JP2000259627A - Device and method for deciding relation between natural language sentences, retrieving device and method utilizing the deciding device and method and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide technology for suitably deciding similarity between two sentences by grammatical level processing and retrieving a sentence based on the decision. SOLUTION: Respective words are extracted from an input character string A consisting of m words from a word a1 up to a word am and a comparing character string B consisting of n words from a word b1 up to a word bn, and the distance t(ai, bj) between respective words is found out by referring to a synonym dictionary 36 or the like. Arithmetic processing is executed by using the value of the distance t(ai, bj) between respective words and the values of omission costs r, q of words (steps S515 to S525) and distances d(ai, bj) between respective word strings are successively found out in all word strings considered on the assumption that the order of words is kept (step S535). After finding out the values of distances d(am, bn) of word strings including all words, the value of similarity s(am, bn) between respective character strings is found out by arithmetic processing using the value of the distance d(am, bn) and the similarity between respective character strings is decided on the basis of the found value.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for handling a natural language, and more particularly, to an apparatus for determining the relationship between two sentences in a natural language sentence, and a method for determining the relationship and a function for performing the determination. It relates to a recording medium.

[0002]

2. Description of the Related Art A language used by humans for communication is called a natural language, as opposed to an artificial language for programming, and its form should be considered as a whole of communication performed using a language. It cannot be grasped as a combination of a small number of principles. There are so-called grammars for each language, but this is just an attempt to organize some of the vast rules that exist in natural language into some easy-to-understand rules, and does not completely describe natural language. It is well known.

[0003] In Japan, such a technology for handling natural languages has been independently developed, for example, in the form of kana-kanji conversion. Based on the input kana characters, in order to obtain the kana-kanji mixed sentence expected by the input person, not only grammatical analysis but also recent methods such as example conversion and conversion using dependency are realized. I have. By using such a method, the kana character "Atsu" can be changed to "Hot summer"
And the case of "tea is hot".

[0004] As other large technologies relating to natural languages, search and machine translation of natural language sentences, and more recently, generation of summary sentences are known. Natural language sentence search is based on a method of searching for a character string that exactly matches the word to be searched in the search target character string, and furthermore, a logical sum and a logical product of search results of a plurality of search words. , Search for conceptual similar words using a thesaurus, etc. have been put to practical use.
As an example, in the case of entering the phrase “Tokyo's delicious shop” as a sentence to be searched (hereinafter referred to as a “search key sentence”) and searching the Internet homepage, simply use a conceptual expression. By preparing a certain "Tokyo" or "Delicious shop" as a search key and performing a compound word search, "Tokyo" or "Tokyo", which is considered to be a synonym of "Tokyo" or an expression indicating a higher or lower concept It is not possible to search for explanatory sentences using such terms. Similarly, for “delicious shop”, it is necessary to prepare words such as “famous shop” and “gourmet” as search keys. Such a search for a natural language sentence is extremely useful, for example, in a search of a homepage of the Internet existing in the whole world and a search of a large amount of articles.

[0005] Proposals for such a search include “video search based on similarity of index sentences” (Ichiro Yamada et al., The 5th National Symposium on International Symposium, 1st Subcommittee, August 1997), Yasuaki Hyodo et al., "Natural Language Processing," Vol 3, No. 2, August 1997. In these papers, we use a "classified vocabulary table" compiled by the National Institute for Japanese Language and Language, and search for nouns and verbs in consideration of the similarity between words. Further, the determination of the similarity between each word and the determination of the similarity of the entire sentence are performed on the premise of the matching of the syntax patterns such as “XX changes △△ to □□”.

On the other hand, several approaches have been proposed for machine translation depending on the similarity between languages. For example, between languages belonging to the same family with the same fundamental rules of grammar, such as between German and French, a certain degree of translation is possible even with a method based on replacing elements constituting a sentence. On the other hand, it is known that, for example, translation between English, which belongs to the inflected word, and Japanese, which is one of the sticky words, is difficult to translate by replacing words.
Therefore, translation of syntax versus syntax has been attempted through a syntax analysis process, but there are many problems such as difficulty in narrowing down the diversity of analysis results, and sufficient results have not been obtained. Under such circumstances, in recent years, a large number of translation examples translated by a translator are collected, and when a sentence to be translated is given, a sentence similar to this sentence is searched, and the translated sentence is referred to. Attention has been paid to a method of performing translation by replacing words. In this case, a sentence having a structure similar to the sentence to be translated (a sentence with high similarity) is searched from a large number of example sentences.

[0007] Considering such processing as searching and translating a sentence in a natural language, and further preparing a summary sentence, it is considered that analysis of the meaning of the sentence in the natural language is ultimately necessary. . Alternatively, as a technique before reaching the meaning, a method of preparing a large number of expression examples expressed in a natural language and referring to them can be considered. Regarding the former, there is difficulty in setting semantic rules, etc., and semantic inference using neural networks and expert systems have been proposed. In the latter case, large-scale example dictionaries or dependency dictionaries have recently become available, and for example, in kana-kanji conversion, it tries to correctly convert "summer is hot" and "tea is hot". A proposal has been made.

[0008]

However, conventional natural language processing has a problem in that it is not possible to perform sufficient search and translation processing. When processing sentences in natural language, if you want to perform high-precision searches,
Eventually, it will be necessary to deal with the meaning represented by the sentence. However, a technique for simply handling the meaning has not yet been put to practical use, and cannot be immediately applied to search and translation at present. On the other hand, the conventional technology of compound word search cannot handle a large amount of natural language data with high accuracy.

As described above, a search method that considers the similarity between words has also been proposed. However, only a conceptual expression such as a noun or a verb is examined, and if the syntax does not match, a similar meaning is shown. There is a high possibility that a sentence will be omitted from the search target. For example, if you use the phrase "Tokyo's delicious shops", you can search for "Great restaurants in the Tokyo metropolitan area" and "Gourmet guides in Tokyo", but you can search for "Good shops in the 23 wards of Tokyo" and "Good restaurants in Tokyo." Is likely not searchable. On the other hand, there is a high possibility that "good restaurants outside Tokyo" will be searched.

The present invention solves these problems and does not enter into the technology of recognizing meaning deeply. Rather, it mainly determines the relationship between natural language sentences by processing at the grammar level. An object of the present invention is to propose a technology for searching for natural language sentences based on the sentence.

[0011]

Means for Solving the Problems and Their Functions / Effects In order to solve at least a part of the above-mentioned problems, the following series of inventions have been made. Is placed. In other words, the natural language sentence relation determination device of the present invention is a sentence in a predetermined language representing content having a certain unity, and the relationship between the first sentence to be determined and the first sentence is determined. A second sentence, and a constituent unit constituting a sentence in the language,
A natural language sentence relation determination device that determines a relationship between the first sentence and the second sentence by using constituent units classified as having united meaning, wherein Of the conceptual expressions classified as structural units representing semantic concepts and the framework expressions extracted as structural units corresponding to the expressions supporting the framework of the sentence structure, at least the relationships between the conceptual expressions and the relationships between the framework expressions are expressed. With reference to the information stored in the relational information storage means, the constituent unit extraction means for extracting the constituent unit from the first sentence and the second sentence, and the relational information storage means, By determining the relationship between the conceptual expressions and the framework expressions among the constituent units constituting the extracted first and second sentences while considering the correspondence in word order. It is summarized in that and a determining relationship determination means the relationship between the second sentence and the first sentence.

Here, "sentence" refers to a linguistic expression that expresses some semantic content by a group of words.
In addition to a complete expression including a predicate, the expression includes an expression in units of a phrase that is a delimiter of a word, an expression composed of a combination of one-word sentences, and the like. For example, the statement "I want to. You know a good store of Tokyo", which was completed that, a phrase such as "Tokyo of good shops" and "good store", "Teacher, Hello" is also a combination of one word sentences such as, " Sentence ".

According to this device, the concept expression and the framework expression included in these sentences are extracted from the first sentence and the second sentence, and at least information indicating the relationship between the concept expressions and the relationship between the framework expressions is obtained. , The relationship between the first sentence and the second sentence is determined in consideration of the correspondence between the word orders of the expressions included in both sentences.

Here, "the relation between the first sentence and the second sentence"
Means a perceived relationship between the first sentence and the second sentence, for example, a class relating to the meaning of a sentence such that the sentence has the same, similar, different, opposite, etc. In addition to nos and nos, it includes similarities and differences regarding the usage of sentences. The usage of this sentence is, for example, sentence sentences and colloquial sentences,
In addition to the grammatical distinctions of stereotypes and honorifics,
Sentences based only on grammatical rules: sentences using standard languages and dialects, sentences created by men and sentences created by women, sentences created by people in their 20s, and sentences created by people in their 50s Including things that cannot be done. A combination of a plurality of elements as such a relationship may be referred to as a “relationship between the first sentence and the second sentence”. As for the relationship between the first sentence and the second sentence, a specific relationship (for example, a similar relationship) may be specified in advance. It may be given by the user.

The concept expression is a structural unit constituting a sentence in a predetermined language, and is an expression representing a semantic concept among constituent units classified as having a collective meaning, and is a noun, a verb, an adjective. And the combination of the concept words. In addition, the framework expression is a structural unit constituting a sentence in a predetermined language, and is extracted as a structural unit corresponding to an expression supporting a framework of a sentence structure among structural units classified as having a unity meaning. It is a thing. The inventor has widely sampled such framework expressions for Japanese, which is one of the sticky words, and said, "Expressions that give a framework of Japanese sentence structure-functional categories and connection rules-" (Report of Fukuoka University Research Institute) No. 63, Showa 58
March, 1988) and "Expression that gives a framework of Japanese sentence structure-arrangement of structural semantic information-" (Fukuoka University Research Institute Bulletin No. 63, March 1983). Framework expressions include relational expressions categorized in these papers (expressions representing inter-concept relations such as case relations and causal relations in one sentence, case particles in Japanese, connective particles, and expressions equivalent to these). Includes advisory expressions in Japanese that give information in a broader sense (speaker or writer's judgment and attitude, tense, aspect, negation, attitude, etc.).

The term "relationship between framework expressions" means a connection recognized between framework expressions, and the type of this relationship is, for example, the same or similar in meaning of framework expressions. In addition to the similarity and similarity regarding the meaning of the framework expression such as, different, opposite, etc., the similarity and similarity regarding the attributes of the framework expression are included. Attributes of this framework expression include, for example, tense and state of the framework expression, degree of estimation, degree of emphasis and limitation, distinction between affirmation and negation, distinction between spoken and spoken words, distinction between ordinary and honorific, etc. In addition to those that can be distinguished from each other, those that cannot be distinguished only by grammatical rules such as the distinction between standard and dialects, the distinction between male and female languages, and the age group using framework expressions can be considered. Of course, a combination of a plurality of the above-described types of relationships may be referred to as a “relationship between framework expressions”.

In such a natural language sentence relation determining apparatus,
When determining the relationship between the first sentence and the second sentence, it is preferable that the difference in the appearance order of the extracted constituent units is allowed as the correspondence of the word order of the constituent units. This is because, in the judgment between natural language sentences, the word order of each word is often not the same. In order to simplify the determination process, it is also preferable to prohibit the intersection of two or more sets of correspondence between constituent units when allowing the difference in the order of appearance in the correspondence in word order.

Further, in the natural language sentence relation determining apparatus, the relation between the conceptual expressions and the relation between the framework expressions among the constituent units constituting the first and second sentences is defined by the relation information. By making a determination with reference to the information stored in the storage means and using the determination result between the conceptual expressions and the determination result between the framework expressions while associating the word order, the first and the second are obtained. The relationship between the two sentences may be determined. According to such a configuration, the relationship between conceptual expressions and the relationship between framework expressions can be determined individually, so that the processing can be simplified.

When determining the relationship between constituent units, for each constituent unit extracted from either the first or second sentence, the corresponding constituent unit does not always exist in the other sentence. It is also conceivable that one constituent unit of one sentence has a plurality of constituent units having a certain relation in the other sentence. In such a case, it is considered that the relationship between the two sentences can be correctly determined as a whole by determining that the corresponding constituent unit is missing. In this determination, a value in the case where the corresponding constituent unit does not exist is set in advance as a missing value, and the constituent units extracted from the first and second sentences are determined while associating the word order. It can be performed by assigning a value based on the relationship as a relationship value to those having a predetermined relationship, and evaluating the assigned relationship value and the set dropout value.
By performing this evaluation, a relationship value between the first and second sentences is obtained, and the determination may be made based on the magnitude of the relationship value.

The missing value when there is no corresponding constituent unit may be set to a uniform value, but may be changed according to the length of the sentence for which the relation is to be determined. If it is considered that a sentence having many constituent units has a high degree of redundancy, it is desirable to set so that the influence of dropout is evaluated to be small. It is also desirable to provide a means for setting the dropout value to a value corresponding to the importance of the constituent unit, since the relationship between sentences can be substantially determined. In addition, a unit that does not exist in the other sentence
A different dropout value may be set depending on whether the expression is a concept expression or a framework expression. If the relationship to be determined is, for example, the similarity between sentences, the omission of the conceptual expression is considered to have a large effect on the determination of the similarity. It is conceivable that the evaluation is made larger than the case of dropping. In addition, it is also preferable to say that the dropout value of the framework expression is set to be larger in the case of determining the relation of the structure of the expression instead of the semantic content.

When judging the relation between constituent units constituting a sentence and judging the relation between sentences, it is sufficient to usually examine the relation between conceptual expressions and between frame expressions. It is thought that there is little need to pre-define the relationship between the two. However, it is desirable to store, as information, the relationship between the concept affix and the concept word, in addition to the relationship between concept words that independently represent the semantic concept. For example, it is desirable to evaluate the concept suffix “new” and the concept word “new” as having a similarity. Conceptual affix "target"
The same applies to the concept word "style".

It has already been described that various relationships can be assumed between the sentences to be determined. If the relationships are similar, information indicating the degree of similarity between conceptual expressions and framework expressions is stored. Can be kept. In this case, the similarity between the first and second sentences is determined by referring to the degree of similarity between the stored conceptual expressions and between the framework expressions. When similarity is determined as the relationship between sentences, the application range is the widest in processing such as searching for an example sentence for searching or translating a natural language sentence, or deforming or compressing (summarizing) a sentence.

The determination of similarity is performed by comparing pairs of conceptual expressions and frame expressions with similarity numerical data representing numerically the semantic similarity between the paired expressions. It is convenient to store the information representing the similarity and to determine the similarity using the magnitude of the similarity numerical value data.

As such similarity numerical data, the value 0
In the range from 1 to 1, a value approaching the value 1 can be used as the degree of similarity increases, and the distance between conceptual expressions and between frame expressions is calculated from the value of the similarity numerical data. The degree of similarity between the first and second sentences may be obtained by specifying the shortest combination. The method of determining the degree of similarity by using the sum of distances provides a concrete image in discussing the similarity between two sentences, and facilitates understanding. Further, it is easy to apply a conventionally known technique such as pattern matching.

A natural language for retrieving a sentence similar to a search key sentence given as a key for a search from a plurality of search target sentences using the natural language sentence relation determining apparatus for judging similarity between sentences. A sentence retrieval device can be configured. That is, the natural language sentence search device includes a natural language sentence relationship determination device that determines the similarity between the sentences described above, a first sentence specifying unit that specifies the search key sentence as a first sentence, One sentence is sequentially selected from a plurality of search target sentences, and the second sentence is selected.
Second sentence specifying means for specifying the first sentence and the second sentence to the natural language sentence relation determination device, and performing the similarity determination. And the determination result of the natural language sentence relation determination device is stored according to the assigned second sentence, and a second sentence most similar to the search key sentence given as the first sentence is
Selecting means for selecting from the plurality of search target sentences.

This natural language sentence retrieval device sequentially selects one sentence from a plurality of retrieval target sentences, determines the similarity between this sentence and the retrieval key sentence by a natural language sentence determination device, and stores the result of this determination. Then, a sentence that is most similar to the search key sentence can be selected from the similarity determination result of each sentence of the plurality of search target sentences. With this configuration, the similarity between two sentences can be determined based on the similarity between concept expressions and framework expressions while considering the correspondence between word orders, and is most similar to the search key sentence. Sentences can be easily searched.

The structure for determining the relationship between sentences in the natural language sentence relationship determination device can be applied to a phrase relationship determination device that determines the relationship between words. That is, for the concept expression that is an expression representing a semantic concept, the word / phrase relationship determination device of the present invention includes a dictionary storing at least information indicating the relationship between the concept expressions, a first word and a second word.
Input means for inputting a word, an extraction means for extracting a word constituting the phrase from the input first word and the second word, and a word constituting the extracted first word. Determining means for determining the relationship between words forming the second word by referring to the dictionary; and determining the relationship between the first word and the second word based on the determination result by the determining means. A determination means for determining whether the first
Or at least one of the second words,
A composite expression which is an expression representing a united semantic concept by combining two or more words, comprising an evaluation unit for evaluating a relationship between the composite expression and an expression corresponding to the composite expression; The gist is that it is means for determining the relationship between the first word and the second word in consideration of the evaluation result by the evaluation means.

With such a configuration, the relationship between the composite expression and the expression corresponding to the composite expression can be accurately determined.

Further, as an invention of a method corresponding to each of the above apparatuses, a natural language sentence relation determination method and a natural language sentence search method can be considered as described in claims 17 to 22.

Similarly, as the invention of a recording medium corresponding to each of the above methods, a recording medium recording a program for determining the relationship between natural language sentences, a recording medium recording a program for searching for a natural language sentence, and the like. To 28, it can be considered.

[0031]

Other Embodiments of the Invention The present invention may be configured as a dedicated machine or may be realized by a highly versatile personal computer or the like. Further, the program for realizing each function stored in the recording medium may be stored and stored in a server or the like connected to a network, and may be used by downloading to an execution machine as needed. . Such a form of the server or a case where the program is transmitted to the public from the server can be regarded as one form of the medium of the present application.

[0032]

DESCRIPTION OF THE PREFERRED EMBODIMENTS In order to further clarify the configuration and operation of the present invention described above, embodiments of the present invention will be described below based on examples. FIG. 1 is a sentence similarity determination device 1A which is an example of the natural language sentence relationship determination device of the present invention.
1 shows a hardware configuration. The inter-sentence similarity determination apparatus 1A according to the first embodiment includes a similarity search engine 10A.
And an external device 90, a similarity search engine 10A
Determines the degree of similarity between the first sentence and the second sentence representing the content having a certain unity and input to the engine 10A (hereinafter, referred to as similarity between character strings). Hereinafter, the first sentence expressed in characters is referred to as “input character string”.
And the second sentence expressed in characters as "contrast character string". The process of determining the similarity between character strings is executed by a computer inside the similarity search engine 10A.

The computer is provided with the following components interconnected by a bus 35, centering on the CPU 22 for executing various arithmetic processing. The ROM 24 is a CPU 2
2 is a memory in which programs and reference data necessary for executing various arithmetic processes are stored in advance, and stores a program relating to execution of determination of similarity between character strings described later. The RAM 26 is a memory for temporarily storing various data necessary for the CPU 22 to execute various arithmetic processes.

Hard Disk Controller (HDC) 3
0 controls signal output to the hard disk 10a as an external storage device. The hard disk 10a stores various programs loaded and executed in the RAM 26 as necessary, programs provided in the form of device drivers and modules, and various dictionaries such as a Japanese language dictionary and a synonym dictionary 36 described later. It is remembered. Of course, various programs similar to those described above, necessary reference data, and the like are stored in the ROM 24, CD-ROM, or the like (not shown), and the computer is executed by loading these various programs and reference data. It is also possible.

The input interface 20 is connected to an external device 90
The output interface 34 controls output of data and character strings to the external device 90 and the printer 92. That is, the similarity search engine 10A is connected to the external device 90 using a cable (not shown), and performs input and output of data and character information with the external device 90. Of course, the input interface 20
A keyboard or a handwritten character recognition board or the like may be connected through the interface to input a desired character string.

Although the input interface 20 inputs characters and data in the form of code information, the input interface 20 may be configured to allow input in other forms. For example, an interface for inputting voice information and information on the shape of characters is provided, and the input information is converted into digital information that can be read by the CPU 22, and this is converted into a character string by voice recognition or character recognition. Can be considered.

Display controller (DC) 28
Controls the signal output to the liquid crystal display 10b as a display device. The serial input / output interface (SIO) 32 is connected to a public telephone line P
The computer 10c can be connected to an external network NW via the public telephone line PTL. Further, it is also possible to access a specific server SV and download necessary programs and data to the hard disk 10a.

In this embodiment, the external device 90 is constituted by a computer 90c comprising a CPU, ROM, RAM and the like, a hard disk 90a, a display 90b, a keyboard 90
It is a desktop personal computer equipped with d and the like. Therefore, the external device 90 is a device that can execute various functions by installing various application programs. For example, a document creation device that implements a word processing function, a data communication device that sends and receives character information through a telephone line, a translation device that translates a sentence created in one language into another language, and a character string that is the same as an input character string Or an information search device for searching for information having a similar character string, a character string search device for searching for a specific character string from a created document, or a desired file from a stored file, An abstract creation device or the like that creates an abstract by a computer can be considered. Of course, any device other than a personal computer can be the external device 90 as long as it has a function of outputting data information including a character string. In this embodiment, the external device 90 is also connected to the public telephone line PTL via a modem (not shown), similarly to the similarity search engine 10A.

Next, the contents of the character string similarity determination processing executed using such hardware will be described. FIG. 2 illustrates a case where the character string similarity determination process is performed.
FIG. 4 is an explanatory diagram showing a flow of information between a similarity search engine 10A and an external device 90. Similarity search engine 10A
Inputs a character string sent from the external device 90 and refers to a comparison character string to be compared with the input character string. In this embodiment, the information of the contrast character string is stored in the external device 9.
0 is stored in the hard disk 90a, so that the reference destination of the comparison character string is the hard disk 90a. Of course, if the comparison character string is stored in a readable manner, the hard disk 1 of the similarity search engine 10A
0a or the server SV can be referred to.

The similarity search engine 10A determines the similarity between the input character string and the comparison character string with reference to the synonym dictionary 36 in the hard disk 10a. This synonym dictionary 3
The content of No. 6 will be described later. Similarity search engine 10
A outputs the determination result of the similarity between the input character string and the comparison character string to the external device 90. Therefore, the external device 90 can use the result of the similarity determination between the character strings for the two character strings held by the external device 90.

Next, the processing procedure of the character string similarity determination processing will be described with reference to the character string similarity determination routine of FIG. The character string similarity determination routine starts when a signal indicating that a character string is input is received from the external device 90.
As shown in FIG. 3, when this routine is started, first,
The character string information sent from the external device 90 is input (step S100), and a process of referring to a comparison character string to be compared for similarity between the character strings is performed (step S120).

Next, for each of the input character string and the comparison character string, a process of extracting constituent units constituting each character string together with grammatical information is performed (step S130). In the present embodiment, an expression having a coherent meaning, for example, a conceptual expression representing a semantic concept or a framework expression supporting a framework of a sentence structure is extracted as one constituent unit. The structural unit as a conceptual expression includes not only a conceptual word but also a unit expressing a semantic concept by a combination of a plurality of words. The structural units as framework expressions include functional expressions such as relational expressions and advisory expressions.

The extraction process of the concept expression and the functional expression is as follows.
This is performed by referring to a Japanese language dictionary that stores grammatical information about this expression. The Japanese language dictionary stores, as grammatical information, whether each expression corresponds to a concept word, an affix, a relational expression, or an auxiliary expression. These grammatical information can refer to the reading as an index.

FIG. 4 shows the concept word and the functional expression.
This will be described with reference to FIG. The concept word refers to a word that expresses some semantic concept by itself, and is mainly an independent word. For example, the noun "bus", the verb "come",
The adjective "beautiful" etc. belongs to the category of concept words. On the other hand, a functional expression is an expression that does not express a semantic concept by itself, but acts to limit the semantic role of the concept represented by the concept word, and is attached to a concept word, such as a particle or auxiliary verb. In addition to words, affixes such as prefixes and suffixes, and ancillary expressions other than these, for example, words representing relational expressions and auxiliary expressions, etc., correspond to this. For example, the particle "ga" representing the subject, the particle "de" representing the place or means, the auxiliary verb "re" representing the passive, the auxiliary verb "like" representing the guesswork, the prefix "new", and the suffix " “Difficult” belongs to the category of functional expression.

The relational expression is an expression that is used between concept words in one sentence and expresses a relationship between concept words such as a case relationship and a causal relationship. In addition to "de", collocations indicating a cause or means such as "by" or collocations indicating a place such as "in" correspond to this. An auxiliary expression is an expression mainly used after a concept word, which is a predicate, to change the semantic content of the predicate itself. For example, collocations such as "may" and "must", and adverb "adidas" in which the adjunct form of the auxiliary verb "beshi" representing guesswork and the final form of the auxiliary verb "da" representing assertion are combined Words correspond to this.

FIG. 4 shows a concrete relationship between the concept word and the functional expression.
Shown in As shown in FIG. 4A, when a character string is composed of only concept words, the meaning content of each concept word is represented separately and independently, so the meaning content represented by the entire character string is Becomes ambiguous. For example, a character string "bus coming" composed of two concept words "bus" and "coming" is "coming bus", "bus coming", "coming by bus", "coming by bus". It is not clear which meaning is shown.

On the other hand, as shown in FIG. 4B, when there is a relational expression that is one of the functional expressions between two concept words, the relational expression has the meaning of the immediately preceding concept word. It works to identify the above work. For example, the concept word "bus", if followed by a functional expression "ga", would mean "bus as subject".
When the relational expression "by" is placed, it means "bus as a means of transportation", and when the relational expression "by" is placed, it is "bus as a means of transportation" or "bus as a place."".
That is, a concept word becomes a sentence structure only when it is linked to a functional expression such as a relational expression.

By specifying the meaning of the preceding concept word by the functional expression in this way, an organic combination of the concept from the preceding concept word to the following concept word occurs, and one meaning content as a whole sentence is obtained. Will be represented. For example, the sentence "The bus is coming" means that the bus as a whole comes and goes to where you are,
The sentence "coming by bus" as a whole means that "a certain person comes to the place where he / she exists by using the bus means of transportation".
In addition, they are themselves called “transportation” and “place”
The word "in the bus", which had two meanings,
The use of the concept word "coming" was subsequently limited to the meaning of "bus as a means of transportation", and as a result, the sentence "coming by bus" was very similar to the sentence "coming by bus". It represents the meaning.

Also, as shown in FIG. 4C, when a predicate, which is one of the functional expressions, is present after the concept word, which is a predicate, the predicate is the immediately preceding predicate. It works to change the meaning of concept words. For example, the concept word "coming", when followed by an advisory expression "may", indicates the inferred meaning of the act of "coming", and the advisory expression "should" When an expression is placed, it means that the act of "coming" is an obligation.

As described above, even if two sentences use the same concept word, the sentence may have a completely different meaning due to a difference in the functional expression used. It may be a statement to have. In other words, the functional expression plays a role that greatly affects the meaning of the sentence.

Returning to FIG. In step S130, constituent units constituting each character string are extracted from the input character string and the comparison character string into concept words, relational expressions, adjunct expressions, and affixes. An inter-word similarity determination process for determining (hereinafter, referred to as an inter-word similarity) is performed (step S140). In the following description, each constituent unit constituting a sentence is referred to as a “word” in a broad sense for convenience of description. Next, based on the determination result, a similarity between word strings constituting the input character string and a word string constituting the comparison character string (hereinafter, referred to as similarity between word strings) is determined. A similarity determination process is performed (step S150). Details of these processes will be described later. Next, a process of outputting a numerical value indicating the determined degree of similarity between word strings to the external device 90 as a determination result is performed (step S160), and this routine ends.

Next, the details of the inter-word similarity determination processing in step S140 of FIG. 3 will be described with reference to FIGS. FIGS. 5 and 6 are flowcharts showing an inter-word similarity determination routine. This routine is started when words constituting the input character string and the comparison character string are extracted from each character string together with grammatical information. Hereinafter, for the sake of convenience of explanation, the input word A
/ Word a2 / word a3 /.../ word ai /.../ word am
(The symbol / indicates a word delimiter, the letter i indicates the ordinal number of a word in the character string A, and the letter j indicates the ordinal number of the word in the character string B. The same applies hereinafter.) A description will be given assuming that a total of n words “word b1 / word b2 / word b3 /.../ word bj /.../ word bn” are extracted from the comparison character string B.

When this routine is started, first, the ordinal number i of the word for the input character string A is set to the value 1 (step S200), and the ordinal number j of the word for the comparison character string B is set to the value 1. (Step S210). As a result, the word similarity determination target is:
Word a1 and word b1 are specified.

Next, it is determined whether or not the words a1 and b1 are both affixes (step S220). If both are not affixes, a process is performed to determine whether or not both are conceptual words (step S220). Step S225). If neither is a concept word, it is determined whether or not both are relational expressions (step S230). If neither is a relational expression, a process is performed to determine whether or not both are adjunct expressions (step S230). Step S235). If the words a1 and b1 are either affixes, both concept words, both relational expressions, and both adjunct expressions, the synonym dictionary 36 is referred to for the word a1 (step S240), and the word b1 is used as a synonym for the word a1. A process for determining whether or not the information has been registered is performed (step S245).

The structure of the synonym dictionary 36 will be described with reference to FIGS. In the present embodiment, the synonym dictionary 36 includes the concept synonym dictionary 36a shown in FIG.
The conceptual expression synonym dictionary 36a includes a functional expression synonym dictionary 36b, and the functional expression synonym dictionary 36b stores information on a concept word and a word having a meaning similar to the concept word (hereinafter referred to as a concept similar word). , Information of a functional expression and a word having a similar meaning to the functional expression (hereinafter referred to as a function similar word). As shown in FIGS. 7 and 8, the concept word synonym dictionary 36a and the functional expression synonym dictionary 36b
Character information and part-of-speech information related to each concept word and each functional expression are stored in alphabetical order corresponding to the index, which is a search heading, and the similarity of concepts corresponding to each of these concept words and each functional expression is stored. Character information of words and functional similar words,
Stores part of speech information. Note that the synonym dictionary 36 may be a single dictionary in which information on concept words and functional expressions is put together, or separate dictionaries are stored for each type of stored words such as concept words, affixes, relational expressions, and functional expressions. The configuration may be provided.

In addition, the concept word synonym dictionary 36a and the functional expression synonym dictionary 36b provide an inter-word similarity degree indicating the degree of meaning similarity between the concept word and each concept similar word and the functional expression and each function similar word. Stores numerical data. In the present embodiment, the inter-word similarity is represented using a numerical value in the range of “0 ≦ G ≦ 1”, and it is defined that the closer the numerical value is to 1, the higher the degree of similarity in meaning is. For example, as for the concept word “bus” described above, information of the word “car” is registered as a concept similar word along with a relatively low inter-word similarity value of “0.3”. Information about the word "I" as a concept similar word,
It is registered with a high inter-word similarity value of “0.9”.

Note that words other than concept words are also registered as concept similar words in the concept word synonym dictionary 36a. For example, an affix used in the sentence in the same meaning and usage as the concept word (hereinafter, referred to as a concept word suffix) is also registered. This concept word suffix includes, for example, "new", "large",
There is a prefix of "real", which is used to modify the noun immediately following it, similar to the adjectives "new", "big", and "actual" (eg, "new plan" and "new" Planning, “great discovery” and “great discovery”, “true story” and “actual story”). Therefore, the concept word synonym dictionary 3
6a has conceptual similarities corresponding to conceptual words as adjectives such as "new", "big", and "actual".
Prefixes such as "new", "large", and "real" are registered. This is the same for the functional expression synonym dictionary 36b. Functional similar words corresponding to functional expressions as affixes such as “new”, “large”, and “real” include “new”, “large”, The concept word "actual" is registered.

In this connection, in step S220, even if both the words ai and bj are not affixes, if one is a concept word suffix and the other is a concept word, both are affixes. Assuming that there is, the synonym dictionary 36 is referred to. Of course, in step S225, if both the word ai and the word bj are not concept words, but one is a concept word and one is a concept word suffix, the words are regarded as concept words and the synonym dictionary 36 is referred to. It may be. Therefore, it is possible to obtain a highly appropriate word similarity between the concept word suffix and the concept word.

Referring back to FIG. If it is determined in step S245 that the word b1 is registered as a synonym for the word a1, the value of the inter-word similarity stored in the synonym dictionary 36 is set to the inter-word similarity between the word a1 and the word b1. A process of storing as degrees is performed (step S25).
0). In the present embodiment, the inter-word similarity between the word ai and the word bj is represented as t (ai, bj). Therefore, the word a1
The inter-word similarity between the word and the word b1 is represented as t (a1, b1).

If it is determined in step S245 that the word b1 is not registered as a synonym for the word a1, or in step S235, the word a1
If the word b1 and the word b1 are neither a concept word (except when one is a concept word suffix), both are affixes, both are relational expressions, and neither are adjunct expressions, the inter-word similarity value is 0 ( (Step S2)
55). The value of the word similarity is recorded in the word information recording table GT on the RAM 26.

Next, a process for obtaining a distance between words (hereinafter, referred to as an inter-word distance) is performed based on the value of the inter-word similarity set in steps S250 and S255 (step S260). In this embodiment, the inter-word distance between the word ai and the word bj is twice the complement of the value of the inter-word similarity between the words ai and bj, that is, 2 ｛1-t (ai, b
j) Expressed as｝. Therefore, the inter-word distance between word a1 and word b1 is expressed as 2 {1-t (a1, b1)}. The value of the inter-word distance is recorded in the inter-word information recording table GT on the RAM 26.

As a result, the inter-word distance is in a numerical range of “0 ≦ G ≦ 2”, and the closer the numerical value is to 0, the shorter the inter-word distance is. For example, in the case of the concept word “bus” and the word “car” having the above-mentioned word similarity value “0.3”, the word distance is “1.4” which is a value indicating a relatively long distance. In the case of the concept word “I” and the word “I” whose inter-word similarity is “0.9”, the inter-word distance is a value indicating a short distance of “0.2”.

Next, the value 1 is added to the ordinal number j of the word in the comparative character string B (step S265), and the ordinal number j is compared with the total number n of words in the comparative character string B. If the ordinal number j exceeds the total number n, The process returns to step S220 and repeats the above processing until it is determined (step S270). If it is determined that the ordinal number j has exceeded the total number n, it is determined that the similarity of the word a1 in the input character string A with each word in the contrast character string B has been determined, and the word of the input character string A Ordinal number i
Is added (step S275), and the ordinal number i is compared with the total word number m of the input character string A (step S280).
If the ordinal number j does not exceed the total number n, step S2
Returning to step 10, the above processing is repeated. If the ordinal number j exceeds the total number n, it is determined that the similarity between all the words in the input character string A and all the words in the comparison character string B has been determined, and this routine ends. .

The case where this inter-word similarity determination processing is actually performed for two character strings will be described. FIG.
Is obtained after the word similarity determination process is performed on the input character string A “About the guideline for concluding a new Japan-US defense agreement” and the comparison character string B “On the new Japan-US cooperation guidelines” 6 shows a state of an inter-word information recording table GT.

The input character string A is stored in step S130 in FIG.
By the word extraction process, the prefix "new", the noun "Japan and the United States", the sa noun "defense", the noun "agreement", the sa noun "conclusion", and the relational expression " For ", a noun" guideline ", a relational expression" about ", and a relational expression" ha "are divided into nine words in total. Hereinafter, each of these words is referred to as a word a
1, word a2, word a3, word a4, word a5, word a
6, description will be made as a word a7, a word a8, and a word a9. On the other hand, the comparison character string B is an adjective "new", a noun "Japan-US", a relational expression "no", a sa noun "cooperation", a noun "guideline", and a relational expression. It is divided into a total of six words “about”. Hereinafter, these words are referred to as word b1 and word b, respectively.
Description will be made as 2, word b3, word b4, word b5, and word b6.

The inter-word information recording table GT records the inter-word similarity and the inter-word distance for all combinations of these nine words and the six words. For example, a column for recording data between the word “new” in the input character string A and the word b1 “new” in the comparison character string B (in the table of FIG. 9, the value of i is 1 and j In the case where the value is 1, the value of “1.0” which is a registered value in the concept word synonym dictionary 36a is used as the value of the inter-word similarity, and the value of “2 × (1−1)” is used as the value of the inter-word distance. 1.0) ”, which is the calculated value of the calculation formula, is recorded.

On the other hand, a column for recording data between the word "a1" of "new" in the input character string A and the word "b2" of "Japan-US" in the comparison character string B (the value of i in the table of FIG. 1 and the value of j is 2), the words a1 and b2 are affixes and nouns, respectively, and the synonym dictionary 36 is not referred to. Is recorded. As a result, as the value of the word distance,
The highest value of “2.0”, which is the operation value of the calculation formula of “2 × (1−0.0)”, is recorded.

Next, the details of the inter-word string similarity determination processing in step S150 in FIG. 3 will be described with reference to FIGS. FIG. 10 is a flowchart showing the inter-word string similarity determination routine A. This routine is a routine for calculating a similarity value between word strings from a value of an inter-word distance of each word constituting a character string determined in the inter-word similarity determination processing in step S140. It starts when the processing ends.

When this routine is started, first, a dropout cost setting process for setting an added value of the distance when a word similar to a word in one word string does not exist in the other character string is performed (step S300). Next, a word string distance calculation process for calculating the distance between word strings is performed (step S320). Finally, a calculation process of calculating the similarity between the character strings using the value of the distance between the word strings is performed (step S340), and this routine is terminated, and the next determination result output processing (step S160 in FIG. 5) Move to). Hereinafter, the three steps of this routine will be described in detail as “dropout cost setting processing”, “inter-word string distance calculation processing”, and “inter-word string similarity calculation processing”, respectively.

First, the drop-out cost setting process in step S300 will be described with reference to FIG. 11 and FIG. FIG. 11 is a flowchart showing a dropout cost setting routine, and FIG. 12 shows a distance graph YG in which coordinate axes are set by dropout cost setting processing. Distance graph Y
G is a graph for distinguishing and indicating words that should be similar to each other and words that are not similar, on the premise of judging the similarity between character strings.
Are composed of words a1 to am, and a vertical axis represents a comparison character string B.
Are assigned to the words b1 to bm.

The drop-off cost setting routine shown in FIG.
It starts with the start of the inter-word string similarity determination routine A in FIG. Thereafter, the words a1 to a in the input character string A
Word string up to i and word b1 to word b in comparison string B
The distance from the word string up to j will be described as d (ai, bj).

When this routine is started, first, the values of the ordinal number i of the word in the input character string A and the ordinal number j of the word in the input character string B are set to 0 (step S400), and d is set. A process of setting the value of (a0, b0), that is, the value of the distance between the input character string A and the comparison character string B before the comparison between the input character string A and the comparison character string B as 0 (zero). Perform (step S410). With this processing, the starting point for measuring the distance between the character strings in the distance graph YG is determined as the origin O on the distance graph YG (see FIG. 12).

Next, after the value of the ordinal number i of the word a in the input character string A is set to 1 (step S420), the value of d (ai, bj) at this time is changed to d ｛a (i−1) ),
A process of setting the value of bj # to the value of the drop cost r of the word ai is performed (step S430). This processing is performed as follows: “If a word having a similar meaning to the word ai does not exist in the word string from the word b1 to the word bj, the word string up to the word immediately before the word ai and the word The distance r is added to the distance from the word string up to bj. " For example, if the value of ordinal i of word a is 1
In the case of, as the value of d (a1, b0), d (a0,
The drop cost r of the word a1 is set to 0 (zero) which is the value of b0).
Is added to the value of “r”. With this processing, the distance from the origin O (O) to the word a1 on the horizontal axis is set as “r” on the distance graph YG (see FIG. 12).

Next, 1 is added to the value of the ordinal number i of the word a in the input character string A (step S435), and whether or not the value of the ordinal number i exceeds the value of the total number m of words of the input character string A is determined. (Step S440), and the value of the ordinal number i is equal to the total number m.
Step S430 until it is determined that the value exceeds
And the above processing is repeated. For example, step S4
If the value of the ordinal i is set to 2 at 35, step S4
By the calculation processing of 30, the value of d (a2, b0) is set to a value of 2r obtained by adding the value r of the dropout cost to the value r of d (a1, b0) previously obtained. As a result, on the distance graph YG, the word a on the horizontal axis is shifted from the origin O (O).
The distance to 2 is set as “2r” (see FIG. 12).

By such repetitive processing, d (a
1, (b0) to d (am, b0) are set. As a result, each word on the horizontal axis of the distance graph YG is equally divided and assigned to the value of the drop cost r of each word of the input character string A (see FIG. 12).

If it is determined in step S440 that the value of the ordinal number i exceeds the value of the total number m, the value of the ordinal number i of the word in the input character string A is set to 0 (zero) and the input character string B
After setting the value of the ordinal number i of the word in it to 1 (step S450), d (ai, bj) at this time
Is set to the value obtained by adding the value of the drop cost q of the word bj to the value of d {ai, b (j-1)} (step S460). This processing is performed as follows: “If a word having a similar meaning to the word bj does not exist in the word string from the word a1 to the word ai, the word string up to the word immediately before the word bj and the word The distance q is added to the distance to the word string up to bj ”. For example, if the value of the ordinal j of the word b is 1, d (a0, b
As the value of 1), a value “q” is set, which is obtained by adding the value of the drop cost q of the word b1 to 0 (zero) which is the value of d (a0, b0). By this processing, the distance from the origin O (O) to the word b1 on the vertical axis is set as “q” on the distance graph YG (see FIG. 12).

Next, 1 is added to the value of the ordinal j of the word b in the comparison character string B (step S465), and whether or not the value of the ordinal j exceeds the value of the total number n of words in the comparison character string B is determined. Is determined (step S470), and the value of the ordinal j is equal to the total number n.
Step S460 until it is determined that the value exceeds
And the above processing is repeated. For example, step S4
If the value of the ordinal j is set to 2 at 65, step S4
By the calculation processing of 60, the value of d (a0, b2) is set to a value of 2q obtained by adding the value q of the dropout cost to the value q of d (a0, b1) obtained earlier. As a result, on the distance graph YG, the word b on the vertical axis is shifted from the origin O (O).
The distance to 2 is set as “2q” (see FIG. 12).

By such a repetitive processing, d (a
0, b0) to d (a0, bn) are set. As a result, each word on the vertical axis of the distance graph YG is equally divided and assigned to the value of the drop cost q of each word of the contrast character string B (see FIG. 12).

If it is determined in step S470 that the value of the ordinal j exceeds the value of the total number n, the routine ends, and the next inter-word string distance calculation processing (step S3 in FIG. 10)
Move to 20). As a result, positions where the words a1 to am and the words b1 to bn are allocated to the horizontal axis and the vertical axis on the distance graph YG are determined.

In this embodiment, the values of the drop costs r and q are set to “1”. However, according to the frequency and importance of the drop of the word in the compared sentences, it is assumed that other numerical values are used. No problem.

The way of expressing the distance between each word string on the distance graph YG to which each word is allocated as described above will be described with reference to FIG. If the input character string A is composed of m words in total and the comparison character string B is composed of n words in total, the length from the origin O as the starting point to the coordinate point (am, bn) as the ending point Is the distance between the character strings of the input character string A and the comparison character string B. The case where the distance is the shortest is indicated by a white arrow, and the case where the distance is the longest is indicated by a hatched arrow. This "when the distance is the longest"
This is a case where no word similar to each word constituting the input character string A exists in the comparison character string B at all. In this case, each word from word a1 to word am and word b1
Therefore, the distance between the input character string A and the contrast character string B is represented by the value of “dropping cost r × m + dropping cost q × n”. It becomes a certain “mr + qn”.

Next, the inter-word string distance calculation processing A in step S320 will be described with reference to FIGS. In this inter-word string distance calculation processing A, each of the input character strings A constituting the input character string A is used by using information about each word such as the value of the inter-word distance between each word and the value of the drop cost r, q of each word. The distance between the word string and each of the word strings forming the comparison character string B is determined. This processing procedure is shown in the word string distance calculation routine A in FIG. This routine starts when the dropout cost setting process ends.

When this routine is started, first, the value of the ordinal number i of the word in the input character string A is set to 1 (step S500), and the value of the ordinal number j of the word in the input character string B is set to 1 (Step S510).

Next, d (ai, bj) is the distance between the word string from the word a1 to the word ai in the input character string A and the word string from the word b1 to the word bj in the comparison character string B.
Find the value of This value is obtained in the following manner. First, processing for obtaining three values X1, X2, and X3 is performed (steps S515, S520, and S525). Value X1
Is the distance d ｛a (i-
1), b (j-1)} and the value of 2 {1-t (ai, bj)}, which is the inter-word distance between the word ai and the word bj (hereinafter, this value is referred to as X1 value).
The value X2 is the word string up to the word immediately before the word ai and the word b
d ｛a (i−1), which is the distance from the word string up to j
It is obtained by adding the value of the drop cost r of the word ai to the value of bjｂ (this value is hereinafter referred to as X2 value). Value X
3 is the distance between the word string up to the word ai and the word string up to the word immediately before the word bj, d ｛ai, b (j−
1) It is obtained by adding the value of the drop cost q of the word bj to the value of｝ (hereinafter, this value is referred to as X3 value)

Next, the smallest value among the values from X1 to X3 is defined as d (ai, bj) and R
It is stored in the word string distance recording table DL on the AM 26 (steps S535 and S540), and this value is stored in the input character string A from the word string from the word a1 to the word ai and the word string from the word b1 in the comparison character string B It is determined as the distance from the word string up to bj. That is, when calculating the distance between the word string from the word a1 to the word ai and the word string from the word b1 to the word bj in the comparison character string B, the distance from the word a1 to the "word immediately before the word ai" Since only the distance or the distance from the word b1 to the “word immediately before the word bj” is considered, two or more sets of similarity between words between the input character string A and the comparison character string B intersect. Inevitably banned.

Next, 1 is added to the value of the ordinal number j of the word b in the comparison character string B (step S550). Is determined (step S560), and the value of the ordinal j is equal to the total number n.
Step S515 until it is determined that the value exceeds
And the above processing is repeated. Thus, the distance between the word string up to the word a1 in the input character string A and each word string from the word b1 to the word bj in the comparison character string B is sequentially obtained.

If it is determined in step S560 that the value of the ordinal j exceeds the value of the total number n, 1 is added to the value of the ordinal i of the word in the input character string A (step S57).
0), until it is determined that the value of the ordinal number i exceeds the value of the total number m of words in the input character string A, step S51.
Returning to 0, the above processing is repeated. Thereby, the distance between each word string from the word a1 to the word ai in the input character string A and each word string from the word b1 to the word bj in the comparison character string B is sequentially obtained. Word am
The value of the distance d (am, bn) between the word string up to and the word string from the word b1 to the word bn is determined.

That is, the distance d (ai, bj) between the word strings
Are all words a1 to a
m, b1 to bn, and at this time, the words a1 to am, b1 to bn set in step S250 in FIG.
The value of the inter-word similarity t (ai, bj) for is used. For example, d (a1, b2), which is the distance between the word string from word a1 to the word string from word b1 to word b2, and the distance between the word string from word a1 to word a2, and the word string from word b1 When d (a2, b2), which is the distance between a word string from word a1 to word a2 and a word string from word b1 to word b2, is determined by d (a2, b1). The value of the inter-word similarity t (a1, b1) is used. Also, d (a7, b3) which is the distance between the word string from the word a1 to the word a7 and the word string from the word b1 to the word b3, the word string from the word a1 to the word a6, and the word string from the word b1 to the word b4 D which is the distance from the word string
(A6, b4), d which is the distance between the word string from word a1 to word a7 and the word string from word b1 to word b4
When calculating (a7, b4), the value of the inter-word similarity t (a6, b3) between the words a6 and b3, which are words having different appearance orders in the character string, is used.

If it is determined in step S570 that the value of the ordinal i exceeds the value of the total number m, the distance d () between the word string forming the input character string A and the word string forming the comparison character string B is determined. am, bn), the routine ends, and the process proceeds to the next character string similarity calculation process (step S340 in FIG. 10).

The contents of the above-described inter-word string distance calculation processing will be described with reference to a specific example using the distance graph YG. Here, the word string distance calculation processing is performed by using the input character string A of “new / Japan / US / defense / agreement / conclusion / for / guideline / about /” and “new / Japan / US / cooperation” / Guideline / "will be described as an example. As described above, the input character string A is composed of nine words from the word a1 to the word a9,
The comparison character string B is a word string including six words from the word b1 to the word b6.

For these two word strings, step S5
When both the value of the ordinal i and the value of the ordinal j are set to 1 by the processing of 00 and step S510, the value of d (a1, b1) is obtained by the arithmetic processing from step S515 to step S535. . FIG. 15 shows the state of the operation buffer ET which temporarily stores the process and the result of this operation. That is, 0 (zero) is obtained as the X1 value by the operation processing of step S515, 2 is obtained as the X2 value by the operation processing of step S520, and 2 is obtained as the X3 value by the operation processing of step S525.
In the operation buffer ET, as the value of d (a1, b1),
0 (zero) of the X1 value which is the minimum value among them is stored.

The value of 0 (zero) of d (a1, b1) is the distance between the word string composed of the word “new” and the word string composed of the word “new” b1. This will be described with reference to the distance graph YG in FIG.

On the distance graph YG, the distance between the word string consisting of the word “new” a1 and the word string consisting of the word “new” b1 is represented by coordinates (a1, b)
It is represented as a reaching route to 1). As shown in FIG. 16, this arrival path includes a first path from the origin O directly to the coordinates (a1, b1), and a first path to the coordinates (a1, b1) via the coordinates (0, b1). The third route which reaches the coordinates (a1, b1) via the second route, the coordinates (0, b1)
There is a route. X1 which is the calculation result of step S515
The value indicates the value of the distance required when passing through the first of the three routes, and is the X2 value that is the calculation result of step S520 and the X2 value that is the calculation result of step S525.
The three values indicate the value of the distance required when passing through the second route and the third route, respectively.

The value of the distance when the coordinates (a1, b1) reach the coordinates (a1, b1) through the first route is 0 (zero), and the value of this distance is determined by passing through the second route and the third route. Is smaller than 2 (= r + q), which is the value of the distance required when Therefore, 0 (zero) of the X1 value when passing through the first route is the value of the distance d (a1, b1) between the word string including the word a1 and the word string including the word b1. The path at which this value is reached is indicated by the hatched arrow in FIG.

Next, a description will be given of the processing contents when the value of the ordinal j is set to 2 by the processing of step S550. By the arithmetic processing from step S515 to step S525, the value 3 is set as the X1 value and the value 3 is set as the X2 value.
Is obtained as the X3 value, and the value 1 of the X3 value, which is the minimum value, is calculated as d (a1, b
The value of 2) is temporarily stored in the calculation buffer ET. This is shown in FIG.

The value of 0 (zero) of d (a1, b1) is determined by the distance between the word string consisting of the word "new" and the word string consisting of the words "new / US" b1 and b2. Become. This will be described with reference to the distance graph YG in FIG.

As shown in FIG. 18, the coordinates (a
The first path from the origin O to the coordinates (a1, b2) via the coordinates (0, b1) includes a reaching path leading to the coordinates (1, b2).
The second route and the coordinates (a) to the coordinates (a1, b2) via the route, the coordinates (0, b1) and the coordinates (0, b2)
There is a third route to the coordinates (a1, b2) via (1, b1). The X1 value that is the calculation result of step S515 indicates the value of the distance required when passing through the first of the three paths, and is the X2 value that is the calculation result of step S520 and the calculation result of step S525. X3
The value indicates the value of the distance required when passing through the second route and the third route, respectively.

That is, the case of passing through the first route means that the word b1 of "new" in the word string of "new / Japan / US" drops from the word string of the word a1 of "new". On the other hand, it means that the word b2 of "Japan and the United States" is regarded as having a similar relationship with the word a1 of "new", and the distance between both word strings is measured. In the case of passing through the second route, the words b1 and b2 of "new" and "Japan-US" are words that are both dropped from the word string consisting of the word a1 of "new", and b2 The word "new" a1 is also
This means that the distance between the word strings is measured assuming that the word is a word dropped from the word string composed of the words "new / Japan / US" b1 and b2. The case of passing through the third route is a word in which the word “new” a1 and the word “new” b1 are similar to each other. The word “b2” of “Japan-US” is a word string composed of the word a1. This means that the distance between word strings is measured assuming that the word is dropped from the word string.

The value of the distance when the coordinates (a1, b2) reach the coordinates (a1, b2) through the third route is 1, and the value of the distance is the value of the distance required when the vehicle passes through the first route. Is smaller than 3 (= q + 2) or 3 (= 2q + r), which is the value of the distance required when passing through the second route. Therefore, the X3 value of 1 when passing through the third route is the value of the distance d (a1, b2) between the word string consisting of the word a1 and the word string consisting of the words b1 and b2. The route at which this value is reached is indicated by the hatched arrow in FIG.

By repeating such arithmetic processing,
A word string consisting of the word "new" a1, a word string consisting of the word "new" b1, a word string consisting of the words b1 and b2 "new / Japan / US", and a word "new / Japan / US / no" A word string consisting of words b1 to b3, a word string consisting of words b1 to b4 of "new / Japan / US / cooperation", a word string of words b1 to b5 of "new / Japan / US / cooperation / guidelines" And the distance between the word string consisting of the word b1 and the word string consisting of the words b6 to "new / Japan / US / cooperation / guidelines /". FIG. 19 shows how the distance values in these cases are stored in the inter-word string distance recording table DL.

After the distance between the word string composed of the word “new” a1 and each word string in the comparison character string B is stored, the processing in steps S570, S580, and S510 is performed to call “new / Japan-US”. Word a1 and word a2
Then, the distance between the word string composed of and the respective word strings in the comparison character string B is calculated. First, the value of the ordinal i is set to 2 and the value of the ordinal j is set to 1.
The distance between the word string consisting of the word string and the word string consisting of the word “new” b1 is calculated.

In this case, the value 1 that is the X2 value, that is, the word “new” and the word “new” b1
Are similar words to each other, and the word b1
Is a word dropped from the word sequence consisting of the word b1.
Is the distance between both word strings d
(A2, b1).

Subsequently, by the processing in step S550,
If the value of the ordinal j is 2, the distance between the word string consisting of the words “new / US” and the word a2 and the word string consisting of the words “new / Japan and the United States” b1 and b2 is It is calculated. FIG. 20 shows the progress of this arithmetic processing and the state of the arithmetic buffer ET storing the result, and FIG.
Are shown below.

The route on the distance graph when the X1 value is adopted is the distance d (a1, b) obtained in FIGS.
While adopting the value 0 (zero) of 1), the word a2 of “Japan / US” in the word string of “new / Japan / US” and “new / US /
The word b2 "Japan-US" in the word string consisting of "Japan-US"
Is a path when it is considered that the relation is similar. on the other hand,
The route on the distance graph when the X2 value is adopted is shown in FIG.
7, the value 1 of the distance d (a1, b2) obtained in FIG.
X3 is a path when it is considered that the word a2 of “Japan / US” in the word string of “New / Japan / US” is omitted from the word string of “New / Japan / US”.
The route on the distance graph when the value is adopted is referred to as “Japan / US” in the word string “new / Japan / US” while adopting the value 1 of the previously obtained distance d (a2, b1). Word b
2 is a path in the case where it is considered that the number 2 is dropped from the word string of “new / US / US”.

As shown by the hatched arrow in FIG. 21, the distance required to reach the coordinates (a2, b2) is X1
The value becomes the shortest when the value is adopted.
(Zero) is a distance d (a) between the word string composed of the words a1 and a2 and the word string composed of the words b1 and b2.
2, b2).

By repeating the above processing, finally, the distances between all the word strings relating to the input character string A and the comparison character string B are obtained. FIG. 22 shows a state where all the obtained values of the distance d (ai, bj) are stored in the inter-word string distance recording table DL. In addition,
The expression in parentheses following the distance value indicates the arithmetic expression that calculated this value, and the “*” mark indicates that the same value was calculated by two or more arithmetic expressions and both values were the minimum value. . FIG.
D (a9, b6) on the inter-word string distance recording table DL
Is a value of the distance between the word string including all the words a1 to a9 in the input character string A and the word string including all the words b1 to a6 in the comparison character string B, that is, , The distance between the character strings of the input character string A and the comparison character string B.

Numerical value 5.8 which is the value of d (a9, b6)
Is the distance between the character strings of the input character string A and the comparison character string B, and the distance graph YG in FIG. 23 and FIG.
This will be described more specifically with reference to FIG. As shown in FIG.
At each coordinate on the distance graph YG, the value of the distance between all the word strings recorded in the inter-word string distance recording table DL of FIG. 22 is described corresponding to the combination of the compared word strings. .

For example, the value “1” described at the coordinates (a2, b1) is the value of the distance d (a2, b1) between the word string “new / US” and the word string “new”. is there. Also, the coordinates (a9, b6) of the intersection of a9 where the last word “ha” of the input character string A is located and b6 where the last word “about” of the comparison character string B is located (hereinafter referred to as the end point coordinates) ) Describes a numerical value 5.8, which is the value of d (a9, b6).

"The distance between the character strings of the input character string A and the comparison character string B is a value of 5.8" means that "from the origin O, which is the starting point, follows any one of the paths and the end point coordinates ( a
In order to head to 9, b6), one has to travel a distance corresponding to the numerical value 5.8. "
One of these paths is shown in FIG. 24 using the path of the arrow. The distance required to move from the start point to the end point is referred to as “overall travel distance”, and the distance required to travel from one coordinate to the next coordinate is referred to as “section travel distance”. "

In FIG. 24, a horizontal arrow indicates a case where the word ai located at the coordinates at the end of the arrow is considered to have been dropped and the X2 value is the value of the distance between word strings. The arrow indicates that the word bj located at the coordinates of the end of the arrow is considered to be dropped, and the X3 value is the value of the distance between word strings. The word ai and the word bj are evaluated as being similar to each other, and the X1 value is the value of the distance between the word strings, respectively. For example, the distance value 4.8 described at (a8, b6), which is one of the coordinates of the end of the oblique arrow, is a value based on an X1 value, that is, an arithmetic expression using the word similarity as an element. (See FIG. 22).

That is, the path from the start point to the end point coordinates shown in FIG. 24 includes the word “new” a1 and the word “new” b1, the word “Japan and the United States” a2 and the word “Japan and the United States” b2, “ Word a6 for "for" and word b3 for "no", word a7 for "guideline" and "guideline"
The word b5 and the word a8 of "about" and the word b6 of "about" are evaluated as mutually similar words, and the word a3 of "defense" a3, the word a5 of "agreement", the word a5 of "conclusion" , And the word a9 similar to the word a9 are evaluated as not existing in the comparison character string B, and the word b4 of "cooperation" is evaluated as not existing in the input character string A. .
FIG. 25 shows the relationship between words represented by paths indicating such paths.

As shown in FIGS. 24 and 25, in this route, between each coordinate from the start point to the end point,
"0 (zero)", "0 (zero)",
"1", "1", "1", "0.6", "1",
The values “0”, “0”, and “1” are consumed, and the total value of the consumed section movement distances is a numerical value 5.8, which is the value of the entire movement distance.

Note that there is a route other than the route shown in FIG. 24 to reach the end point coordinates. It is also possible to narrow down.

The details of the word string distance calculation processing A have been described above. Next, the character string similarity calculation processing A (step S34 in FIG. 10), which is started upon completion of this processing.
The content of 0) will be described with reference to a character string similarity calculation routine A in FIG. First, a value obtained by adding the dropping cost r for the total number m of words in the input character string A to the dropping cost q for the total number n of words b in the comparison character string B, that is, the input character string A and the comparison character The maximum value of the distance from column B is determined, and this value is set to U. Next, the distance calculated between word strings was calculated.
Distance d (am, bn) between word strings including all words
Is divided by the value of U, and this value is defined as V. Next, the complement of the value of V is obtained, and this value is compared with the input character string A and the contrast character string B.
This routine ends as the degree of similarity s (am, bn) between the character strings, and proceeds to the next processing.

In the above example, the total value of the drop costs r and q for 15 words is 15, and the value of the distance d (a9, b6) between word strings including all words is 5.8. Therefore, the similarity between the character strings has a value of 0.61. It is determined that the closer the value is to 1, the higher the semantic similarity between the character strings. The similarity between the character strings determined using the numerical values is output as the determination result (FIG. 3).
Step S160), the similarity determination process between character strings is terminated.

The inter-sentence similarity determining apparatus 1A of the first embodiment described above refers to the synonym dictionary 36 for the similarity between the concept words and the functional The similarity between expressions is determined, and using the value of the similarity, the similarity between character strings is determined in consideration of the correspondence in word order. Accordingly, it is possible to accurately determine the degree of similarity between two sentences having different word orders.

Further, since the similarity between character strings is determined using the value of the similarity between the relational expressions and the adjunct expressions, the relational expressions and the auxiliary expressions expressing concept words having the same semantic concept and mutually different frameworks are obtained. The similarity determination can be accurately performed even between sentences composed of predicate expressions.

Further, in this embodiment, it is forbidden that two or more sets of similarity between words intersect between character strings while considering the correspondence of word order between sentences. Therefore, the processing involved in determining the similarity between character strings can be simplified.

Further, when the similarity between word strings is determined, the drop costs r and q are used to determine whether a word similar to one word exists in another word string or not. Assuming that the value in the case where the similarity is high is determined as the similarity between the word strings. Therefore, the similarity between two sentences can be correctly determined as a whole sentence.

In the similarity judgment processing between character strings described above, when a word similar to a word in one character string does not exist in the other character string, all words are recognized regardless of the type of the word. The distance between word strings was calculated using the same dropout costs r and q. On the other hand, it is also possible to adopt a configuration in which a different value of the cost of dropping is set in accordance with the type and importance of the word considered to be dropped. Hereinafter, such a configuration will be described with reference to FIGS.

FIG. 27 shows an inter-word string similarity determination routine B.
It is a flowchart which shows. In this routine, FIG.
Of the word string similarity routine A, the routine performs a word importance setting process (step S6).
00) is different from the inter-word string similarity routine A in which this is not performed. Corresponding to this difference, the next processing, word string distance calculation processing B (step S620) and character string similarity calculation processing B are also related to the values of the drop-out costs r and q used in the calculation. The processing differs from the processing A for calculating the distance between word strings in FIG. 14 and the processing A for calculating the similarity between character strings in FIG.

The procedure and contents of the word importance setting process are shown in a word importance setting routine of FIG. In this routine, almost the same process as in the dropout cost setting routine of FIG. 11 is performed.
And the same number.

In this routine, for the input character string A, the value of d (ai, 0) when the ordinal number i of the word ai is 1 to the total number m is changed to the value of d {a (i-1), 0}. A process is performed to set the value of ai to the value obtained by adding the value w (ai) (step S730). This processing is performed as follows: “If a word having a similar meaning to the word ai does not exist in the word string from the word b1 to the word bj, the word string up to the word immediately before the word ai and the word b1 to the word Add importance value w (ai) of word ai to the distance from word string to bj "
Means that

For example, when the value of the ordinal i of the word a is 1, the value of d (a0, b0) is set to 0 (zero) as the value of d (a1, b0), and the importance of the word ai is set to 0 (zero). Value w (ai)
Is added to the value “w (ai)”. By this process, as shown in FIG. 29, on the distance graph YG, the distance from the origin O (O) to the word a1 on the horizontal axis is set as "w (ai)".

Further, regarding the comparison character string B, the word bj
The value of d (0, bj) from ordinal j of 1 to the total number n is
The value of the importance level w of the word bj is added to the value of d {0, b (j-1)}.
A process is performed to set the value to which (bj) has been added (step S765). Therefore, the distance graph Y
On G, the distance from the origin O (O) to the word b1 on the vertical axis is set as “w (bj)” (see FIG. 29).

Steps S730 and S765
(Steps S740 and S7)
70), values from d (a1, b0) to d (am, b0) and values from d (a0, b0) to d (a0, bn) are set. As a result, the words a1 to am and b1 to bn on the horizontal axis and the vertical axis of the distance graph YG are, as shown in FIG. 29, the importance values w (ai) and w (bj) of each word.
Are allocated at intervals according to

The importance value w (ai), w (b) of each word
j) is stored in the above-mentioned national language dictionary, and is extracted together with the character information and grammatical information of the word when the word is extracted from each character string in step S130 of FIG. FIG. 30 shows how the importance values w (ai) and w (bj) of each word are stored in the Japanese language dictionary.

In this embodiment, the importance of a word is set to “0 ≦ w
(Ai), w (bj) ≦ 1 ”, and is defined as the degree to which the sentiment is influenced more as the numerical value approaches 1. Further, the importance of a concept word such as a noun is set to a value larger than the importance of a functional expression such as a relational expression, and the importance is different depending on the type of the functional expression. For example, the main particle "ha", which is connected after the particle, and the case particle "no", which is located between noun and noun and indicates the same status, do not change the meaning of the sentence even if omitted, The importance value is lower than other functional expressions.

When the word importance setting values w (ai) and w (bj) to be added when a word is lost are set and the word importance setting routine is completed, the inter-word sequence shown in FIG. The distance calculation routine B is started. In this routine, almost the same processes as those in the inter-word string distance calculation routine A of FIG. 14 are performed, and therefore, the last two digits of the corresponding step numbers are set to the same numbers as in FIG.

In this routine, the smallest value among the X1, X2, and X3 values is determined as the distance between word strings, almost in the same manner as the distance calculation routine between word strings A.
In the process of calculating the value and the X3 value, the word importance values w (ai) and w (bj) are used instead of the dropout cost r used in the word string distance calculation routine A (steps S820 and S825). ). Therefore, the distance d (a
Which of the three values is the minimum value determined as i, bj) is determined by the word string distance calculation routine A.
Will differ from the results in

FIG. 32 shows the similarity between the word strings of the input character string A described above “On the guideline for concluding a new Japan-US defense agreement” and the comparison character string B described in “Regarding the new Japan-US cooperation guidelines”. 7 shows a state of the inter-word string distance recording table DL after the determination processing B is performed. FIG.
Is compared with the result of the word string distance calculation routine A shown in FIG.
The word a8 of "about" and the word a of "ha"
9 and the word b3 of “no” in the comparison character string B, and the word b6 of “about”, when words similar to these are considered to be absent in the other character string,
A value smaller than 1 set as the dropout cost r, q is added to the distance of the word string to the immediately preceding word. Therefore, the calculated values X2 and X3 when these words are missing are likely to be the minimum of the three values. As a result, the X2 and X3 values are smaller than those in the table of FIG. The value is the distance d (ai, b) between word strings.
j) increases, the value of the distance between the word string including all the words a1 to a9 in the input character string A and the word string including all the words b1 to a6 in the comparison character string B, That is, the distance d between the character strings of the input character string A and the comparison character string B
The value of (a9, b6) is also a smaller value of 3.6.

FIG. 33 shows a state of a distance graph YW in which the values of the distances between all the word strings recorded in the inter-word string distance recording table DL are described. The absence of the word b3 “no” meaning the movement from the coordinates (a3, b2) to the coordinates (a3, b3), or the coordinates (a9, b) from the coordinates (a8, b6)
The “section moving distance” due to the lack of the word “a” indicating “moving” to 6) is smaller, and it can be seen that this contributes to the reduction in the overall moving distance.

FIG. 34 shows the relationship between the words indicated by the paths indicated by the arrows on the distance graph YW. As compared with FIG. 25 showing the case where the dropout costs r and q are uniform as described above, FIG.
In the case of No. 4, "Agreement" which was regarded as a dropout in FIG.
The word a4 and the word b6 of “for” which are considered to have a similarity to each other while the word a4 and the word b4 to “cooperation” are considered to have a similarity to each other.
And the word b3 of “no” are considered to be missing from the other character string, respectively. In the former, the added value when the word b3 of “no” is dropped is from value 1 to value 0.2.
In the value of d (a4, b4)
This is due to the fact that one value has become the smallest value. The latter consists of the words a6 for "for" and "no"
When the word b3 is dropped, the added value is reduced from the value 1 to the value 0.4 and the value 1 from the value 0.2, so that the value of d (a6, b3) is the smallest value of the X3 value. (See FIG. 32).

The contents of the character string similarity calculation processing B, which is performed after the end of the above-described word string distance calculation processing B, are shown in a character string similarity calculation routine B of FIG. In the inter-character string similarity calculation routine A shown in FIG. 26, the maximum value of the distance between the input character string A and the comparison character string B is compared with the drop-out costs r and q for m words in the input character string A. String B
A value obtained by adding a drop-out cost q for the total number n of words b in the list, and the distance d between word strings including all words
By dividing the value of (am, bn), the distance d (am, am) actually obtained with respect to the longest distance assumed between the character strings
bn) was calculated. In this regard, in the case of this configuration in which the cost in the case of a word drop is set according to the importance of the word, the maximum value of the distance between the input character string A and the contrast character string B is determined by the input character string and the contrast character string. Is the sum of the importance values w (ai) and w (bj) for the individual words constituting Therefore, this value is set to U (step S121).
0), and calculating the one's complement of the value obtained by dividing the value of d (am, bn) by this value (step S123)
0, S1250), the similarity s (am, bn) between the character strings of the input character string A and the comparison character string B is obtained.

In the above-described example, the total value of the importance values w (ai) and w (bj) for each of the 15 words is 15, and the distance d (a9) between the word strings including all the words. , B
Since the value of 6) is 3.6, the degree of similarity between character strings is 0.76. Therefore, the input character string A and the comparison character string B are different from the case where the drop-out costs r and q are set uniformly.
It is determined that the semantic similarity between the character strings is higher.

As described above, by adopting a configuration in which a different value is set for the dropout cost according to the type and importance of a word, an appropriate inter-sentence can be set in accordance with the nature of the sentence to be compared. Can be determined. For example,
When judging the similarity of the meanings of the Japanese character strings, the meaning of the sentence is greatly different due to the difference of the concept word, so it is desirable to increase the cost of dropping the concept word. On the other hand, when the external device 90 is a translation device that translates between Japanese and English, the structure of the sentence greatly affects the meaning of the sentence, so that accurate translation can be ensured by increasing the cost of dropping the functional expression. .

In the above-described character string similarity determination processing, similarity between character strings is determined by focusing on the presence or absence or similarity of a functional expression in an input character string and a comparison character string.
This processing can be applied to the determination of similarity regarding a character string including a compound expression. Here, the compound expression is
It refers to an expression that represents one unified semantic concept by combining two or more words. For example, expressions such as “analysis method” and “new product” correspond to compound expressions. Hereinafter, this application example will be described.

First, a character string composed of a compound expression "analysis method" is input to the similarity search engine 10A as an input character string, and a character string not containing the compound expression of "analysis method" is input as a contrast character string. The following describes the case where The concept word synonym dictionary 36a includes information indicating that the concept word “analysis” and the concept word “analysis” are similar, together with the inter-word similarity value of 0.7, the concept word “method” and the “method”. Is similar to the concept word "
Each is stored together with the value 0.7 of the inter-word similarity.

The similarity between word strings is determined based on the value of the inter-word similarity (step S150 in FIG. 3). In this application example, the continuation of the words constituting the character string is converted into a complex expression. If applicable, the input character string and the contrast character string are taken into account, taking into account the similarity between words and the similarity between the compound expression in the input character string and the expression corresponding to the compound expression in the contrast character string. Is determined. For example, in the case of the above example, the similarity between expressions is determined not only for the correspondence between "analysis-analysis" and "method-method" but also for the correspondence between "analysis method-analysis method".

Next, the grammar information stored in the Japanese language dictionary is referred to for the word “no” existing only in the comparison character string. The Japanese language dictionary stores, as grammatical information, the fact that the type of the word "no" is "a particle representing an equivalence" and that it can be omitted when used between predetermined nouns. . By obtaining such information, it is determined that the similarity between the expressions “analysis method-analysis method” is the same as the similarity between the expressions “analysis method-analysis method”. In this case, the omission cost q is not added because a word similar to the word “no” does not exist in the input character string. Therefore, the similarity between the word string “analysis method” and the word string “analysis method”
The value is the same as the similarity between the word string “analysis method” and the word string “analysis method”.

Note that the words omitted in the compound expression are not limited to functional expressions such as “no” and “for”, but are referred to as “analysis method” and “analysis method”, for example. A part of the stem or inflectional ending of the sa variant used in connection with a certain noun such as the word "to" in the case, or an adjective such as "shi" in the case of "new product" and "new product" May be a part of the stem or the ending of the word.

With such a configuration, the similarity between a character string expressed without omitting a part of the stem of a particle or a declinable word or an inflected ending and a compound expression omitting these and synonymously. Can be accurately determined.

It should be noted that the input character string and the comparison character string may have expressions other than “analysis method” and “analysis method”. For example, character strings “material analysis method” and “material analysis method” may be used. Further, as in the case where the comparison character string is the “analysis method”, the case where the concept word forming the input character string matches the concept word forming the comparison character string may be used.

The above configuration can be applied to a case where the expression in the contrast character string corresponding to the compound expression of the input character string is another compound expression. For example, when a composite expression “analysis method” is input as an input character string and a composite expression “analysis method” is input as a contrast character string, the corresponding relationships “analysis-analysis” and “method-method” are input. In addition, the similarity between the expressions “analysis method-analysis method” is determined, and the value of the inter-word similarity 0.7 between “analysis-analysis” and the inter-word similarity between the “method-method” are determined. Value 0.7
It is also desirable to add a predetermined value as a value of the degree of similarity between expressions to determine that "the two character strings are similar to each other more than the degree to which the words are similar".

In the above description, the case where a phrase, that is, a segment of a word is input as a character string has been described as an example. However, the input character string may be a character string containing a compound expression. The same effect as described above can be obtained even when a complete expression provided is input as a character string. In this way, the similarity between the compound expression used in one sentence and the expression in the other sentence corresponding to this compound expression is
Since the determination can be made accurately, the similarity of the entire sentence becomes more accurate.

Next, a second embodiment of the present invention will be described. The second embodiment relates to the data search device 1B, and is realized by the same configuration as the hardware configuration of the first embodiment. FIG. 36 shows an outline of the data search device 1B.
It was shown to. The data search device 1B includes a search engine 10B that performs a search while determining the similarity of a character string, and an external device 90 that inputs a character string. The internal configurations of the search engine 10B and the external device 90 are the same as in the first embodiment.

Compared with the inter-sentence similarity determination apparatus 1A of the first embodiment, the similarity determination engine 10A of the first embodiment determines the similarity between two given sentences and outputs the same. In contrast, the search engine 10B of the second embodiment inputs a sentence of a natural language (hereinafter referred to as a search key sentence) provided from the external device 90, detects a sentence having a high similarity to this sentence, and The difference is that it has a function of outputting this.
A plurality of search target sentences (hereinafter, referred to as data) to be searched are data stored in the hard disk 90a of the external device 90.

FIG. 37 shows an outline of the processing in the second embodiment. When this routine is started, the search engine 10B first performs a process of inputting a character string constituting a search key sentence to be searched (step S900).
This processing (Step S900) is the same as the character string input processing (Step S100 in FIG. 3) in the first embodiment. After inputting the search key sentence in this way, a process of extracting one search target sentence from the data stored in the external device 90 is performed (step S930). Next, the similarity between one extracted retrieval target sentence and the retrieval key sentence is determined, and a process of detecting a similar character string is performed (step S94).
0). This process is the same as the inter-word similarity determination process (FIG. 3, step S140) and the inter-word string similarity determination process (FIG. 3, step S150) of the first embodiment. That is, the distance between word strings constituting each character string is obtained from the similarity between conceptual words and the similarity between functional expressions, and the similarity between character strings is calculated from the distance between word strings including all words. The similarity between the distance search key sentence and one extracted search target sentence is calculated as a numerical value. The calculated similarity numerical data is temporarily stored in a predetermined area on the RAM 26.

After the above processing, it is determined whether or not the sentence to be searched still remains in the hard disk 90a (step S950).
And the process is executed again from the process of extracting the search target sentence.
When there is no more search target sentence for determining the similarity, reference is made to a predetermined area on the RAM 26 in which numerical data of the similarity with each search target sentence is stored. The determined search target sentence is output as a search result (step S960). Since the similarity determination is expressed as a numerical value, all the search target sentences showing the similarity of a certain numerical value or more may be output. Alternatively, a predetermined number of sentences to be searched may be output in descending order of similarity. Further, all the search target sentences may be rearranged in the order of the similarity value, and the sentence with the highest similarity may be sequentially output.

The data search apparatus 1B of the second embodiment determines the similarity between each of a plurality of search target sentences and a search key sentence, and outputs a sentence having a high similarity as a search result. be able to. When determining the similarity with the search key sentence, the similarity between the concept words is added to the similarity between the functional expressions, which are expressions that give the framework of the sentence, and the word order correspondence between the sentences is determined. Since the judgment is made in consideration of the above, the search of the natural language sentence can be performed with high accuracy.

It should be noted that the external device 90 includes the search engine 10
Upon receiving the search result by B, the search result may be simply displayed, or a process such as translation may be performed using the search result. In the former configuration, for example, a configuration is conceivable in which, from among abstracts of a large number of papers, papers having a high similarity to a search key sentence to be searched are displayed. Alternatively, the present invention is also applied to a configuration in which a search is made for a homepage having a high degree of similarity to the search key sentence to be searched from among a large number of summaries describing the outlines of a huge number of homepages on the Internet, and this is displayed. be able to. the latter,
That is, in the case of translation, the sentence to be translated (search key sentence)
On the other hand, a technique is known in which a sentence very similar to this sentence is searched for in a prepared translated sentence, and a translated sentence is obtained by replacing a concept word in the obtained translated sentence.
Therefore, if the search engine 10B searches for one of the translations prepared in advance, and then replaces the concept word of the searched translation with the translation of the concept word of the sentence to be translated, the translation is performed. good. When detecting a translation with a high degree of coincidence from the translations prepared in advance,
Since it is considered that the similarity of the functional expression, which is the framework expression, is more important than the similarity of the concept word, the search engine 10B of this embodiment is extremely useful in this regard. When the search engine 10B determines the similarity, which of the concept word similarity and the functional expression similarity should be emphasized may be appropriately adjusted by an application. It is also preferable that the concept words are heavier when searching for papers or homepages, and that the functional expressions are heavier when searching for translated sentences.

The embodiment of the present invention has been described with reference to the first and second examples. In the inter-word similarity determination processing and the inter-word string similarity determination processing of the present embodiment, the similarity regarding the structure between character strings is compared with the similarity between functional expressions included in the character strings. This determination method is based on the focus on the "characteristics that form a structure of a sentence linked to a concept word" and "the role that greatly affects the meaning of a sentence" of functional expressions. . Therefore, the inter-word similarity determination processing and the inter-word string similarity determination processing are not limited to the method described in the present embodiment, and other methods that focus on expressions supporting a sentence structure framework such as functional expressions are used. It is also possible to adopt a judgment method. For example, the presence / absence of a functional expression in a character string, the position and type of the functional expression used in the character string, and the like may also be used as elements of similarity determination between character strings.

In this embodiment, the inter-sentence similarity determination device 1A is configured by providing the similarity search engine 10A, the search engine 10B, and the like as a separate device from the external device 90. Engine 10A or 10
B may be integrated to form the inter-sentence similarity determination device 1A or the data search device 1B. For example, by installing a program for executing the character string similarity determination processing and the similar character string detection processing in the external device 90, or by downloading the program to the external device 90 through the public telephone line PTL, the external device 90 itself outputs a sentence. The same function as the inter-similarity determination device 1A can be realized.

The similarity search engine 10A or the search engine 10B of this embodiment inputs a character string from the external device 90 and outputs the similarity determination result regarding the input character string to the external device 9.
Although it is configured to output to 0, it is possible to input a character string by providing an input means such as a keyboard in the search engine 10A or the search engine 10B itself, or to display a determination result using a display means such as a display. It does not matter even if it has a simple configuration.

As another mode for carrying out the present invention,
FD, CD-ROM, R
A recording medium such as an OM chip can be considered. By installing the information stored in the recording medium into the computer, the computer is enabled to execute a character string similarity determination program or the like based on an instruction from the CPU. 1A and a function similar to that of the data search device 1B are realized. Therefore, the same effects as described above can be obtained.

These media are read by, for example, the flexible disk drive FDD shown in FIG. 1, transmitted to the similarity determination engine 10A, etc., and developed and executed in the internal main memory. Instead of using such a medium, a program stored in the server SV may be read from a modem via the network NW, developed in the main memory, and executed.

In this embodiment, the semantic similarity between the character strings is determined, but it is also possible to determine the relationship between other character strings. For example, by focusing on the difference between the functional expressions in the two character strings, it is possible to determine the relationship between the two character strings such as emphasis and limitation, a difference in the degree of estimation, a difference in tense, and the like. Alternatively, the user may arbitrarily select some of these relationships, and the sentence similarity determination device 10 may execute the determination on the selected relationship.

Further, in this embodiment, when determining the similarity between word strings, the similarity between words is prohibited from intersecting between character strings, but this intersection is allowed. It is also possible to adopt a configuration. For example, when the intersection is allowed within the range of two word widths, in the word string distance calculation processing shown in FIG. 14, d is the word string distance to the word two words before the attention word ai, bj. ｛A (i-2), b (j-
2) The value of 、, the value of 2 {1-t (ai, b (j-1))}, which is the inter-word distance between word ai and word b (j-1), and word ai, word b (j -1) is the inter-word distance between the word a (i-1) and the word bj in a relationship intersecting with 2
The sum of the three values {1-t (a (i-1), bj)} is determined as the X4 value, and the minimum value among the X1, X2, X3, and X4 values is calculated up to the word ai. And the distance d (ai, bj) between the word string of the word and the word string to the word bj. In the same way, when the intersection is allowed within the width of four words, the minimum value may be selected from 72 candidates from the X1 value to the X72 value. Thus, it is also preferable to extend the present invention to a configuration that allows intersection within a certain width.

The embodiments of the present invention have been described based on the embodiments. However, the present invention is not limited to these embodiments, and may be implemented in various modes without departing from the gist of the present invention. Of course you can.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram illustrating a hardware configuration of an inter-sentence similarity determination apparatus according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing a flow of information between a similarity search engine 10A and an external device 90 when an inter-character-string similarity determination process is executed.

FIG. 3 is a flowchart illustrating a character string similarity determination routine;

FIG. 4 is a block diagram illustrating roles of a concept word and a functional expression.

FIG. 5 is a flowchart illustrating an inter-word similarity determination routine.

FIG. 6 is a flowchart illustrating an inter-word similarity determination routine.

FIG. 7 is an explanatory diagram showing the structure of a concept word synonym dictionary 36a.

FIG. 8 is an explanatory diagram showing the structure of a functional expression synonym dictionary 36b.

FIG. 9 is an explanatory diagram showing a state where an inter-word similarity and an inter-word distance are recorded in an inter-word information recording table GT.

FIG. 10 is a flowchart showing an inter-word string similarity determination routine A.

FIG. 11 is a flowchart illustrating a dropout cost setting routine.

FIG. 12 is an explanatory diagram showing a distance graph YG in which coordinate axes have been set by drop-out cost setting processing.

FIG. 13 is an explanatory diagram showing how to represent the distance between word strings on a distance graph YG.

FIG. 14 is a flowchart showing a word string distance calculation routine A.

FIG. 15 is an explanatory diagram showing a process of calculating a value of d (a1, b1) and a state of a calculation buffer ET storing a result.

FIG. 16 is an explanatory diagram showing a path from an origin O to coordinates (a1, b1) on a distance graph YG.

FIG. 17 is an explanatory diagram showing a process of calculating a value of d (a1, b2) and a state of a calculation buffer ET storing a result.

FIG. 18 is an explanatory diagram illustrating a path from an origin O to coordinates (a1, b2) on a distance graph YG.

FIG. 19 is an explanatory diagram showing a state of an inter-word-string distance recording table DL in which a distance value between a word string composed of a word a1 and each word string of a comparison character string B is stored.

FIG. 20 is an explanatory diagram showing a process of calculating a value of d (a2, b2) and a state of a calculation buffer ET storing a result.

FIG. 21 is an explanatory diagram showing a route from an origin O to coordinates (a2, b2) on a distance graph YG.

FIG. 22 is an explanatory diagram showing a state of an inter-word-string distance recording table DL that stores a distance value between each word string of an input character string A and each word string of a comparison character string B.

FIG. 23 is an explanatory diagram showing a state in which a distance value between all word strings recorded in the inter-word string distance recording table DL is assigned to each coordinate of the distance graph YG.

FIG. 24 is an explanatory diagram showing a distance between character strings between an input character string A and a comparison character string B using a path on a distance graph YG.

FIG. 25 is an explanatory diagram showing, as a result of the inter-word string distance calculation processing A, different words that are determined to have a similarity to words that have been dropped.

FIG. 26 is a flowchart showing an inter-character-string similarity calculation routine A.

FIG. 27 is a flowchart showing an inter-word string similarity determination routine B.

FIG. 28 is a flowchart showing a word importance setting routine.

FIG. 29 is an explanatory diagram showing a distance graph YW in which coordinate axes have been set by drop-out cost setting processing.

FIG. 30 is an explanatory diagram showing importance values of respective words stored in the Japanese language dictionary.

FIG. 31 is a flowchart showing a word string distance calculation routine B.

FIG. 32 is an explanatory diagram showing a state of an inter-word-string distance recording table DW that stores a distance value between each word string of the input character string A and each word string of the comparison character string B.

FIG. 33 is an explanatory diagram showing the distance between the character strings of the input character string A and the comparison character string B using a path on the distance graph YW.

FIG. 34 is an explanatory diagram showing, as a result of the inter-word-string distance calculation processing B, different words that are regarded as having a similar relationship to the word regarded as being dropped.

FIG. 35 is a flowchart showing an inter-character string similarity calculation routine B.

FIG. 36 is an explanatory diagram showing a data search device 1B as a second embodiment.

FIG. 37 is a flowchart showing a data search routine.

[Explanation of symbols]

 1A ... sentence similarity determination device 1B ... data search device 10A ... similarity search engine 10B ... search engine 10a ... hard disk 10b ... liquid crystal display 10c ... computer 20 ... input interface 22 ... CPU 24 ... ROM 26 ... RAM 34 ... output interface 35 bus 35 synonym dictionary 36a concept synonym dictionary 36b functional expression synonym dictionary 90 external device 90a hard disk 90b display 90c computer 90d keyboard 92 printer 94 modem FDD flexible disk device NW network PTL: Public telephone line SV: Server

Claims (28)

[Claims] 1. A sentence in a predetermined language representing contents having a certain unity, wherein a first sentence to be determined and a first sentence
And a second sentence whose relationship with the first sentence is determined. The first sentence is used by using a constituent unit that constitutes a sentence in the language and that is classified as having a collective meaning. A natural language sentence relation determining apparatus for determining a relation between a sentence of the first sentence and the second sentence, wherein, among the classified constituent units, a conceptual expression classified as a structural unit representing a semantic concept; Relation information storage means for storing at least information representing the relationship between the conceptual expressions and the relationship between the framework expressions, with respect to the framework expressions extracted as constituent units corresponding to the expressions supporting the framework; and the first sentence and the second sentence A constituent unit extracting unit for extracting the constituent unit from a sentence, and referring to the information stored in the relation information storage unit, among the constituent units constituting the extracted first and second sentences. Said A nature determining means for determining the relationship between the first sentence and the second sentence by determining the relationship between the thought expressions and the framework expressions in consideration of the correspondence in word order. Language sentence relation determination device. 2. The framework expression in which the relationship information storage means stores the relationship is a relationship expression that is an expression representing a relationship between concepts such as a case relationship and a causal relationship with respect to a concept expression representing a semantic concept. The natural language sentence relation determination device according to claim 1. 3. The framework expression in which the relation information storage means stores the relation is an advisory expression that gives broadly defined information such as judgment, attitude, and tense on the sentence. Natural language sentence relation determination device. 4. The natural language sentence relation determination device according to claim 1, wherein the relation determination means determines the correspondence while associating the difference in the appearance order of the extracted constituent units as the correspondence of the word order. A natural language sentence relation determination device which is a means for performing. 5. The natural language sentence relation determination device according to claim 1, wherein the relation determination means allows the difference in the appearance order of the extracted constituent units as the correspondence relation of the word order, and A natural language sentence relation determination device, which is a means for performing the determination by prohibiting the intersection of two or more sets of correspondence relations between constituent units. 6. The natural language sentence relation determining device according to claim 1, wherein the relation determining means determines a relation between the conceptual expressions among the constituent units constituting the first and second sentences. A first determination unit that determines by referring to information stored in the relation information storage unit; and a relationship between the framework expressions among the constituent units constituting the first and second sentences, A second determination unit that determines by referring to information stored in the relation information storage unit; and a determination result obtained by the first and second determination units is used while associating the word order. A natural language sentence relation determining apparatus, comprising: a comprehensive determining means for determining a relation between first and second sentences. 7. The natural language sentence relation determination device according to claim 6, wherein the comprehensive determination unit is configured to correspond to each of the constituent units extracted from the first or second sentence. A missing value setting unit that presets a value in the case where the word does not exist in the other sentence as a missing value, and among the constituent units extracted from the first and second sentences while associating the word order with each other. A relationship value assigning means for assigning a value based on the relationship as a relationship value to those having a predetermined relationship, evaluating the assigned relationship value and the set dropout value, A natural language sentence relation determining apparatus, comprising: an inter-sentence relation value calculating means for obtaining a relation value between two sentences. 8. The drop-out value setting means sets the drop-out value as:
8. The natural language sentence relation determination device according to claim 7, wherein the setting unit sets a value according to the importance of the constituent unit. 9. The drop-out value setting unit is a unit that sets a different drop-out value depending on whether a constituent unit that does not exist in the other sentence is the conceptual expression or the framework expression. The natural language sentence relation determination device according to 1. 10. The natural language sentence relation determination device according to claim 6, wherein the relation information storage means includes, as a relation between the concept expressions, a relation between concept words that independently represent a semantic concept, The relationship between a conceptuality affix and a concept word is stored as the information, and the first determination unit stores the relationship between the concept expressions in the relationship information storage unit in addition to the relationship between the concept words. A natural language sentence relation determining apparatus, which is means for determining a relation between the stored conceptuality affix and a conceptual word. 11. The natural language sentence relation determination device according to claim 1, wherein the relation information storage means stores information indicating a degree of similarity as a relation between the conceptual expressions and between the framework expressions. Rejection information storage means, wherein the relation determination means refers to the degree of similarity between concept expressions and framework expressions stored in the resemblance information storage means, and refers to the degree of similarity between the first and second sentences. A natural language sentence relation determination device, which is a similarity determination unit that determines whether or not the determination is made. 12. The natural language sentence relation determination device according to claim 11, wherein the similarity information storage means includes a pair of the conceptual expressions and a pair of the framework expressions and a semantic relationship between the paired expressions. Is stored in an external storage device as information indicating the degree of similarity in the external storage device, and the similarity determination means uses the magnitude of the similarity degree numerical data. A natural language sentence relation determining device for determining the similarity. 13. The natural language sentence relation determining device according to claim 12, wherein the similarity information storage means sets the similarity numerical value data in a range of values from 0 to 1 as the degree of similarity increases as the degree of similarity increases. 1
The similarity determination unit calculates a distance between the conceptual expressions and the framework expressions from the value of the similarity numerical value data, and a sum of the distances is shortest. A natural language sentence relation determining apparatus, comprising: means for specifying a combination; and means for calculating a sum of the distances in the specified combination as a degree of difference between the first and second sentences. 14. A natural language sentence search apparatus for searching a sentence similar to a search key sentence given as a search key from a plurality of search target sentences, wherein the natural language sentence relationship determination according to claim 11. A device, first sentence specifying means for specifying the search key sentence as a first sentence, and sequentially selecting one sentence from the plurality of search target sentences,
A second sentence specifying means for specifying the sentence as a second sentence; and a determination for giving the specified first sentence and the second sentence to the natural language sentence relation determination device to perform the similarity determination. Execution means, and a determination result of the natural language sentence relation determination device is stored according to the assigned second sentence, and a second sentence most similar to a search key sentence given as the first sentence Selecting means for selecting from among the plurality of search target sentences. 15. A dictionary storing at least information indicating a relationship between the concept expressions, which is an expression representing a semantic concept, input means for inputting a first word and a second word, and Extracting means for extracting a word constituting the phrase from the input first and second phrases; and extracting a word constituting the extracted first phrase and a word constituting the second phrase. Word / phrase relation judgment comprising: judgment means for judging a relation by referring to the dictionary; and judgment means for judging a relation between the first word and the second word based on a judgment result by the judgment means. The apparatus, wherein at least one of the first phrase or the second phrase includes a complex expression that is an expression representing a united semantic concept by combining two or more words, and , An expression corresponding to the compound expression Evaluation means for evaluating the relationship between the first word and the second word in consideration of the evaluation result by the evaluation means. apparatus. 16. The phrase relation determining apparatus according to claim 15, wherein: a storage unit that stores information related to a word in advance; and a word included in an expression corresponding to the compound expression does not exist in the compound expression. A word specifying unit that specifies a word; and a type specifying unit that specifies the type of the specified word by referring to the storage unit. The evaluation unit determines whether the type of the word specified by the type specifying unit is If the word is of a predetermined type, similar to the case where the word corresponding to the word exists in the compound expression, the phrase relation determination means for evaluating the relationship between the compound expression and the expression corresponding to the compound expression. apparatus. 17. A sentence in a predetermined language representing contents having a certain unity in a predetermined language, and a first sentence to be determined and a second sentence for which a relationship with the first sentence is determined are input. A method of determining a relationship between the first sentence and the second sentence by using a constituent unit constituting a sentence in the language, which is classified as having a collective meaning, At least, among the classified structural units, a concept expression classified as a structural unit representing a semantic concept, and a framework expression extracted as a structural unit corresponding to an expression supporting a framework of a sentence structure, at least the concept expression Storing information representing a relationship between expressions and framework expressions; extracting the constituent unit from the first sentence and the second sentence; referring to the stored information; 1
And the relationship between the conceptual expressions and the framework expressions among the constituent units constituting the second sentence is determined in consideration of the correspondence between word orders, whereby the first sentence and the second sentence are determined. Natural language sentence relationship determination method for determining the relationship with 18. The natural language sentence relation determination method according to claim 17, wherein the determination of the relation between the first sentence and the second sentence is performed in the first and second sentences. A first process of determining the relationship between the conceptual expressions among the units by referring to the stored information; and a frame process of the framework expressions among the structural units constituting the first and second sentences. A second process of determining a relationship with reference to the stored information, and integrating the determination results of the concept expression and the framework expression while associating the word order with the first and second processes. A natural language sentence relation determining method including a third process of determining a relation between sentences. 19. The natural language sentence relation determining method according to claim 18, wherein the third processing is performed for each constituent unit extracted from the first or second sentence and has a corresponding configuration. An omission value setting process for setting a value in the case where the unit does not exist in the other sentence as an omission value, and a correspondence between the word units and the constituent units extracted from the first and second sentences Among them, a relationship value providing process of providing a value based on the relationship as a relationship value to those having a predetermined relationship, and evaluating the provided relationship value and the set dropout value, And a second sentence relation value calculating process for obtaining a relation value between the second sentences. 20. The natural language sentence relationship determination method according to claim 17, wherein information indicating a degree of similarity is stored as a relationship between the conceptual expressions and the framework expressions, and The determination of the sentence relation is to determine the similarity between the first and second sentences by referring to the degree of similarity between the stored conceptual expressions and between the stored frame expressions. Method. 21. A method for searching a sentence similar to a search key sentence given as a key for search from a plurality of search target sentences, wherein the search key sentence is specified as a first sentence, From the plurality of search target sentences, sequentially select one sentence,
21. Identifying as a second sentence, executing the natural language sentence relation determination method according to claim 20 using the first sentence and the second sentence, and determining a determination result by the natural language sentence relation determination method. A natural language sentence search for saving a second sentence most similar to the search key sentence given as the first sentence from the plurality of search target sentences, storing the sentence according to the given second sentence. Method. 22. A first word and a second word, which are words representing contents having a certain unity, are input, and the input first word and the second word constitute the word. A word is extracted, and the relationship between the word constituting the extracted first phrase and the word constituting the second phrase is converted into at least the relationship between the concept expressions which are expressions representing the semantic concept. A word relationship determination method for determining a relationship between the first word and the second word based on a result of the determination with reference to a dictionary storing information representing the first word and the second word. At least one of the phrase and the second phrase includes a compound expression that is an expression representing a united semantic concept by combining two or more noun words, and corresponds to the compound expression and the compound expression. Evaluate the relationship with the expression Results considering the phrase relationship determining method of determining the relationship between the second word and the first word. 23. A sentence in a predetermined language representing contents having a certain unity, and a first sentence to be determined and a second sentence for which a relationship with the first sentence is determined are input. A program that determines the relationship between the first sentence and the second sentence by using a constituent unit that constitutes a sentence in the language and that is classified as having a collective meaning. A recording medium recorded in a computer-readable manner, among the classified structural units, a conceptual expression classified as a structural unit representing a semantic concept, and a structural unit corresponding to an expression supporting a framework of a sentence structure. A function of recording at least information indicating the relationship between the concept expressions and the relationship between the framework expressions with respect to the extracted framework expressions, and extracting the constituent units from the first sentence and the second sentence; Referring to the stored information, the extracted first
And the relationship between the conceptual expressions and the framework expressions among the constituent units constituting the second sentence is determined in consideration of the correspondence between word orders, whereby the first sentence and the second sentence are determined. And a function of determining the relationship with the recording medium. 24. The recording medium according to claim 23, wherein the function of determining a relationship between the first sentence and the second sentence includes: A first function of determining a relationship between the conceptual expressions with reference to the stored information; and a relationship between the framework expressions in the structural units constituting the first and second sentences. And a second function of determining the first and second functions by referring to the stored information, and determining results of the concept expression and the framework expression while associating the word order. And a third function of determining a relationship between sentences. 25. The recording medium according to claim 24, wherein, as the third function, for each constituent unit extracted from the first or second sentence, a corresponding constituent unit is the other constituent unit. A function of setting a value that does not exist in a sentence in advance as a missing value; and a predetermined relationship among constituent units extracted from the first and second sentences while associating the word order. A function of assigning a value based on the relationship as a relationship value to the objects, and evaluating the assigned relationship value and the set dropout value to obtain a relationship value between the first and second sentences. A recording medium that records functions. 26. The recording medium according to claim 23, wherein information representing a degree of similarity is recorded as a relationship between the conceptual expressions and between the framework expressions. A recording medium that records a function of determining the similarity between the first and second sentences by referring to the degree of similarity between the stored conceptual expressions and between the stored framework expressions as a function of determining a relationship. 27. A computer-readable program for searching a plurality of search target sentences in a predetermined language representing a content having a certain unity and searching for a sentence similar to a search key sentence given as a search key Among the constituent units that are classified as having a unitary meaning, and are constituent units that constitute a sentence in the language.
For the conceptual expressions categorized as structural units representing semantic concepts and the framework expressions extracted as structural units corresponding to the expressions supporting the framework of the sentence structure, at least information representing the relationship between the conceptual expressions and the relationship between the framework expressions is provided. A function of recording and specifying the search key sentence as a first sentence; a function of sequentially specifying one sentence from the plurality of search target sentences as a second sentence; and a function of the first sentence. And a function of extracting the constituent unit from the second sentence, and referring to the stored information to obtain the extracted first unit.
And the relationship between the conceptual expressions and the framework expressions among the constituent units constituting the second sentence is determined in consideration of the correspondence between word orders, whereby the first sentence and the second sentence are determined. And a function of determining the relationship between the first sentence and the second sentence according to the specified second sentence.
And a function of selecting a second sentence most similar to the search key sentence specified as the sentence from the plurality of search target sentences. 28. A first word and a second word, which are words representing contents having a certain unity, are inputted, and the inputted first word and the second word constitute the word. A word is extracted, and the relationship between the word constituting the extracted first phrase and the word constituting the second phrase is converted into at least a relationship between the concept expressions which are expressions representing a semantic concept. And a computer-readable recording medium storing a program for determining a relationship between the first word and the second word based on a result of the determination with reference to a dictionary storing information representing Wherein at least one of the first phrase and the second phrase includes a composite expression that is an expression representing a group of semantic concepts by combining two or more noun words; The expression and the composite table To evaluate the relationship between the expressed correspond to, taking into account the results of this evaluation, a recording medium recording a function of determining a relationship between the second word and the first word.