target NP5, the analysis target NP
Since NP5 is a general type, it is determined that NP5 is a non-verb phrase based on the analysis target classification knowledge in FIG. 19 (step 1).
00-analysis target classification determination process).

Since the head node of the analysis target NP5 is N6 in FIG. 16 based on the knowledge on the syntax tree, a semantic frame N6 * corresponding to N6 is set, and a head relation link is provided from NP5 * to N6 *. (Step 300-Non-verb phrase analysis process, Step 310-Subject frame extraction process). Step 31
Of the head frames set at 0, the head frame to be processed thereafter is determined as N6 * (step 320-head frame determination process for processing), and the semantic analysis is recursively called using N6 * as an input (step 330- Head frame analysis process). Based on the information attached to the analysis target N6, since the analysis target N6 is a part of speech, the semantic analysis for the analysis target frame N6 * is terminated (step 100—analysis target classification determination process).

Based on the knowledge of the position on the parse tree, it is found that the modifier for modifying the head node N6 of the analysis target NP5 is not on the parse tree of FIG. 16 (step 340—modifier phrase frame extraction process). Since there is no modifier phrase frame set in, it is determined that there is no modifier phrase frame to be processed (step 350-processing target modifier phrase frame determination process). Since all of the head frames set in step 310 have been processed, it is determined that there is no head frame to be processed, and the semantic analysis for the analysis target frame NP5 * is terminated (step 320—processing head head frame determination process).

According to FIG. 20 (G), the modifier phrase for modifying the head node NP5 of the analysis target NP3 is PN4 in FIG.
Since the semantic frame corresponding to PN4 has not been set, a semantic frame PN4 * is set, and a modification-related link is established from NP5 * to PN4 * (step 340—modifier phrase frame extraction process). Of the modifier frames newly set in step 340,
The qualifier frame to be processed is determined to be PN4 * (step 3
50-Process of Determining Target Modifier Phrase Frame), Semantic Analysis Recursively Called Using PN4 * as Input (Step 360-
Modifier frame analysis process). From the information attached to the analysis target PN4, since the analysis target PN4 is a part of speech, the analysis target frame
The semantic analysis for PN4 * is completed (step 100—analysis target category determination process). Then, since all of the modifier frames newly set in step 340 have been processed,
It is determined that there is no modifier frame to be processed (step 350-process of determining a modifier frame).

Since all of the head frames set in step 310 have been processed, it is determined that there is no head frame to be processed, and the semantic analysis for the analysis target frame NP3 * is terminated (step 320—determination of head head frame to be processed). process). According to FIG. 20, the head node NP3 of the analysis target NP2
It is found that there is no qualifying phrase in the syntax tree of FIG. 16 (step 340-qualifying phrase frame extracting process), and since there is no qualifying phrase frame newly set in step 340, the qualifying phrase frame to be processed is It is determined that there is no (step 350-process of modifying target object phrase frame).

Subsequently, among the subject frames set in step 310, a subject to be processed is selected from those which have never been processed, and the subject frames to be processed thereafter are set to NP8 *.
(Step 320-subject frame determination process) and recursively invoke semantic analysis with NP8 * as input (step 330-subject frame analysis process). Analysis object
Since the type of the analysis target NP8 is an interrogative phrase based on the information attached to the NP8, it is determined that the NP8 is a non-verb phrase based on the analysis target classification knowledge in FIG. 19 (step 100-analysis target classification determination process). FIG. 20 (F) shows that the analysis target NP8
Is set to NP10 in FIG. 16, a semantic frame NP10 * corresponding to NP10 is set, and a head relation link is established from NP8 * to NP10 * (step 300-non-verb phrase analysis process, step 310 -Head frame extraction process). Of the head frames set in step 310, the head frame to be processed thereafter is determined as NP10 * (step 32).
0-process to determine subject head frame), recursively invoke semantic analysis with NP10 * as input (step 330-head frame analysis process). Since the type of the analysis target NP10 is more general than the information attached to the analysis target NP10, it is determined that the NP10 is a non-verb phrase based on the analysis target classification classification knowledge of FIG. 19 (step 100-analysis target classification determination process).

The head node of the analysis target NP10 is N11 in FIG. 16 based on the knowledge on the position on the syntax tree. A semantic frame N11 * corresponding to N11 is set, and a head relation link is provided from NP10 * to N11 * (step 300-non-verb phrase analysis process,
Step 310-head frame extraction process). Step 3
Of the head frames set in step 10, the head frame to be processed thereafter is determined to be N11 * (step 320—processing head frame determination process), and the semantic analysis is recursively called using N11 * as an input (step 330-). Head frame analysis process). From the information attached to the analysis target N11,
Since N11 is a part of speech, the semantic analysis for the analysis target frame N11 * is terminated (step 100—analysis target category determination process).

From the knowledge of the position on the syntax tree, it can be seen that the modifier for modifying the head node N11 of the analysis target NP10 is not on the syntax tree of FIG. 16 (step 340—modifier phrase frame extraction process). Since there is no modifier phrase frame newly set in step 340, it is determined that there is no modifier phrase frame to be processed (step 350-processing target modifier phrase frame determination process). Then, since all of the head frames set in step 310 have been processed, it is determined that there is no head frame to be processed, and the semantic analysis for the analysis target frame NP10 * is terminated (step 320—processing head head frame determination process). ).

According to FIG. 20 (G), the modifier for modifying the head node NP10 of the analysis target NP8 is PN9 in FIG.
Since the semantic frame corresponding to PN9 has not been set, a semantic frame PN9 * is set, and a modification-related link is established from NP10 * to PN9 * (step 340—modification phrase frame extraction process).
Of the modifier frames newly set in step 340, the modifier frame to be processed is determined to be PN9 * (step 350—processing target modifier frame determination process), and the semantic analysis is recursively called using PN9 * as an input. (Step 36
0-modifier frame analysis process). From the information attached to the analysis target PN9, since the analysis target PN9 is a part of speech, the semantic analysis for the analysis target frame PN9 * is completed (step 100).
-Analysis target classification determination process). Since all of the modifier frames newly set in step 340 have been processed, it is determined that there are no modifier frames to be processed (step 35).
0—Processing target modifier frame determination process). Step 31
Since all of the subject frames set at 0 have been processed,
It is determined that there is no subject frame to be processed, and the semantic analysis for the analysis target frame NP8 * is terminated (step 320).
-Determining subject head frame to be processed).

As shown in FIG. 20, the head node of the analysis target NP2
It can be seen that the modifier that modifies NP8 is not on the syntax tree of FIG. 16 (step 340—modifier phrase frame extraction process). Since there is no modifier phrase frame newly set in step 340, it is determined that there is no modifier phrase frame to be processed (step 350-processing target modifier phrase frame determination process). Since all of the head frames set in step 310 have been processed, it is determined that there is no head frame to be processed, and the semantic analysis for the analysis target frame NP2 * is terminated (step 320-process for determining the head frame to be processed).

Since all the target case frames newly set in step 250 have been processed, it is determined that there is no deep case frame to be processed (step 260-process for determining a deep case frame to be processed). Next, an extraction target is selected from the deep cases described in FIG. 22, which are the semantic information of the head verb node V18, which have not been extracted yet, and the deep case to be extracted is determined as a location case (step 240-Extraction target deep case determination process).

As shown in FIG. 22, surface case information (surface case: at object, constraint condition: noun phrase and semantic information: occupation) is used as a search key with deep case: place case, form: active, precondition: none. (Step 250-deep case frame extraction process, step 251-surface case information obtaining process).
From FIG. 20 (D), the object of at in the analysis target S1 is as shown in FIG.
6 (step 252-syntax tree node acquisition process 1). The constraint in the surface case information is noun phrase and semantic information: occupation, and NP23 is obtained from the information attached to NP23.
Since the meaning information is occupation in the noun phrase, the collation result is true (step 253-constraint condition collation process 1).

Since the semantic frame corresponding to NP23 has not been set, the semantic frame NP23 * is set, and a link of the location relation is set from V18 * to NP23 * (step 254-deep case frame generation process 1). Of the newly set locative frames, the locative frame to be processed is NP23
* (Step 260-deep case frame determination process to be processed), and semantic analysis is recursively called using NP23 * as an input (step 270-deep case frame analysis process).

Since the type of the analysis target NP23 is generally based on the information attached to the analysis target NP23, it is determined that the NP23 is a non-verb phrase based on the analysis target classification knowledge shown in FIG. process). According to FIG. 20H, the head node of the analysis target NP23 is NP24 in FIG. Set the meaning frame NP24 * corresponding to NP24 and set NP23
Create a head link from * to NP24 * (Step 30)
0-non-verb phrase analysis process, step 310-subject frame extraction process).

Of the head frames set in step 310, the head frame to be processed thereafter is determined to be NP24 * (step 320—processing head frame determination process).
The semantic analysis is recursively called using NP24 * as an input (step 330-head frame analysis process). Subsequently, since the type of the analysis target NP24 is general based on the information attached to the analysis target NP24, it is determined that the NP24 is a non-verb phrase based on the analysis target classification classification knowledge of FIG. process). According to FIG. 20H, the head node of the analysis target NP24 is NP26 in FIG. A semantic frame NP26 * corresponding to NP26 is set, and a head relation link is established from NP24 * to NP26 * (step 300-non-verb phrase analysis process, step 310-head frame extraction process).

[0143] Of the head frames set in step 310, the head frame to be processed thereafter is determined as NP26 * (step 320-process for determining the head frame to be processed).
The semantic analysis is recursively called using NP26 * as an input (step 330-head frame analysis process). Since the type of the analysis target NP26 is more general than the information attached to the analysis target NP26,
Based on the analysis target category classification knowledge of FIG. 19, it is determined that NP26 is a non-verb phrase (step 100-analysis target category determination process).

The head node of the analysis target NP26 becomes N27 in FIG. 16 based on the position knowledge on the syntax tree. A semantic frame N27 * corresponding to N27 is set, and a head relation link is provided from NP26 * to N27 * (step 300-non-verb phrase analysis process,
Step 310-head frame extraction process). Step 3
Of the head frames set in step 10, the head frame to be processed thereafter is determined to be N27 * (step 320—process head frame determination process), and semantic analysis is recursively called using N27 * as an input (step 330-). Head frame analysis process). Then, from the information attached to the analysis target N27,
Since the analysis target N27 is a part of speech, the semantic analysis for the analysis target frame N27 * is terminated (step 100—analysis target classification determination process).

It can be understood from the syntax tree position knowledge that the modifier for modifying the head node N27 of the analysis target NP26 is not on the syntax tree of FIG. 16 (step 340—modifier phrase frame extraction process). Since there is no modifier phrase frame newly set in step 340, it is determined that there is no modifier phrase frame to be processed (step 350-processing target modifier phrase frame determination process).

Since all of the head frames set in step 310 have been processed, it is determined that there is no head frame to be processed, and the semantic analysis for the analysis target frame NP26 * is completed (step 320—determination of head frame to be processed). process). Next, according to FIG. 20 (I), the modifier to modify the head node NP26 of the analysis target NP24 is DET25 in FIG. Since the meaning frame corresponding to DET25 has not been set,
A semantic frame DET25 * is set, and a link relating to the modification is provided from NP26 * to DET25 * (step 340—modifier phrase frame extraction process). Of the modifier frames newly set in step 340, the modifier frame to be processed is determined to be DET25 * (step 350—processing target modifier frame determination process), and the semantic analysis is recursively called using DET25 * as an input. (Step 360-Modifier phrase frame analysis process). Based on the information attached to the analysis target DET 25, the analysis target DET 25 is a part of speech, so the semantic analysis for the analysis target frame DET 25 * is terminated (step 100—analysis target classification determination process).

Since all of the modifier frames newly set in step 340 have been processed, it is determined that there is no modifier frame to be processed (step 350-process of determining a modifier frame to be processed). Since all of the head frames set in step 310 have been processed, it is determined that there is no head frame to be processed, and the semantic analysis for the analysis target frame NP24 * is terminated (step 320—process for determining the head frame to be processed).

Next, according to FIG. 20 (I), the analysis target NP23
Modifiers that modify the head node NP24 of are NP30 and NP30 in FIG.
NP34. Since the semantic frame corresponding to NP30 has not been set, the semantic frame NP30 * is set, and a modification-related link is made from NP24 * to NP30 *. Also, since the meaning frame corresponding to NP34 has not been set, set the meaning frame NP34 *,
A link relating to the modification is provided from NP24 * to NP34 * (step 340—modifier phrase frame extraction process). Of the modifier frames newly set in step 340, the modifier frame to be processed is determined to be NP30 * (step 350—processing target modifier frame determination process), and the semantic analysis is recursively called using NP30 * as an input. (Step 360-Modifier phrase frame analysis process). Since the type of the analysis target NP30 is more general than the information attached to the analysis target NP30, it is determined that the NP30 is a non-verb phrase based on the analysis target classification classification knowledge of FIG. 19 (step 100-analysis target classification determination process). According to the position knowledge on the parse tree, the head node of the analysis target NP30 is N3 in FIG.
It becomes 1. Set a meaning frame N31 * corresponding to N31,
Create a head link from NP30 * to N31 * (Step 3)
00-non-verb phrase analysis process, step 310-subject frame extraction process).

Of the head frames set in step 310, the head frame to be processed thereafter is determined as N31 * (step 320—process head frame determination process).
The semantic analysis is recursively called using N31 * as an input (step 330-head frame analysis process). According to the information attached to the analysis target N31, since the analysis target N31 is a part of speech, the semantic analysis for the analysis target frame N31 * is completed (step 1).
00-analysis target classification determination process). Based on the knowledge of the position on the syntax tree, it can be understood that the qualifying phrase for modifying the head node N31 of the analysis target NP30 is not on the syntax tree of FIG. 16 (step 3).
40-Modifier phrase frame extraction process). Since there is no modifier phrase frame newly set in step 340, it is determined that there is no modifier phrase frame to be processed (step 350-processing target modifier phrase frame determination process).

Since all of the head frames set in step 310 have been processed, it is determined that there is no head frame to be processed, and the semantic analysis for the analysis target frame NP30 * is terminated (step 320—determination of head frame to be processed). process). Subsequently, among the modifier frames newly set in step 340, a processing target is selected from those which have never been processed, and the modifier phrase frame to be processed is set to NP.
34 * (Step 350-Modification phrase frame determination process to be processed), and semantic analysis is recursively called using NP34 * as an input (Step 360-Modification phrase frame analysis process). From the information attached to the analysis target NP34, the analysis target NP34
Since NP34 is a general type, it is determined that NP34 is a non-verb phrase class from the analysis target classification classification knowledge of FIG. 19 (step 10).
0-analysis target classification determination process).

The head node of the analysis target NP34 is N35 in FIG. 16 based on the position knowledge on the syntax tree. A semantic frame N35 * corresponding to N35 is set, and a head relation link is established from NP34 * to N35 * (step 300-non-verb phrase analysis process,
Step 310-head frame extraction process). Step 3
Of the head frames set in step 10, the head frame to be processed thereafter is determined to be N35 * (step 320-processing head frame determination process), and the semantic analysis is recursively called using N35 * as an input (step 330-). Head frame analysis process). From the information attached to the analysis target N35,
Since N35 is a part-of-speech, the semantic analysis for the analysis target frame N35 * is terminated (step 100—analysis target classification determination process).

It can be seen from the syntax tree position knowledge that the modifier for modifying the head node N35 of the analysis target NP34 is not on the syntax tree in FIG. 16 (step 340—modifier phrase frame extraction process). Since there is no modifier phrase frame newly set in step 340, it is determined that there is no modifier phrase frame to be processed (step 350-processing target modifier phrase frame determination process).

Since all of the head frames set in step 310 have been processed, it is determined that there is no head frame to be processed, and the semantic analysis for the analysis target frame NP34 * is terminated (step 320—determination of head frame to be processed). process). Then, since all of the modifier frames newly set in step 340 have been processed, it is determined that there are no modifier frames to be processed (step 350-processing target modifier frame determination process).

Since all of the head frames set in step 310 have been processed, it is determined that there is no head frame to be processed, and the semantic analysis for the analysis target frame NP23 * is completed (step 320—determination of head frame to be processed). process). Then, since all the place case frames newly set in step 250 have been processed, it is determined that there is no deep case frame to be processed (step 260-target deep case frame determination process).

Subsequently, an extraction target is selected from the deep cases described in FIG. 22, which are the semantic information of the head verb node V18, which have never been extracted, and the deep case to be extracted is determined as the nominative case. (Step 240-extraction target deep case determination process). From FIG. 22, it is found that the deep case: nominative case, the form: active, the precondition: none, and the search key is the surface case information (surface case:
Nominal case, constraint condition: noun phrase and semantic information: person) are acquired (step 250-deep case frame extraction process, step 251-surface case information acquisition process). From FIG. 20 (C), the nominative case as the surface case in the analysis target S1 is NP14 in FIG. 16 (step 252-syntax tree node acquisition process 1). The constraint condition in the surface case information is noun phrase and semantic information: person, and since the NP14 is a noun phrase and the semantic information is person than the information attached to NP14, the collation result is true (step 253-
Constraint condition matching process 1). And, since the meaning frame corresponding to NP14 has not been set, the meaning frame NP14 * is set,
Link the nominative relationship from V18 * to NP14 * (Step 2)
54—Deep case frame generation process 1).

Of the nominative frames newly set in step 250, the nominative case to be processed is determined to be NP14 * (step 260-process of determining deep frame to be processed), and semantic analysis is recursively called using NP14 * as an input. (Step 270-Deep Case Frame Analysis Process). Since the type of the analysis target NP14 is a pronoun from the information attached to the analysis target NP14, it is determined that the NP14 is a non-verb phrase based on the analysis target classification classification knowledge of FIG. 19 (step 100-analysis target classification determination process). Analysis target based on knowledge of position on parse tree
The head node of NP14 is PN15 in FIG. A semantic frame PN15 * corresponding to PN15 is set, and a head relation link is established from NP14 * to PN15 * (step 300-non-verb phrase analysis process, step 310-head frame extraction process).
Of the head frames set in step 310, the head frame to be processed thereafter is determined as PN15 * (step 3).
20—Process of Determining Subject Subject Frame), Recursively invoke semantic analysis with PN15 * as input (step 330-).
Head frame analysis process). From the information attached to the analysis target PN15, since the analysis target PN15 is a part of speech, the analysis target frame PN
The semantic analysis for 15 * is completed (step 100-analysis target category determination process).

From the knowledge of the position on the syntax tree, it can be understood that the modifier for modifying the head node PN15 of the analysis target NP14 is not on the syntax tree of FIG. 16 (step 340—modifier phrase frame extraction process). Since there is no modifier phrase frame newly set in step 340, it is determined that there is no modifier phrase frame to be processed (step 350-processing target modifier phrase frame determination process).

Since all of the head frames set in step 310 have been processed, it is determined that there is no head frame to be processed, and the semantic analysis for the analysis target frame NP14 * is terminated (step 320—determination of head frame to be processed). process). Since all of the nominative case frames newly set in step 250 have been processed, it is determined that there is no deep case frame to be processed (step 260-processing target deep case frame determination process).

Then, since all the deep cases described in FIG. 22, which are the semantic information of the head verb node V18, have been processed, it is determined that there is no deep case to be extracted (step 24).
0—Determination of deep case to be extracted). Then, step 21
Of the head verb frames set at 0, a processing target is selected from those that have never been processed, and the head verb frame to be processed thereafter is determined to be V20 * (step 22).
0-Processing subject head verb frame determination process). Then, the state of the analysis target S1 is determined to be active based on the information attached to VP17, which is an upper node of the head verb node V20 (step 230—state determination process). From the information attached to the head verb node V20, V20 is determined. Is the content shown in FIG. The deep case to be extracted is determined as the target case from the deep case described in FIG. 22 (step 240-extraction target deep case determination process). From FIG. 22, it is found that the deep case: target case, state: active, precondition: none, and the search key are the surface case information (surface case: object case, constraint condition: noun phrase and semantic information:
(Step 250-deep case frame extraction process, step 251-surface case information obtaining process).

From FIG. 20 (B), it can be seen that the object case as the surface case in the analysis target S1 is not on the syntax tree in FIG. 16 (step 252-syntax tree node acquisition process 1). FIG.
From 2, the deep case: target case, state: active, prerequisite: no case node is used as a search key, and the surface case destination information (surface case: moved case, constraint condition: noun phrase and semantic information) : Object) (step 255-surface layer destination information acquisition process). From FIG. 21A, the target case after the movement is NP2 in FIG. 16 (step 256—procedure tree node acquisition process 2). The constraint condition in the surface case destination information is a noun phrase and semantic information: thing, based on the information attached to NP2.
Since NP2 is also a noun phrase and has semantic information, the collation result is true (step 257-constraint condition collation process 2). NP2
Since the semantic frame NP2 * corresponding to is already set, a link of the target case is established from V20 * to NP2 * (step 258-).
Deep case frame generation process 2).

Since there is no target case frame newly set in step 250, it is determined that there is no deep case frame to be processed (step 260-process for determining deep case frame to be processed). Figure 2 which is the semantic information of the head verb node V20
An extraction target is selected from the deep cases described in No. 2 which have never been an extraction target, and the deep case to be extracted is determined as a location case (step 240-extraction target deep case determination process).
According to FIG. 22, deep case: place, form: active, precondition:
With the search key of “none”, surface case information (surface case: object of at, constraint condition: noun phrase and semantic information: occupation) is acquired (step 250-deep case frame extraction process, step 251-surface case information acquisition process). From FIG. 20 (D),
The object of at in the analysis target S1 is NP23 in FIG. 16 (step 252-syntax tree node acquisition process 1). The constraint condition in the surface case information is noun phrase and semantic information: occupation, and NP
Since the NP23 is a noun phrase and the semantic information is occupation from the information attached to 23, the collation result is true (step 253-constraint condition collation process 1). NP23 meaning frame NP23 *
Is already set, a link of the location relation is established from V20 * to NP23 * (step 254-deep case frame generation process 1).

Since there is no place case frame newly set in step 250, it is determined that there is no deep case frame to be processed (step 260-process for determining deep case frame to be processed). Figure 2 which is the semantic information of the head verb node V18
An extraction target is selected from the deep cases described in No. 2 which have never been an extraction target, and the deep case to be extracted is determined as the main case (step 240-extraction target deep case determination process).

FIG. 22 shows that the deep case: the nominative case and the state:
Active case, prerequisite: none, and using the search key as surface key information (surface case: nominative, constraint: noun phrase and semantic information:
(Step 250-deep case frame extraction process, step 251-surface case information obtaining process). FIG.
From (C), the nominative case as the surface case in the analysis target S1 is NP14 in FIG. 10 (step 252-syntax tree node acquisition process 1). The constraint condition in the surface case information is noun phrase and semantic information: person, and since the NP14 is a noun phrase and the semantic information is person than the information attached to NP14, the collation result is true (step 253-constraint condition). Collation process 1). Since the semantic frame NP14 * corresponding to NP14 has been set, a link of the nominative relation is established from V20 * to NP14 * (step 254-deep case frame generation process 1).

Since there is no main case frame newly set in step 250, it is determined that there is no deep case frame to be processed (step 260-process for determining deep case frame to be processed). FIG. 22 which is the semantic information of the head verb node V20
Are processed, it is determined that there is no deep case to be extracted (step 240-extraction case deep case determination process).

Since all of the head verb frames set in step 210 have been processed, the semantic analysis for the analysis target frame S1 * is terminated (step 220—process for determining the processing target head verb frame). As described above, the semantic structure of FIG. 17 or FIG. 18 is obtained as an output result when the syntax tree of FIG. 16 is input.

Further, the present invention constructs each component shown in FIG. 4 and FIG. 5 as a program, and can be connected to a computer used as a semantic analyzer, a floppy disk, a CD-ROM or the like. The present invention can be easily realized by storing it in a portable storage medium and installing it in a computer when implementing the present invention. In addition, the information search device shown in FIG. 3 can also construct each component as a program and perform the information search of the present invention on a computer.

FIG. 23 is a block diagram showing an example of a hardware configuration of the computer. The computer system includes a CPU 401 for executing processing, a memory 402 for storing programs and data, a memory 402 or a CP.
External storage device 403 for storing programs and data used in U401, display 40 for displaying data
4, a keyboard 405 for inputting data or instructions, and an interface 406 for communicating with another computer system or the like via a network.
The above program is executed by the CPU 401.

For example, when this computer is used as the information search device of the present invention, the interface 4
It is possible to input a natural language sentence from another computer or the like via 06, and output search target information to another computer or the like via the interface 406. Alternatively, a natural language sentence may be input from the keyboard 405 and the search target information may be output to the display 404. Various databases referred to in the search process are stored in the external storage device 40.
3 is stored. Further, a database of another computer can be referred to via the interface 406.

As described above, the present invention has been specifically described based on the embodiments. However, the present invention is not limited to the above embodiments, and can be variously modified without departing from the gist thereof.

[0170]

As described above, according to the present invention, in order to perform the analysis from the top down, starting from the head of the sentence,
Only the phrase of the specific designated deep case necessary for later keyword extraction is to be analyzed, and unnecessary analysis inside the deep case is not performed, so that the execution time of the analysis can be reduced.

In addition, since knowledge that prescribes, for each type of phrase, its head, the surface case of the head, and the position of the qualifier that modifies the head on the parse tree is held, the flow of the entire semantic analysis is reduced. It is easy to understand visually, and maintenance such as correction of the processing method becomes easy.

[Brief description of the drawings]

FIG. 1 is a diagram for explaining the principle of a semantic analysis method according to the present invention.

FIG. 2 is a flowchart showing the operation of the information search method of the present invention.

FIG. 3 is a configuration diagram of an information search device of the present invention.

FIG. 4 is a configuration diagram of a semantic analysis device of the present invention.

FIG. 5 is a configuration diagram of a deep case frame extraction unit of the present invention.

FIG. 6 is a flowchart showing the operation of the semantic analysis method of the present invention.

FIG. 7 is a flowchart of a deep case frame extraction process of the present invention.

FIG. 8 is a diagram for explaining types and symbols of phrases and parts of speech according to the first embodiment of the present invention.

FIG. 9 is a diagram illustrating a syntax tree of a natural language sentence according to the first embodiment of this invention.

FIG. 10 is a diagram illustrating a semantic structure output by the semantic analysis processing according to the first embodiment of this invention.

FIG. 11 is a diagram visually illustrating a semantic structure output by a semantic analysis process when a syntax tree according to the first embodiment of the present invention is input.

FIG. 12 is a diagram illustrating an example of analysis target category classification knowledge stored in the analysis target category classification knowledge storage device according to the first embodiment of this invention;

FIG. 13 is an example of syntax tree position knowledge stored in the syntax tree position knowledge storage device according to the first embodiment of this invention.

FIG. 14 is a diagram illustrating an example of a value of semantic information of a verb node of a syntax tree according to the first embodiment of this invention.

FIG. 15 is a diagram illustrating types and symbols of phrases and parts of speech according to the second embodiment of the present invention.

FIG. 16 is a diagram illustrating a syntax tree of a natural language sentence according to the second embodiment of this invention.

FIG. 17 is a diagram illustrating a semantic structure output by the semantic analysis processing according to the second embodiment of this invention.

FIG. 18 is a diagram visually illustrating a semantic structure output by a semantic analysis process when a syntax tree according to the second embodiment of the present invention is input.

FIG. 19 is a diagram illustrating an example of analysis target category classification knowledge stored in the analysis target category classification knowledge storage device according to the second embodiment of this invention.

FIG. 20 is an example of knowledge relating to a normal position of a surface case, of knowledge of positions on a syntax tree stored in a storage device for knowledge of positions on a syntax tree according to a second embodiment of the present invention.

FIG. 21 does not have a node corresponding to the surface case for the subject in the input syntax tree among the syntax tree location knowledge stored in the syntax tree location knowledge storage device according to the second embodiment of the present invention. In this case, it is an example of knowledge that defines the position of the destination of the surface case.

FIG. 22 is a diagram illustrating an example of a value of semantic information of a verb node of a syntax tree according to the second embodiment of this invention.

FIG. 23 is a diagram illustrating a configuration example of a computer.

[Explanation of symbols]

 Reference Signs List 1 input device 2 information retrieval device 3 overall control device 4 natural language processing device 5 morphological analysis device 6 morphological analysis unit 7 word dictionary 8 syntax analysis device 9 syntax analysis unit 10 phrase structure rule table 11 semantic analysis device 12 semantic analysis unit 13 various Knowledge storage device 14 Keyword extraction device 15 Keyword extraction unit 16 Extraction rule table 17 Category name inference device 18 Category name inference unit 19 Various knowledge databases such as thesaurus 20 Word extraction device 21 Word extraction unit 22 Word table to be extracted 23 Address analysis device 24 Address Analysis unit 25 Address table 26 Search device 27 Search unit 28 Guidance information database 29 Display device 100 Analysis target classification determination unit 200 Verb phrase class analysis unit 210 Head verb frame extraction unit 220 Processing head head verb frame determination unit 230 State determination unit 2 0 Extraction target deep case determination unit 250 Deep case frame extraction unit 251 Surface case information acquisition unit 252 Syntax tree node acquisition unit 1 253 Constraint condition matching unit 1 254 Deep case frame generation unit 1 255 Surface case movement destination information acquisition unit 256 Syntax tree Node acquisition unit 2 257 Constraint condition matching unit 2 258 Deep case frame generation unit 2 260 Processing target deep case frame determination unit 270 Deep case frame analysis unit 300 Non-verb phrase class analysis unit 310 Subject frame extraction unit 320 Processing subject head frame determination unit 330 Subject frame analysis unit 340 Modifier phrase frame extraction unit 350 Modification phrase frame determination unit to be processed 360 Modifier phrase frame analysis unit 400 Analysis target category classification knowledge storage device 401 CPU 402 Memory 403 External storage device 404 Display 405 Keyboard 406 Interface 00 on the syntax tree position knowledge storage device 1000 semantic analysis apparatus

Claims (27)

[Claims] 1. A semantic analysis method for analyzing a semantic structure of a natural language sentence, wherein a step of inputting a syntax tree of the natural language sentence, and a subclassification of a phrase attached to a node to be analyzed in the syntax tree. From the type information, it is determined whether the analysis target is a verb phrase having a verb as a head or a non-verb phrase mainly having a noun as a head with reference to the analysis target classification knowledge. If the analysis target is the verb phrase, a verb phrase analysis process for analyzing the relationship between the verb and the deep case of the verb, and the analysis target is the non-verb phrase If so, a non-verb phrase analysis process of analyzing a modification relation including a noun phrase and an adjective phrase, a semantic frame corresponding to a certain node of the syntax tree, a head relation connecting the semantic frames, a deep case A set of links of relations and modification relations, Semantic analysis method characterized by comprising a process of natural outputting the semantic structure of the language sentence. 2. The method according to claim 1, wherein the analysis target category determination step further includes a step of determining whether or not the analysis target is a part of speech, and if the analysis target is a part of speech, the semantic analysis is terminated. Semantic analysis method. 3. The verb phrase class analyzing step obtains a verb node to be the subject to be analyzed with reference to the position knowledge on the parse tree, sets a head verb frame corresponding to the head verb node, and A subject verb frame extraction process of linking a subject relation verb from the analysis target frame corresponding to the analysis target to the head verb frame; and referring to information on a state attached to the node including the analysis target, to determine a state of the analysis target. An extraction target deep case determining step of determining a deep case to be extracted from the deep case described in the verb semantic information associated with the head verb node; and the extraction target deep case determination A parse tree node corresponding to the deep case determined in the process is obtained, a deep case frame corresponding to the parse tree node is set, and from the head verb frame to the deep case frame. A deep case frame extracting step of linking a stratigraphic relationship; and, as a result of the deep case frame extracting step, if the deep case frame is set, a deep case recursively calling the semantic analysis using the deep case frame as an input. When the deep case frame analysis process is completed, the extraction target deep case determination process is performed again, and as a result of the processing in the extraction target deep case determination process, if there is no deep case to be extracted, 2. The semantic analysis method according to claim 1, wherein the semantic analysis including the verb phrase class analyzing step is terminated. 4. The verb / phrase analysis step refers to the knowledge of the position on the parse tree, finds a verb node to be the subject to be analyzed, and sets a head / verb frame corresponding to the head / verb node. A subject verb frame extraction process of linking a subject relationship from the analysis target frame corresponding to the analysis subject to the subject verb frame, and which of the subject verb frames in the subject verb frame is to be analyzed in the subsequent processing. A subject verb frame determining process to be determined; a state determining process of determining a state of the analysis target by referring to information of a state accompanying the node including the analysis target; a verb accompanying the head verb node; From the deep case described in the semantic information, an extraction target deep case determination step of determining a deep case to be extracted, and the extraction target deep case determination step Determining a syntactic tree node corresponding to the stratum case, setting a deep case frame corresponding to the syntactic tree node, and linking a deep case relationship from the head verb frame to the deep case frame, a deep case frame extracting step; A deep case frame to be processed to determine which deep case frame to analyze among the deep case frames set in the deep case frame extracting process; and the deep case determined in the deep case frame to be processed determining process. A deep case frame analyzing step of recursively calling the semantic analysis with a case frame as an input.If the deep case frame analyzing step is completed, the processing target deep case frame determining step is performed again, and the processing target deep case frame is performed. As a result of the determination process, if there is no deep case frame to be processed, the extraction target deep case determination process is performed again. If there is no deep case to be extracted as a result of the processing of the extraction target deep case determination step, the processing of the processing target head verb frame determination step is performed again, and the processing result of the processing target head verb frame determination step ,
The semantic analysis method according to claim 1, wherein if there is no subject verb frame to be processed, the semantic analysis including the verb phrase class analysis process is terminated. 5. The deep case frame extracting step is characterized in that the state determined in the state determining step and verb semantic information associated with the head verb node are referred to to determine the deep case frame in the extraction target deep case determining step. A surface case information obtaining step of obtaining a surface case corresponding to the obtained deep case; a syntax tree node obtaining step of obtaining a syntax tree node corresponding to the surface case based on the surface case and the position knowledge on the syntax tree; A constraint condition matching step of matching information including a phrase, a type, and semantic information attached to the tree node with constraint knowledge of the deep case described in the verb semantic information; and a result of the constraint condition matching step. If true, generate a deep case frame that sets a deep case frame corresponding to the parse tree node and links a deep case relationship from the head verb frame to the deep case frame Semantic analysis method according to claim 3 comprising a degree. 6. As a result of the process of obtaining the surface case information,
If the surface case could not be obtained, or if the syntax tree node could not be obtained, or if the syntax tree node for which the result of the constraint matching process was true was not obtained, the surface case movement was performed. The semantic analysis method according to claim 5, wherein deep case frames are generated by acquiring the preceding information. 7. The non-verb phrase class analyzing step obtains a phrase or a part of speech node as a head of the analysis target by referring to the position knowledge on the parse tree, sets a head frame corresponding to the head node, A subject frame extraction process of linking a subject relationship from the analysis target frame corresponding to the analysis subject to the subject frame; a subject frame analysis process of recursively calling the semantic analysis with the subject frame as an input; With reference to the position knowledge, a modifier node for modifying the head node is obtained, a modifier frame corresponding to the modifier node is set, and a modifier link that links a modifier relationship from the head frame to the modifier frame is provided. A frame extracting step; a qualifying phrase frame extraction determining step of determining whether a qualifying phrase frame is set as a result of the qualifying phrase frame extracting step; As a result of the processing of the frame extraction determination step, if the modifier frame is set, the modifier frame analysis step of recursively calling the semantic analysis with the modifier frame as an input, Is completed, the qualifying phrase frame extraction process is performed again. If the qualifying phrase frame is not determined as a result of the qualifying phrase frame extraction determination process, the semantic analysis including the non-verb phrase class analysis process is terminated. The semantic analysis method according to claim 1, wherein 8. The non-verb phrase class analysis step obtains a phrase or part-of-speech node serving as the subject to be analyzed with reference to the positional knowledge on the parse tree, sets a subject frame corresponding to the subject node, A subject frame extraction process of linking a subject relationship from the analysis subject frame corresponding to the analysis subject to the subject frame, and a subject subject frame for determining which subject frame of the subject frames is to be analyzed in the subsequent processing. A determining step; a head frame analyzing step of recursively calling the semantic analysis with the head frame as an input; and a modifier node for modifying the head node with reference to the position knowledge on the parse tree. Setting a modifier frame corresponding to the node and providing a modifier-related link from the subject frame to the modifier frame; A processing target modifier phrase frame determining step of determining which modifier phrase frame is to be analyzed among the modifier phrase frames set in the frame extraction process; and the modifying phrase determined in the processing target modifier phrase frame determining step. A modifier frame analysis step of recursively calling the semantic analysis with a frame as an input. When the modifier frame analysis step is completed, the processing of the processing target modifier frame determination step is performed again, and the processing target modifier phrase is obtained. As a result of the frame determination step, if there is no modifier frame to be processed, the processing of the processing target head frame determination step is performed again. As a result of the processing of the processing target head frame determination step, there is no head frame to be processed. 2. The semantic analysis method according to claim 1, wherein the semantic analysis including the non-verb phrase analysis step is terminated. 9. An information retrieval method using the semantic analysis method according to claim 1, wherein the information retrieval method is executed on a computer having a storage device and an interface unit, and inputs a natural language sentence via the interface unit. Performing a morphological analysis and a syntax analysis on the natural language sentence,
Generating the syntax tree of the natural language sentence; performing the semantic analysis using the syntax tree as an input to generate a semantic structure of the natural language sentence; extracting a keyword representing a user's intention from the semantic structure Acquiring the category name and the address information corresponding to the natural language sentence by using the keyword and referring to at least one knowledge database stored in the storage device; and Acquiring information to be searched by referring to an information database including information to be searched classified by the category name and the address information using the information; and transmitting the obtained information to another computer via the interface unit. Transmitting or displaying the information on a display connectable to the computer. 10. A semantic analysis device for analyzing a semantic structure of a natural language sentence, a means for inputting a parse tree of the natural language sentence, and an analysis target category for determining whether the verb phrase or the non-verb phrase. The analysis target classification classification knowledge storage device in which the classification knowledge is stored, and attached to the node to be analyzed in the syntax tree, from the information of the classification that is the sub-classification of the phrase, referring to the analysis target classification classification knowledge, Analysis target classification determining means for determining whether the analysis target is a verb phrase class having a verb as a head or a non-verb phrase mainly having a noun as a head, and a syntax for obtaining a node of a syntax tree A syntax tree position knowledge storage device in which tree position knowledge is stored, and if the analysis target is the verb phrase, a verb and a verb phrase class analyzing means for analyzing a relationship between the verb and the deep case of the verb. If the object of analysis is the non-verb phrases A non-verb phrase class analyzing means for analyzing a modification relation including a noun phrase and an adjective phrase; a semantic frame corresponding to a certain node of the syntax tree; a head relation, a deep case relation, and a modification connecting the semantic frames; Means for outputting a semantic structure of the natural language sentence, comprising a set of relational links. 11. The semantic analysis according to claim 10, wherein said analysis target classification determining means further comprises means for determining whether or not the analysis target is a part of speech, and if the analysis target is a part of speech, terminates the semantic analysis. Semantic analyzer. 12. The verb phrase analysis means obtains a verb node to be the subject to be analyzed with reference to the position knowledge on the parse tree, and sets a head verb frame corresponding to the head verb node, A subject verb frame extracting means for linking a subject relation from the analysis target frame corresponding to the analysis target to the head verb frame; and by referring to information on a state attached to a node including the analysis target, the analysis target State determination means for determining a state; extraction target deep case determination means for determining a deep case to be extracted from the deep case described in the verb meaning information attached to the head verb node; and the extraction target deep case A parse tree node corresponding to the deep case determined by the determining means is obtained, a deep case frame corresponding to the parse tree node is set, and the deep case frame is determined from the head verb frame. A deep case frame extracting means for linking a deep case relationship to a system, and if the deep case frame is set as a result of the processing of the deep case frame extracting means, the semantic analysis is recursively performed using the deep case frame as an input. When the processing of the deep case frame analyzing means is finished, the processing of the extraction target deep case determining means is performed again, and as a result of the processing in the extraction target deep case determining means, The semantic analysis device according to claim 10, wherein if there is no deep case to be extracted, the semantic analysis including the processing of the verb phrase class analyzing means is ended. 13. The verb phrase class analyzing means obtains a verb node to be the subject of the analysis with reference to the position knowledge on the parse tree, sets a main verb frame corresponding to the main verb node, A head verb frame extracting means for linking a head relation from the analysis target frame corresponding to the analysis target to the head verb frame; and analyzing the deep case of any head verb frame among the head verb frames in the subsequent processing. A processing target main verb frame determining means for determining whether the state is to be analyzed by referring to information of a state associated with the node including the analysis target; and a state determination means for determining the state of the analysis target, An extraction target deep case determining means for determining a deep case to be extracted from the deep case described in the verb semantic information; and A deep case frame extracting means for obtaining a parse tree node corresponding to the determined deep case, setting a deep case frame corresponding to the parse tree node, and linking a deep case relationship from the head verb frame to the deep case frame Processing target deep case frame determining means for determining which deep case frame to analyze among the deep case frames set in the processing of the deep case frame extracting means; and Deep case frame analysis means for recursively calling the semantic analysis with the deep case frame determined in the processing as input, and when the processing of the deep case frame analysis means ends, the processing target deep case frame determination is performed again. Means, and as a result of the processing of the deep case frame determining means for processing, if there is no deep case frame to be processed, The processing of the extraction target deep case determination means is performed, and as a result of the processing of the extraction target deep case determination means, if there is no deep case to be extracted, the processing of the processing target main verb frame determination means is performed again. As a result of the processing of the target head verb frame determining means, if there is no head verb frame to be processed,
The semantic analysis device according to claim 10, wherein the semantic analysis including the processing of the verb phrase class analyzing means is ended. 14. The deep case determining unit according to claim 1, wherein the deep case frame extracting unit refers to the state determined in the processing of the state determining unit and verb meaning information attached to the head verb node. Surface case information obtaining means for obtaining a surface case corresponding to the deep case determined in the processing of the above, and a syntax tree node obtaining means for obtaining a syntax tree node corresponding to the surface case based on the surface case and the position knowledge on the syntax tree Constraint information matching means for matching information including a phrase, a type, and semantic information associated with the syntax tree node with constraint information of the deep case described in the verb semantic information; and If the result of the processing by the means is true, a deep case frame corresponding to the parse tree node is set, and a deep case relation link is established from the head verb frame to the deep case frame. Semantic analysis apparatus according to claim 12 consisting of a deep case frame generating means. 15. As a result of the processing of the surface case information obtaining means, if the surface case cannot be obtained, or if the parse tree node cannot be obtained, or the result of the processing of the constraint condition matching means. The semantic analysis device according to claim 14, further comprising: means for acquiring surface case destination information and generating a deep case frame when a syntax tree node for which is true is not obtained. 16. The non-verb phrase class analyzing means obtains a phrase or a part of speech node to be the subject to be analyzed with reference to the position knowledge on the parse tree, and sets a subject frame corresponding to the subject node. A subject frame extracting means for linking a subject relationship from the analysis target frame corresponding to the analysis subject to the subject frame, a subject frame analyzing means for recursively calling the semantic analysis by using the subject frame as an input, and the syntax tree With reference to the upper position knowledge, a modifier node that modifies the head node is obtained, a modifier frame corresponding to the modifier node is set, and a modification relation link is provided from the head frame to the modifier frame. Phrase frame extraction means, as a result of the processing of the decoration phrase frame extraction means, decoration phrase frame extraction determination means for determining whether a decoration phrase frame is set, And a modifier frame analyzing means for recursively calling the semantic analysis by using the modifier frame as an input if the modifier frame is set as a result of the processing of the modifier frame extraction determining means. When the processing of the frame analyzing means is completed, the processing of the modifier phrase frame extracting means is performed again. As a result of the processing of the modifier frame extraction determining means, if the modifying phrase frame is not set, the non-verb phrase analysis is performed. The semantic analysis device according to claim 10, wherein the semantic analysis including the processing of the means is terminated. 17. The non-verb phrase class analyzing means refers to the positional knowledge on the parse tree to find a phrase or a part of speech node to be the subject of the analysis, and sets a subject frame corresponding to the subject node. A subject frame extracting means for linking a subject relationship from the analysis target frame corresponding to the analysis subject to the subject frame, and a processing subject subject which determines which subject frame of the subject frames is to be analyzed in the subsequent processing. Frame determining means; head frame analyzing means for recursively calling the semantic analysis with the head frame as an input; referring to the parse tree position knowledge to obtain a modifier node for modifying the head node; A modifier frame extracting means for setting a modifier frame corresponding to the phrase node and linking a modifier relation link from the subject frame to the modifier frame; Processing target modifier phrase frame determining means for determining which modifier phrase frame is to be analyzed among the modifier phrase frames set in the processing of the phrase frame extracting means, and the processing is performed by the processing target modifier phrase frame determining means. Modifier frame analysis means for recursively calling the semantic analysis with the modifier frame input as input, and when the processing of the modifier frame analysis means ends, the processing of the processing target modifier phrase frame determining means is performed again. If there is no modifier frame to be processed as a result of the processing of the processing target modifier frame determining means, the processing of the processing target head frame determining means is performed again, and the result of the processing of the processing target head frame determining means is performed. 11. The method according to claim 10, wherein if there is no subject frame to be processed, the semantic analysis including the processing of the non-verb phrase class analyzing means is ended. Semantic analysis device. 18. An information retrieval apparatus using the semantic analysis device according to claim 10, wherein: a storage device and an interface unit; means for inputting a natural language sentence via the interface unit; Perform morphological analysis and parsing,
Means for generating a syntax tree of the natural language sentence; means for performing the semantic analysis by using the syntax tree as an input to generate a semantic structure of the natural language sentence; extracting a keyword representing a user's intention from the semantic structure Means for referring to at least one knowledge database stored in the storage device using the keyword, and acquiring category name and address information corresponding to the natural language sentence; and Means for obtaining search target information by referring to an information database including search target information classified by the category name and the address information using information, and obtaining the obtained information to another computer via the interface unit. Means for transmitting or displaying on a display connectable to the computer. 19. A recording medium on which a semantic analysis program for analyzing a semantic structure of a natural language sentence is recorded, wherein a procedure for inputting a syntax tree of the natural language sentence is provided, and , From the information of the type that is the sub-classification of the phrase, referring to the analysis target classification classification knowledge, the analysis target is a verb phrase having a verb as a head, or a non-verb phrase having a noun as a head. An analysis target category determination procedure for determining whether or not the verb phrase is the verb phrase, a verb and a verb phrase analysis procedure for analyzing a relationship between the verb and the deep case of the verb; For the non-verb phrases, a non-verb phrase analysis procedure for analyzing a modification relation including a noun phrase and an adjective phrase; a semantic frame corresponding to a certain node of the syntax tree; and connecting the semantic frames. Head relations, deep case relations, and qualification relations Consisting of a set of links, a recording medium recording the semantic analysis program characterized by comprising the steps of outputting the semantic structure of the natural language text. 20. The analysis method according to claim 19, wherein the analysis target category determination step further includes a step of determining whether or not the analysis target is a part of speech. If the analysis target is a part of speech, the semantic analysis is terminated. A recording medium on which a semantic analysis program is recorded. 21. The verb / phrase analysis procedure includes: obtaining a verb node to be the subject of the analysis with reference to the position knowledge on the parse tree; setting a head / verb frame corresponding to the head / verb node; A subject verb frame extraction procedure for linking a subject relation verb from the analysis target frame corresponding to the analysis target to the head verb frame; and An extraction target deep case determining step of determining a deep case to be extracted from the deep case described in the verb semantic information associated with the head verb node; and the extraction target deep case determination A syntax tree node corresponding to the deep case determined in the procedure is obtained, a deep case frame corresponding to the syntax tree node is set, and the deep case frame is determined from the head verb frame. A deep case frame extracting procedure for establishing a link of a deep case relationship; and, as a result of the deep case frame extracting procedure, if the deep case frame is set, a deep case for recursively calling the semantic analysis using the deep case frame as an input. When the deep case frame analysis procedure is completed, the processing of the extraction target deep case determination procedure is performed again. As a result of the processing in the extraction target deep case determination procedure, the deep case to be extracted is obtained. 20. A recording medium recording the semantic analysis program according to claim 19, wherein if there is no, the semantic analysis including the verb phrase analysis procedure is terminated. 22. The verb-phrase analysis procedure includes: obtaining a verb node to be the subject to be analyzed with reference to the position knowledge on the parse tree; setting a head-verb frame corresponding to the head-verb node; A subject verb frame extraction procedure for linking a subject relation from the analysis target frame corresponding to the analysis subject to the subject verb frame, and which of the subject verb frames in the subject verb frame is to be analyzed in the subsequent processing. A processing subject main verb frame determining procedure to be determined; a state determining procedure for determining a state of the analysis target by referring to information of a state accompanying the node including the analysis target; and a verb accompanying the head verb node. An extraction target deep case determination procedure for determining a deep case to be extracted from the deep case described in the semantic information, and a determination in the extraction target deep case determination procedure. A deep case frame extracting step of finding a parse tree node corresponding to the obtained deep case, setting a deep case frame corresponding to the parse tree node, and linking a deep case relationship from the head verb frame to the deep case frame. A processing target deep case frame determining step of determining which deep case frame to analyze among the deep case frames set in the deep case frame extracting step; and a processing target deep case frame determining step. A deep case frame analysis procedure for recursively calling the semantic analysis with the deep case frame as an input.When the deep case frame analysis procedure is completed, the processing of the target deep case frame determination procedure is performed again. If there is no deep case frame to be processed as a result of the processing of the deep case frame to be processed determination procedure, Performs the processing order, the extraction target deep case determination procedure of processing results, if there is no deep cases to be extracted, again performs processing of the processing target head-verb frame determination procedure,
As a result of the processing of the processing subject head verb frame determination procedure,
20. A recording medium storing the semantic analysis program according to claim 19, wherein if there is no subject verb frame to be processed, the semantic analysis including the verb phrase analysis procedure is terminated. 23. The deep case frame extraction procedure, wherein the state determined in the state determination step and the verb meaning information attached to the head verb node are referred to to determine in the extraction target deep case determination procedure. A surface case information obtaining step of obtaining a surface case corresponding to the obtained deep case; a syntax tree node obtaining step of obtaining a syntax tree node corresponding to the surface case based on the surface case and the syntactic tree position knowledge; A constraint condition matching procedure for matching information including a phrase, a type, and semantic information associated with a tree node, and the constraint condition knowledge of the deep case described in the verb semantic information; and a result of the constraint condition matching procedure. If true, a deep case frame corresponding to the parse tree node is set, and a deep case frame linking a deep case relationship from the head verb frame to the deep case frame. Recording medium recording the semantic analysis program according to claim 21 comprising the steps. 24. As a result of the processing of the surface case information obtaining procedure, if the surface case cannot be obtained, or if the parse tree node cannot be obtained, or if the result of the constraint matching procedure is true. 24. The recording medium storing the semantic analysis program according to claim 23, further comprising a step of acquiring surface case movement destination information and generating a deep case frame when a parse tree node is not obtained. 25. The non-verb phrase class analysis procedure obtains a phrase or part-of-speech node serving as the subject to be analyzed with reference to the position knowledge on the parse tree, sets a subject frame corresponding to the subject node, A subject frame extraction procedure for linking a subject relationship from the analysis target frame corresponding to the analysis subject to the subject frame; a subject frame analysis procedure for recursively calling the semantic analysis with the subject frame as an input; With reference to the position knowledge, a modifier node for modifying the head node is obtained, a modifier frame corresponding to the modifier node is set, and a modifier link that links a modifier relationship from the head frame to the modifier frame is provided. A frame extraction procedure; a modifier frame extraction determination procedure for determining whether a modifier frame is set as a result of the processing of the modifier frame extraction procedure; As a result of the processing of the decoration phrase frame extraction determination procedure, if the decoration phrase frame is set, the decoration phrase frame analysis step of recursively calling the semantic analysis with the decoration phrase frame as an input, When the processing of the procedure is completed, the processing of the qualifying phrase frame extraction procedure is performed again. As a result of the processing of the qualifying phrase frame extraction determining procedure, if the qualifying phrase frame is not set, the non-verb phrase class analysis procedure is performed. 20. A recording medium on which the semantic analysis program according to claim 19 is terminated. 26. The non-verb phrase class analysis procedure obtains a phrase or part-of-speech node serving as the subject to be analyzed with reference to the positional knowledge on the parse tree, sets a subject frame corresponding to the subject node, A subject frame extraction procedure for linking a subject relationship from the analysis subject frame corresponding to the analysis subject to the subject frame, and a subject subject frame for determining which subject frame of the subject frames is to be analyzed in the subsequent processing. A determination procedure; a head frame analysis procedure for recursively calling the semantic analysis using the head frame as an input; and a modifier node for modifying the head node with reference to the positional knowledge on the parse tree. A modifier frame extraction procedure for setting a modifier frame corresponding to a node and linking a modifier-related link from the head frame to the modifier frame; Among the modifier frames set in the processing of the frame extraction procedure, a processing target modifier frame determination procedure for determining which modifier frame is to be analyzed, and the processing is performed in the processing target modifier frame determination procedure. A modifier frame analysis procedure for recursively calling the semantic analysis with the modifier frame as an input.If the processing of the modifier frame analysis procedure is completed, the processing of the processing target modifier frame determination procedure is performed again. As a result of the processing target modifier frame determination procedure, if there is no modifier phrase frame to be processed, the processing of the processing target head frame determination procedure is performed again, and as a result of the processing of the processing target head frame determination procedure, 20. The semantic analysis program according to claim 19, wherein the semantic analysis including the non-verb phrase class analysis procedure is terminated if there is no subject frame to be executed. A recording medium on which a program is recorded. 27. A recording medium storing an information search program having the semantic analysis program according to claim 19, wherein the information search program is executed on a computer having a storage device and an interface unit, and Inputting a natural language sentence, performing morphological analysis and syntax analysis on the natural language sentence,
A step of generating a syntax tree of the natural language sentence; a step of performing the semantic analysis using the syntax tree as an input to generate a semantic structure of the natural language sentence; extracting a keyword representing a user's intention from the semantic structure Using the keyword, referring to at least one knowledge database stored in the storage device, and acquiring the category name and address information corresponding to the natural language sentence; and A procedure for obtaining search target information by referring to an information database including search target information classified by the category name and the address information using the information; and transmitting the obtained information to another computer via the interface unit. Transmitting or displaying the information on a display connectable to the computer. Recording the recording medium.