JP2948159B2 - Database device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION The present invention is intended for general users
The present invention relates to a natural language interface for making it easy to find and register information. More specifically, the present invention relates to a natural language understanding method for realizing such a natural language interface, and a knowledge base method using the same. In particular, the present invention relates to a method of constructing a knowledge database in which a variety of things and facts are stored and can be searched by inference from a natural language.

[0002]

2. Description of the Related Art Conventionally, a natural language interface for a database is known as a human interface using a natural language. For example, see the literature [G. Hendrix, et al., "Developing a Natural Langedge Interface to Complex Data", ACM, Transaction Database System, Vol. 3, 197.
8, P.P. 105-147 (GG Hendrix et al.,
"Developing a Natural Language Interface to Compl
ex Data, "ACM Trans. Database Systems, Vol. 3, 197
8, pp. 105-147)]. In these systems,
A data model for a database (a method for expressing the relationship between stored data items) and a grammar and a dictionary for interpreting a natural language are set independently. That is, when adding a natural language interface to an existing database, it is necessary to construct a grammar and a dictionary anew. Alternatively, if the target database changes, the grammar and dictionary for the natural language interface need to be changed.

Conventionally, a database provided with a natural language interface is a relational database (Rela
database, and a formal language for search used therein, such as a quasi-standard SQL language (Structure
d Query Language (structured query Langedge) is weak in the ability to describe higher-level knowledge. Usually, a question sentence in a natural language is translated into such a formal language as an intermediate language, so that there is a problem that the function of the entire system is limited by the expressive power of this formal language.
In particular, relational databases are effective for uniform data, but are not adequately suited for heterogeneous databases that handle a wide variety of things and object-oriented databases. I can not say. For example, it is not suitable for describing an object from an ambiguous user's memory and retrieving information about the object from the description.

Further, in these systems, input of data (new knowledge and information) uses a method other than natural language, and is performed by experts. Therefore, there is a problem that it is difficult for the end user to directly input and register.

[0005]

SUMMARY OF THE INVENTION The problem to be solved by the present invention is to solve the above problems in databases and knowledge bases so that end users can use natural language even from ambiguous or fragmentary memory. The purpose is to make it possible to retrieve desired information from a description. Another object of the present invention is to enable a user to register new information and knowledge by using a natural language in the same manner.

[0006]

In order to achieve the above object, the present invention provides a common knowledge expression base for a knowledge base and a natural language interface, thereby searching and registering knowledge and information in natural language. Can be easily realized.

More specifically, the present invention provides a knowledge expression method (data in a database) called a “concept relation model” that expresses a system of things and facts by “concepts” and “relationships” as a method of constructing a knowledge base. (Corresponding to the model), and a method for storing linguistic knowledge in the knowledge base is also provided. Here, the “concept” is a data item in a computer that represents an object, an event, or an abstract concept.
“Relation” is a data item defined between the concepts.
Concepts are represented by nodes (nodes), and relationships are represented by links (edges). The knowledge represented by the concept relation model thus constitutes a network of concepts. Here, this is called a concept network.

That is, in the knowledge base according to the present invention,
It is characterized in that knowledge to be originally stored and knowledge to express it linguistically are stored together in a concept network, and that the natural language interface shares the knowledge. Therefore, in principle, it is not necessary to newly construct a dictionary or the like for the natural language interface.

The present invention also provides a natural language comprehension system for performing inference from facts and the like stored in a knowledge base to interpret the meaning of a question sentence expressed in a natural language. In particular, it usually gives a semantic interpretation method for compound noun phrases consisting of a series of nouns that we often use. In order to interpret a compound noun phrase, the system needs to infer the relation between nouns. In the present invention, however, a method for inferring only the concept and the relation stored in the knowledge base is provided.

[0010] Furthermore, the knowledge expression method according to the present invention limits the language-dependent parts so that it can be easily applied to a plurality of languages. Furthermore, it enables the coexistence of expressions in multiple languages. Therefore, a means is provided that enables searching and registration in the same knowledge base in, for example, English and Japanese.

Japanese Patent Application Laid-Open No. 61-220027 discloses a technique for expressing the semantic relationship between words using a concept network and using the same for natural language processing. However, the present invention indicates that data to be searched is stored using a concept network, and that knowledge information for interpreting a natural language is also extracted from the same concept network. Therefore, the amount of knowledge information automatically increases as the data increases.

Thus, the problem of maintenance of the knowledge file which must be performed separately from the database when the knowledge file for interpretation is provided separately can be solved.

[0013]

The principle of the present invention will be described below.

First, a description will be given of a knowledge expression method based on a basic concept relation model. FIG. 2 shows a part of the concept network for explanation. In the figure, an ellipse represents a concept (node), and an arrow represents a relationship (link). Node 201, "UNIVERSAL" is a root node (root) that represents everything in the knowledge base. Each node can define one or more character strings as concept names. For example, synonyms and foreign languages can be added.

On the other hand, the links connecting the nodes include:
Inclusive relations defined between concepts whose attributes are inherited (IS-
A link 202), a “general relationship” (node 203) generally defined between concepts, and a “specific relationship” (link 204) as a specific example of the general relationship. The subsumption relation indicates the class (classification) of an object. Therefore, a concept network consisting of concepts and inclusive relationships is a taxonomic hierarchy (taxonom
ic hierarchy).

For example, in the concept network of FIG. 2, the concept tree represents the following knowledge.

(PAPER-MATERIAL (is-a THING)) (BOOK (is-a PAPER-MATERIAL)) (BOOK # 0051 (is-a BOOK)) (LIVING-THING (is-a THING)) (PERSON ( is-a LIVING-THING)) (NEWTON (is-a PERSON)) Here, these are called frame representations. here,
Symbolic expression used in LISP language
It is represented by These can be expressed in ordinary English, and can be written as Paper material is a thing. Book is a paper material. In Japanese, "books are things.""Books are books.""People are living things."

An example of a general relationship is shown in FIG.
1 (BOOK) and a concept 203 defined between the concept 212 (PERSON). This indicates that there may be a relationship between "book" and "person" such as "author" or "work". The general relationship is (BOOK (is-written-by PERSON)) (PERSON (has-written BOOK)) or right or left, such as "the person who wrote the book" or "the book that the person wrote." You can read in both directions.

The specific relationship is a relationship shown by a broken line link 204 in FIG. 2, and shows one specific example of a general relationship (also referred to as an actual value in a database). For example, in FIG.
As a specific example, a concrete relationship 204 is defined between the concept 213 and the concept 214. When writing in frame expression, Becomes In natural language, BOOK # 0051 is a book which is written by NEWTON. NEWTON is a person who has written BOOK # 0051. "BOOK # 0051 is a book written by NEWTON.""NEWTON wrote BOOK # 0051. It is a person. "

The above knowledge is stored in the following data structure in the method of the present invention. First, the concept and its name are stored in the concept name table 221 (FIG. 3).
The table 221 has three columns 222, 223, 2
Consists of 24. Column 222 is the unique number C of the concept.
#, And the name CNAME can be defined in the column 223. The language LANG of the same name is defined in column 224. For example, when the value of LANG is "J", the name is Japanese, and when it is "E", it is English.

A plurality of concept names can be defined for the same language. Therefore, column 223
Allows data to be repeated as a data structure. For example, the name of concept C # 0004 is “book” and “BOO”
K ", the concept name" book "can be additionally defined.

Next, the subsumption relation of the concept is represented by a subsumption relation table 231 shown in FIG. Column 232 and Column 2
Reference numeral 33 denotes a concept unique number C # and a CS #, respectively, and indicates that the superordinate concept (Superconcept) of the concept C # is the concept SC #. For example, the second record of the table 231 indicates that the superordinate concept of the concept C # 0002 (“book”) is the concept C # 0002.
# 0001 ("thing"). "Relationships" such as attributes defined for each concept are inherited from higher to lower levels through links of subsumption relations. It is possible to define a superordinate concept of
Therefore, multiple inheritance of attributes is realized.

A wide variety of relationships between concepts other than inclusive relationships can be defined in a general relationship definition table 241 shown in FIG. Each individual general relationship represents a type of relationship. Basically, there is no limit on the number of such relationship types, and it is possible to define any number of general relationships.

The general relation definition table 241 mainly defines "reading" when "relation" is expressed in a natural language.
Column 244 is the “reading” LR when the relationship is read from left to right, and column 245 defines the “reading” RL from right to left. These columns allow repetition as a data structure and can define multiple readings. In addition, the reading language can be designated by the column 246 as in the concept name table.
Therefore, it is possible to express in a plurality of languages.

In the example of FIG. 5, the relationship "AUTHORSHIP" is PERSON who is author of BOOK PERSON who is the author of BOOK PERSON who wrote BOOK PERSON who has written BOOK or BOOK whose author is PERSON BOOK by PERSON BOOK from PERSON It can be expressed in BOOK of PERSON and natural language (English in this case). The same is true for Japanese (omitted).

The existence of a relationship between concepts is stored in a relationship table 251 shown in FIG. As described above, the relation link includes the general relation and the specific relation, which are distinguished by the column 256 of the table 251. General relation when the value of column CLASS is GR, I
NST indicates a specific relationship. In the example of FIG.
The record represents the general relationship 203 of FIG. 2, and the second record represents the concrete relationship 204 of FIG. In addition, column C #
L defines the concept on the left, and column C # R defines the concept on the right. Here, whether a certain concept is placed on the right or on the left depends on the definition, and either may be defined as long as there is no contradiction between the table 241 and the table 251.

Next, the principle of the natural language understanding system based on the above knowledge expression system will be described.

First, a method of understanding the meaning of a compound noun phrase which is the most important in an object-oriented knowledge base will be described. Here, the compound noun phrase is a noun phrase composed of a sequence of a plurality of nouns including some adjectives. For example, the following is an example of a compound noun phrase.

Supercomputer article (1) ElectronicsWeek article (2) Japanese personal computer company (3) American personal computer software packages (4) Here, understanding the meaning explicitly seeks the relationship between each adjective and noun. That is.

For example, the compound noun phrases (1) and (2) have the same structure but different meanings and should be interpreted as follows. (1) is "article whose subject i
s supercomputer "and (2) is" articl
e which is part of ElectronicsWeek ". In other words, in (1), article and supercomputer are related
connected by subject-is, (2), article
It is necessary to infer that ElectronicsWeek is connected with the relationship is-part-of. When you write in a frame format,
It is important to automatically extract the following structures.

(ARTICLE (subject-is SUPERCOMPUTER)) (5) (ARTICLE (is part-of ElectronicsWeek)) (6) In the natural language understanding method according to the present invention, based on the knowledge shown in FIG. Interpret the meaning as follows. First,
The knowledge that enables this inference is, as a general relation,
Relationship RS # 0011: (ARTICLE (subject-is UNIVERSAL)) (7a) (UNIVERSAL (is subject-of ARTICLE)) (7b) and Relationship RS # 0012: (ARTICLE (is-part-of JOURNAL)) (8a) ) (JOURNAL (has-part-of ARTICLE)) (8b) must be defined. In other words, it is necessary to memorize that "the subject of the article can be anything" and "the article is part of a magazine (the article is published in a part of the magazine)". is there.

As the inclusive relation, (SUPERCOMPUTER (is-a THING)) (9) (THING (is-a UNIVERSAL)) (10) (ElectronicsWeek (is-a JOURNAL)) (11) is stored. Need to be.

Thus, the "supercomputer article" can be interpreted. First, from SUPERCOMPUTER, tracing the subsumption relation to the higher rank, it can be seen that (SUPERCOMPUTER (is-a UNIVERSAL)) (UNIVERSAL (is-subject-of ARTICLE)). As a result, it is inferred that the relationship of (SUPERCOMPUTER (is-subject-of ARTICLE)) or (ARTICLE (subject-is SUPERCOMPUTER)) may exist due to attribute inheritance. That is,
"Supercomputers can be the subject of articles"
Is elected. In this case, there is no other interpretation, so the interpretation: "article whose subject is supercomputer" is adopted.

The meaning interpretation of the compound noun phrase (2) is a little complicated. In this case, as can be seen from FIG. 7, (ElectonicsWeek (is-a JOURNAL)) (JOURNAL (has-part-of ARTICLE)) and (ElectronicsWeek (is-a UNIVERSAL)) (UNIVERSAL (is-subject -of ARTICLE)), so (Electronics Week (has-par
(t-of ARTICLE)) and (Electronics Week (is-subj)
ect-of ARTICLE)). That is, it can be understood that there are two interpretations of "article which is part of ElectronicsWeek" and "article whose subject is ElectronicsWeek".

When a plurality of interpretation candidates are present, the present invention uses a heuristic technique of evaluating the likelihood of interpretation based on which interpretation has many specific cases.

Specifically, in the above example, the concept "ARTICLE"
And the subordinate concepts of "ElectronicsWeek", including myself, are searched for the relation RS # 0011 and the relation RS # 0012.
And how many specific relationships are registered. In the example of FIG. 7, 0 is registered for the former and one specific relationship is registered for the latter. That is, there is no article on "ElectronicsWeek", but there is one article on "ElectronicsWeek", ARTICLE # 0101. Therefore, the relation RS # 0012 (is-part-of) is selected as a more appropriate interpretation. That is, "article which is part of El
ectronicsWeek ". As described above, the interpretation of compound noun phrases is based on the inference process of the relationship between two nouns. That is, a compound consisting of three or more words described below. The basic processing of the interpretation of the noun phrase is the extraction of the relationship between the two words, and the following description will be made using the compound noun phrase (3) as an example.

First, a concept corresponding to each word is selected while checking whether or not a word constituting the compound noun phrase includes a concept name composed of a compound word. That is, words are sequentially cut from the beginning, and it is checked whether or not the words are registered by referring to the concept name table.

In the case of compound noun phrase (3), Then, a partial word string is cut out from the compound noun phrase, and it is checked whether each is a concept name.

Here, in this method, adjectives are registered as synonyms of a concept whose name is the corresponding noun form.
One feature of adjectives is that they are treated as the same concept.
For example, the adjective "Japanese" is registered as a synonym for the concept "JAPAN" (Japan) or the concept "Japanese people" (Japanese) and is treated as the same concept.

Therefore, as a result of this processing, "personal comp
If "uter" is defined as one concept name, "PERSONAL-COMPUTER", the compound noun phrase (3) is first recognized as (JAPANESE-PEOPLE PERSONAL-COMPUTER COMPANY). However, in the following description, for the sake of simplicity, the latter that is understood to be a meaningless interpretation is omitted.

That is, at this stage, it has been found that the compound noun phrase is substantially a combination of three concepts. This can be expressed as follows using parentheses ().

(Japanese (personal computer) company) (12) Therefore, the next processing is to examine how these three concepts are related to each other. In this case, it can be seen that there are two possibilities.

(Japanese ((personal computer) company)) (13) ((Japanese (personal computer)) company) (14) First, in the case of (13), between COMPANY and PERSONAL-COMPUTER, and between COMPANY and JAPAN We need to infer two relationships that can be connected. In this case, according to the relation inference method described above, (COMPANY (produces PERSONAL-COMPUTER) (is-located-in JAPAN)) (15a) (COMPANY (has-developed PERSONAL-COMUTER) (is-located-in JAPAN ) (15b) is extracted, where COMPANY and PER are used to evaluate the priority (likelihood: likelihood) of multiple interpretations.
The total number of SONAL-COMPUTER and the concrete examples of the two relations of COMPANY and JAPAN (concrete relations defined in the lower concepts) are counted for each interpretation (15a) and (15b), and this is counted. Weight. As the overall evaluation, the total number of specific relations (the weight of relations) is divided by the number of general relations to normalize. In the examples of (15a) and (15b), the number of general relations is two.

Next, a function is extracted for the second possibility (14). In this case, COMPANY and PERSONAL-CPMPUTE
It is necessary to find R and two relations of PERSONAL-COMPUTER and JAPAN. For the former, two relationships are found (in the assumed knowledge base), such as (COMPANY (produces PERSONAL-COMPUTER)) (16a) and (COMPANY (has-developed PERSONAL-COMPUTER)) (16b). Similarly, for the latter, (PERSONAL-COMPUTER (is-produced-by (COMPANY (is Located-in JAPAN)))) (17a) (PERSONAL-COMPUTER (was-developed-by (COMPANY (is-located-in JAPAN) ))) (17b) is found, here, PERSONAL-COMPUTER and JAPAN
Since there is no relation directly connecting with, the concept COMPANY which relates these indirectly is automatically found.

In the method of the present invention, when a relation to directly connect cannot be found in this way, the general concept (set) of all the general relations defined in those higher concepts is traced back to each of these two concepts for which the relationship should be found. Is extracted. Since we are working on two concepts, we have two sets of general relations. This is the three tables 231, 241, 25
1 can be extracted by searching. Next, the concepts connected to the extracted general relationship are listed. As a result, two sets of concepts are obtained. By calculating the intersection of these sets, a concept that mediates the two concepts, that is, the concept COMPANY in the above example, is finally obtained.

After all, the meaningless combinations are discarded, and the following two interpretations are obtained.

(COMPANY (produces (PERSONAL-COMPUTER (was-developed-by (COMPANY (is-located-in JAPAN)))))))) (18a) (COMPANY (has-developed (PERSONAL-COMPUTER (is-produced- by (COMPANY (is-located-in JAPAN)))))) (18b) Ultimately, the following four interpretations will be automatically obtained.

"Company which produces personal computer which is located in Japan" (19a) "company which has developed personal computer which is located in Japan" (19b) "company which produces personal computer which was developed by company which is located in Japan "(19c)" company which has developed personal computer which is produced by company "(19d) In the above-mentioned English expression, the relationship of the relation pronoun is indicated by indentation. If the level is deeper, it modifies the noun in the line immediately above.

The same applies to the compound noun phrase (4). By matching each word to the concept name, you can first see that it consists of three concepts, such as (USA PERSONAL-COMPUTER SOFTWARE-PACKAGE). By inferring the relationship between these concepts, the following interpretation is obtained.

"Software package which runs on personal computer which was developed by company which is located in USA" (20) The natural language comprehension system according to the present invention has (Syntactic analysis unit) enables a more powerful noun phrase interpretation function and at the same time realizes extensive dialogue with the knowledge base in natural language. The sentence interpreter uses the syntax rules, but Listing 1
Indicates an example of the syntax rule. One syntax rule is that non-terminal symbols (non-terminal symbols;
), And the right-hand side, which is its expanded form, and three parts of the function to be evaluated when the expansion is successful (eg, equation (21a)). In the list 1, the evaluation function when the analysis is successful is omitted. The right side of each rule is composed of a sequence of terminal symbols (terminal symbols) and non-terminal symbols. Here, the symbol "|" indicates juxtaposition. Also, the non-terminal symbol <s>
Represents the whole sentence.

−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−− List 1: <s>:; = (21a) <define-command> | (21b) <search-command> | (21c) <browse-command> (21d) <system-command>;: = load from <?> | save to <? update | quit | stop | <show-word> chart | edit <?><show-word>;: = show | print <search-command>;: = <find-word><noun-phrase> (22 ) <find-word>;: = look up | look for | search for | lookup | find (23) <noun-phrase>;: = <noun-group> | (24a) <noun-phrase><joiner><noun-phrase> (24b) <noun-group>;: = <noun-compound> | (25a) <art><noun-compound> | (25b) <def-art><noun-compound> (25c) <noun-compound>;: = <noun-string> (26) <noun-string>;: = <concept> | (27a) <concept><noun-string> (27b) <joiner>;: = <prep><rel-pronoun><link> (28) <define-command>;: = define relation <new-phrase>. | (29a) define <?> as <noun-phrase>. | (29b) <noun- phrase> is <noun-phrase>. | (29c) <noun-phrase><link><noun-phras (29d) <noun-group> is also known as <?>. (29e) <new-phrase>;: = <noun-compound><?><noun-compound> (30) <?>;: = <null> | <null><?> (31) −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−− -The sentence interpreter consists of a vocabulary analyzer by bottom-up processing and a parser by top-down processing. The syntax analyzer in this method outputs all sentence structures that match the syntax rules as candidates. From there, only the meaningful interpretation (the interrelation between words) is evaluated and selected by the compound noun phrase analysis described above.

Lexical analysis (lexica) which is a bottom-up process
l analysis) is a lexical dictionary (lexico) that describes concept names, relational descriptors, prepositions, relational pronouns, and articles.
n).

The vocabulary dictionary can be automatically constructed by editing processing from the concept network based on the knowledge expression method described above. Specifically, the concept names are extracted from the table 221 (FIG. 3). As a result of lexical analysis,
Words in the input sentence that match the concept name are <concept
Recognized as the nonterminal symbol>.

The relation descriptor is a “relation” or “reading” thereof in the same knowledge expression method. For example, “is author of”, “wrote”, “has written”, “autho”
ris "etc. These are the general relationship definition table 2
41 (FIG. 5). Words or word strings in the input sentence that match these are recognized as non-terminal <link>.

Prepositions are "by", "from", "of", etc.
Its meaning (the type of relation it represents) is automatically extracted from the general relation definition table 241 (FIG. 5). These prepositions are recognized as non-terminal <prep> symbols. Here, in the case of being recognized as a relation descriptor and a preposition, the type of the "relation" and the direction of the non-terminal symbol are output as a lexical analysis result together with the non-terminal symbol.
This makes it possible to interpret the meaning of a noun phrase such as "book by Newton", as described later (syntax rules (24b),
(See (28)).

The vocabulary dictionary further includes a relational pronoun <rel
-pronoun> (which, who, that), indefinite article <art> (a,
an) and the definite article <def-art> (the). Words that could not be recognized by lexical analysis are non-terminal symbols <nu
ll>

The syntax analysis is performed based on the vocabulary analysis as the bottom-up preprocessing. The syntax analysis method will be described with reference to FIG. In the figure, three words 401, 402, and 403 are analyzed as input sentences. The arrow flowing from the upper part to the lower part (for example, the arrow 411) indicates the bottom-up processing, and the arrow flowing from the lower part to the upper part (for example, the arrow 41).
2) represents a top-down process.

As a result of the lexical analysis, the word 401 (pers
onal) has no matching concept, word 402
The interpretation of (computer) is that there can be two concepts, “computer” and “magazine computer”, word string 401
-402 (personal computer) matches the concept "personal computer", word 403 (comp
any) is output to match the concept "company".

The syntax analyzer receives these results, and expands the entire sentence with reference to, for example, the syntax rules shown in Listing 1. FIG. 8 shows only from the middle,
In the figure, <noun-phrase> 421 is converted to <noun-compound> 422 according to equation (25a), and <noun-phrase> 421 according to equation (26).
noun-string> Expands to 423. Furthermore, <noun-strin
g> is expressed by <concept> 426 and <noun-stri
g> 427. The latter <noun-string> is (2
It is expanded to <concept> 428 by the expression 7a). After all, in the example of FIG. 8, the interpretation of the input sentence (noun phrase) is (personal c
It can be seen that there can be a compound noun phrase (noun compound) composed of two concepts (<concept>) of (omputer) and (company).

Here, in FIGS. 8 to 11, the symbols shown in List 1 are simplified as follows in order to simplify the drawings. That is, <noun-phrase>, <noun
-group>, <noun-compound>, <noun-string>, <joiner>,
<rel-pronoun> and <find-word> are <n-phrase
>, <n-group>, <n-comp>, <n-string>, <join>, <rpron
>, <find>.

In this syntactic analysis, all syntactically possible sentence structures are extracted. Examples of such parsing algorithms include, for example, Chart Barza [Reference: H. Thompson, Etoal, "Implementing" Natural
Langedge Barthers ", Artificial Intelligence: Tours, Technique, And Applications, Edited By The Ossia And M Eisenstadt, Harpa And And
Rho, New York, 1984, PP, 245-3
00: Thompson, H. et al., "Implementing Natural L
anguage Parsers, "in Artificial Intelligence: Tool
s, Techniques, and Applications, editedby T. O'She
a and M. Eisenstat, Harper and Row, New York, NY,
1984, pp.245-300] is known.

All the extracted syntax candidates are evaluated by the evaluation formulas described in the syntax rules. A feature of the present invention is that knowledge stored in a knowledge base is used for this evaluation. In the example of FIG. 8, the concept "personal com
puter "and the concept" company "to form a meaningful compound noun phrase. This evaluation method uses the knowledge registered in the concept network described above. Will be obtained in the order of likelihood.

FIG. 9 and subsequent figures show the results of parsing the entire search sentence. The search sentence is expanded into a command word and a noun phrase using the syntax rules (21c) and (22). FIG. 9 is an example including the noun phrase of FIG. FIG. 10 is an example of analysis of a search sentence including a noun phrase including a relative pronoun. Here, it can be seen that “reading” of “relation” in the conceptual relation model shown in FIG. 5 functions as a predicate of the relation pronoun phrase. Further, FIG.
Is an example of a search sentence including a preposition. Where the preposition o
n is the relation "subject"("whose subject is") and the preposition a
The fact that t means the "affiliation"("who works at") is described in the general relationship definition table shown in FIG. 5, as already described.

According to the principle described above, it is possible to semantically analyze a natural language sentence including a sequence of nouns (compound noun phrases) for which no relation is explicitly described, using a knowledge base and minimal syntax rules. .

[0065]

【Example】

(Embodiment 1) An embodiment of the present invention will be described below. FIG.
FIG. 4 is a block diagram of a hardware configuration of a knowledge base system according to one embodiment of the present invention.

The system is mainly composed of a central control unit (CPU) 101 and a main memory unit (MMU) in FIG.
102, a work memory unit (WMU) 103, integrated disk units 121 and 122 having system files and user files, a printer 131, a communication control unit (CCU) 141 for connecting to a local area network (LAN), and a user A display device (CRT) 151, a keyboard 152, and a mouse 153 for dialogue with the user.

The integrated disk unit is composed of a magnetic disk and an optical disk. For example, in the case of this embodiment, the first disk unit (DKU1) 121 is composed of a magnetic disk and stores information requiring high speed. I do.
On the other hand, the second disk unit (DKU2) 122 is made up of an optical disk, and particularly stores a large amount of information.

The system file stores a group of programs provided by a system such as an operating system, a communication management program, a file management program, a knowledge base management program, or an integrated office operation support system, and a group of files such as a database required for the system. ing. On the other hand, a user file includes a group of application programs such as a personalized system-provided program and a program created by a user, a group of files such as a database required for the program, and an integrated knowledge base (described later) created by a user and the like. Is stored.

The operation of the system is performed by accessing necessary data to the main memory unit and the file under the control of the program loaded in the main memory unit. At the same time, the user can interact with the system using the input means 152 and 153 while viewing the information displayed on the display device 151.

The software part of the knowledge base system according to the present invention includes a program group consisting of a concept network editor, a concept network browser, a natural language sentence interpretation program, and a multimedia file management program;
It consists of an integrated knowledge base integrating the information used by them. Hereinafter, each of them will be sequentially described.

First, the concept network editor edits, searches, and manages an object-oriented knowledge base based on the concept relation model described above. The knowledge and information are stored and managed by the concept name table, the inclusion relation table, the general relation definition table, and the relation table shown in FIGS. The data stored in these is called a concept network here.

The concept network editor provides the following functions using a command language (formal language) based on a predetermined specification.

(1) Concept registration: A new concept is registered by designating a concept name and a classification hierarchy position.

(2) Registration of general relation: A new relation type is registered by designating two concepts and a linguistic expression of the relation.

(3) Registration of concrete relation: Facts, relations, attributes, etc. are registered by designating two concepts and the kind of relation.

(4) Editing a concept name: correcting the name; deleting / adding synonyms / aliases.

(5) Editing of general relations: change of two related concepts; modification and addition of linguistic expressions.

(6) Deletion of concepts (7) Deletion of concrete relations (8) Change of hierarchical classification system (9) Addition of superordinate concept: Classify the same concept doubly (or more); Multiple inheritance of attributes, etc. Grant.

(10) Registration / deletion of a structured concept: Registration and deletion of an abstract concept (described later) in which the concept is modified by the concept.

(11) Update of vocabulary dictionary: Information necessary for vocabulary analysis is added / edited from the concept network to the vocabulary dictionary. □ (12) Search for concepts: Search from substrings, keywords, and structured concepts.

(13) Network traversal : movement of a concept of interest by a classification hierarchical tree and association.

(14) Addition / deletion of multimedia information:
Associatively add or delete information in multimedia files to any concept.

(15) Read multimedia information: Read multimedia information added to a concept from a multimedia file.

Here, the structuring concept in the above function will be described. The structured concept is an abstract concept in which a certain concept is modified (limited in meaning) by another concept and a relation, and is stored in a storage unit different from the concept network. Specifically, the compound noun phrase shown in the examples so far corresponds to this. For example, the structured concept D # 0011 You can register as More specifically, the following internal data structure: Is stored as a unique number. Here, the first element is a structured concept unique number, the second element is a structured concept name, and the third element is a structured concept definition formula. However, in the above notation example, C #
0401 is the concept "article", C # 0501 is the concept "supercompute"
r "and RS # 0033 are unique numbers of the relation" subject ".

This structured concept is used when retrieving information from abstract or descriptive storage. The present system has a function of searching for a specific concept from a structured network using a reasoning process from a concept network. In the example above, the concrete concept (concept
From the set of "articles", the subject is a concept "
It is possible to extract only those concepts that are subsumed by "supercomputer". This process is called concept matching, which is realized by reverse inference.

As described above, by providing the means for storing the structured concept and the search means therefrom, the user can arbitrarily register his / her own viewpoint in the system, and further,
It is possible to easily search for a desired concept.

In the function of the concept network editor, the multimedia information function will be further described. For example, photos input by an image scanner can be registered and stored in the multimedia file. The concept network editor has the function of associatively adding the photo to a concept. More specifically, this is realized by creating and managing a correspondence table between concept unique numbers and multimedia file member names. After searching for the same concept, the user can display the photograph. In addition to photographs, document images, audio information, and the like can be recorded in the multimedia file.

As a method of configuring the concept network editor, the invention “Information storage method” (Japanese Patent Application No. 60-6067)
Since the method disclosed in 8) can be used, the description is omitted here.

Next, the concept network browser will be described. The concept network browser is a visual interface (visual interface) for the concept network editor, and improves operability by adopting a direct operation method using a multi-window and a mouse. The browser can display the contents of the knowledge base in an arbitrary window in a concept hierarchical tree, frame format, table format, or even natural language format. Also, by directly pointing an object (concept and relation) on the display with, for example, a mouse, and further selecting and instructing a command from a menu, an arbitrary function of the above concept network editor is called and executed. I can do it. The concept network browser translates such a user's direct instruction into a command language, which is a formal language, and activates the concept network editor.

Next, a natural language sentence interpretation program will be described with reference to a functional block diagram and FIG. One of the main features of the present invention resides in the present natural language sentence interpretation system, and its principle has been described above. The natural language sentence interpretation program has an input unit 502 for inputting a word string 501 from a user, a vocabulary analysis unit 503 for analyzing attributes of each word of the word string, and a grammar file storing syntax rules. It comprises a syntax analyzer 504 for analyzing the syntax while referring to 514, and a compound noun analyzer 506. The operation, the natural language text interpreter receives a natural language sentence 501 from the concept network browser, returns the interpretation result 505 in the concept network browser.

In the present embodiment, the vocabulary analysis unit includes the structured concept file 511, the vocabulary dictionary file 512, and the concept name table 5 of the concept network type knowledge base 513.
The analysis is executed with reference to 22. Here, as described above, the structured concept file 511 stores a set of three sets of a structured concept unique number, a structured concept name, and a structured concept definition expression. The vocabulary dictionary 512 includes a preposition table 601 shown in FIG. 12 and a relation descriptor table 611 shown in FIG. As another embodiment, there may be a configuration in which a concept name table is also included in the vocabulary dictionary for high-speed processing.

The preposition table 601 stores, for each preposition, a unique number of a general relation represented by the preposition, for each direction of the relation. In the example of FIG. 12, the preposition at
Represents a “relation” such as a case where the general relationship RS # 0123 is read from left to right or a case where RS # 0209 is read from right to left. The relation descriptor table 611 stores, for each relation descriptor, its general relation unique number and its directionality. These two tables are automatically constructed or updated from the general concept definition table shown in FIG. 5 by storing all possible prepositions separately so that they can be easily understood from the above description. You can do it. This function is provided by the vocabulary dictionary editing unit 507 in FIG. 1, and is actually a part of the function of the concept network editor.

Here, the description of the function of the vocabulary analysis unit is partially supplemented. In the explanation of the principle of the invention, the structured concept was not described in order to facilitate understanding.However, the vocabulary analyzer recognizes the word and word string that matched the structured concept name as a non-terminal symbol <concept>, In the case of a structured concept, the structured concept definition expression is output as an evaluation value.

For this purpose, the vocabulary analyzer sequentially extracts partial word strings from the input word strings as in the principle explanation,
It is checked whether or not each partial word string matches not only the normal concept name but also the structured concept name. As a result, if any of them match,
<Concept> is a recognition result as a non-terminal symbol.
When matching with a normal concept name, the unique number of the same concept is used as the evaluation value, and when matching with the structured concept name, the definition formula of the same structured concept is used as the evaluation value.

Further, as a result of the vocabulary analysis, when it is recognized as a preposition and a relation descriptor, the corresponding non-terminal symbol is set as a recognition result, and all the corresponding "relations" are set as evaluation values. These evaluation values are output as a vocabulary analysis result accompanying the non-terminal symbol as the recognition result.

As described in the principle explanation, the structure analysis unit 504 receives the lexical analysis result as an input, and
First, generate all possible sentence structures while referring to the syntax rules of. Purging (in a narrow sense) a possible sentence structure according to syntax rules is called parsing. At this purging stage, no semantic relevance was considered. Next, the syntactic analysis unit 504 causes the compound noun analysis unit 506 to perform semantic compatibility analysis and semantic analysis of the compound noun phrase. As a result, semantically compatible sentences are selected in the order of likelihood from the set of sentence structures generated as candidates.

As a result of the parsing, a semantically particularly important element is the noun phrase <noun-phrase>. Therefore, the parsing of the noun phrase will be described in more detail with reference to the examples of FIGS. Will be described.

First, in the case of FIG. 9, the noun phrase “personal computer companies” is analyzed as a structure shown below as a result of parsing.

(<Noun-phrase> D # 1009 "personal computer companies"((<concept> (C # 1001 "personal computer")) (<concept> (C # 1022 "company")))) (32) Here, in the equation (32), the first element is a non-terminal symbol,
The second element indicates a structured concept number given as needed, the third element indicates a corresponding word string of the input sentence, and the fourth and subsequent elements indicate a parsing result. The parsing result juxtaposes multiple interpretations if they exist. In the equation (32), since the candidate is a single candidate, the fifth and subsequent elements do not exist. In addition, the fourth element generally constitutes a sequence composed of concepts and relationships. (3
In the example of the expression 2), there are two concept columns. An example including the relationship will be described later using Expression (35).

As shown in equation (32), these concepts and relations are represented by columns of concepts or relations specifically matched with the identifiers (non-terminal symbols). (3
In the example of equation 2), each concept has only a single matching result. An example of a plurality of matching results is shown in Expression (37).

Now, the parsing unit checks the parsing result (3
2) In response to the expression, the last two lines of the expression are replaced with the compound noun analysis unit 5
Send to 06. The compound noun analysis unit infers the semantic relationship between the two concepts C # 1011 and C # 1022 from a knowledge base represented by a concept network by the method described in the principle explanation. As a result (in the assumed knowledge base)
The following two interpretations, Expressions (33) and (34), are obtained.
In this example, relations not explicitly described are extracted by inference processing.

[0102] These are expressed as follows in natural language.

[0103] Next, an example of FIG. 10 will be described. In this case, the purging result is as shown in equation (35). That is, a noun phrase is recognized as two concepts connected by a joiner.

(<Noun-phrase> D # 1010 "person who is author of a book"((<concept> (C # 1101 "person")) (<joiner> (RS # 0001 RL "is-author-of ")) (<concept> (C # 1122" book "))))) (35) Here, the compound noun analysis unit receives the last three lines of the expression (35) and receives the concepts C # 1101 and C # 1122. Is related (RS
# 0001 RL), and finally obtain the interpretation of equation (36). In this case, equation (35) shows only a single candidate, but if the above inference failed, such a semantic interpretation would be based on knowledge stored in the knowledge base. Means that it is impossible.

(D # 1010 "person who is author of a book" (C # 1101 ((RS # 0001 RL) C # 1122))) (36) When the expression (36) is displayed in a natural language, it is as follows. become.

"Person who is author of book" Finally, an example of FIG. 11 will be described. In the parsing result shown in Expression (37), the second joiner is obtained by interpreting the preposition at with reference to the preposition table 601 shown in FIG. 12, and as a result, a plurality of interpretations are listed. I have.

(<Noun-phrase> D # 1011 "articles on people at companies"((<concept> (C # 0401 "article")) (<joiner> (RS # 0033 LR "subject-is"))) <concept> (C # 0601 "person")) (<joiner> (RS # 0123 LR "is-developed-at") (RS # 1054 LR "is-located-in") ・ ・ ・ (RS # 0209 RL "works-at") ・ ・ ・) (<concept> (C # 1022 "company"))) (37) The compound noun analysis unit 506 inputs the fourth element of the expression (37) and returns "concept-join". The validity of the semantic relationship is evaluated for each of the three sets of “child-concepts” according to the concept described in the principle description.
There is no relation "is-developed-at" or "is-located-in" between 601 (person) and concept C # 1022 (company), and only the relation "works-at" is semantically valid. Something is inferred from concrete facts stored in the knowledge base. In this case, the interpretation of equation (38) is finally obtained.

(D # 1011 "articles on people at companies") (C # 0401 ((RS # 0033 LR) (C # 0601 ((RS # 0209 RL) C # 1022)))))) (38) Interpretation (38 The expression is expressed as follows in natural language.

[0109] Next, a multimedia file management program which is the next element of the knowledge base system according to the present invention will be described. While a knowledge base stores objects or concepts symbolically, multimedia files are
The information itself is stored as media. In other words, each performs optimal encoding for information compression, but photographs, documents,
A plurality of different types of media information (multimedia information) such as drawings, texts, and voices are stored and managed according to a predetermined description method. The description method of the multimedia information is not described here because a multimedia content architecture (for example, ODA: office document architecture) to be standardized is known.

The multimedia information is managed in a predetermined unit, and forms one object (information unit). For example, objects that we usually recognize as one thing, such as “document”, “article”, “photograph”, “chart”, “single-segment voice message”, etc., can be objects. This object may be a single media or a multimedia. These objects are managed by a multimedia file management program using management information including a file member name assigned to each information unit, a storage position and a storage record length in a physical file, a description format (for example, an encoding method), and the like. A plurality of objects can be structured to form another object. Conversely, an object can be managed as a structure of a plurality of partial objects.

On the other hand, each information unit stored in the multimedia file can be associated with an arbitrary concept node of the concept network as described above. As a result, it is possible to intelligently search for the information body stored in the multimedia file from the concept network type knowledge base by using inference.

Next, the integrated knowledge base will be described. As described above, the knowledge base system according to the present invention includes a concept network type knowledge base based on a concept relation model, a grammar file storing syntax rules, and a linguistic dictionary which can be automatically created from the knowledge base by editing. (Vocabulary dictionary 512), a concept dictionary (structured concept file 511) that the user can define for each individual, and a multimedia file. In addition, the system has storage means for associating concepts in the knowledge base with objects in the multimedia file. Such a knowledge base that can organically store knowledge / information by symbolic / logical expressions, linguistic knowledge, and media information is referred to herein as an integrated knowledge base. These pieces of information are stored in the integrated disk unit according to the respective characteristics.

Although the present invention has been described with particular reference to the case of English, in principle, the present invention can be similarly extended to other languages such as Japanese. In particular, as described in the description of the knowledge expression method, according to the present invention, the system can be easily extended to a multilingual system. Specifically, among the knowledge expressing means, only the concept name table 221 (FIG. 3) and the general concept definition table 241 (FIG. 5) depend on the language, and at the same time, those tables are defined by fields (columns) defining the language. LAN
G, 224 and 246.

Further, the present invention is not limited to the contents of the syntax rules shown in the embodiment (List 1), and can be adapted to a wider or different system by changing and expanding the syntax rules. It is self-evident that natural language can be accepted. In this sense, the method of the present invention has a wide field of application. For example, although not described in detail, even in the list 1, (29a) to (29e) indicate registration of a new concept using natural language (29b) and registration of a general relationship (29b).
a) or registration of a new fact (29d) or the like is possible.

Further, in the description of the embodiment, the system combined with the multimedia file has been described. However, it is needless to say that it is effective to use only the knowledge base system independently of the multimedia file.

[0116]

As described above, according to the present invention, a user can learn a formal language using a natural language without using a formal language which is difficult to learn, and from a descriptive search request recalled from an ambiguous memory. It is possible to extract desired information from within. Normally, our memories fade over time, but the memories that remain at that time are fragmentary and abstract. Alternatively, it is rare that the name is memorized, and it is also known that the name becomes descriptive such as "this kind" or "that kind". The method of the present invention provides a method of searching from a fragmentary or abstract description, and in that sense, has an effect more than simply searching in a natural language.

Further, according to the present invention, the knowledge necessary for understanding a natural language is stored together with the original knowledge of the system in the knowledge base. It is easier than the conventional method of providing a natural language interface. Specifically, the user does not need to be particularly conscious of the consistency between the natural language interface and the knowledge base, but only needs to pay attention to the registration of information that is to be stored and managed in the knowledge base.

Further, according to the present invention, it is possible to register new knowledge using a natural language. Typically,
In a system having a complex function, coexistence of a visual (direct operation) interface and a linguistic interface is important. In this method, a linguistic interface function is integrated with a visual interface, and almost all functions are provided by natural language and direct operation. In that sense, the user can select a favorite one on the spot, and is excellent in operability and usability.

Further, the present invention provides a means for registering an alias such as a synonym and a different notation for a concept, and for storing and registering an alias for a relation name ("reading"). In addition, it is possible to appropriately execute search processing even for expressions of various search requests.

The present invention also discloses a method of integrating multimedia information into a knowledge base, and according to this, it is possible to open up an application field beyond the conventional knowledge base using symbolic expressions.

Alternatively, if a foreign language system is constructed in accordance with the method described as an extended example, it is possible to realize, for example, a search in English and output in Japanese.

[Brief description of the drawings]

FIG. 1 is a functional block diagram of a natural language sentence interpretation program according to the present invention.

FIG. 2 is a conceptual network diagram illustrating a knowledge expression method using a conceptual relationship model.

FIG. 3 is a diagram showing knowledge storage based on the model, and showing concept names.

FIG. 4 is a diagram showing knowledge storage by the same model, showing a subsumption relationship.

FIG. 5 is a diagram showing knowledge storage by the same model, showing a general relationship definition.

FIG. 6 is a diagram showing knowledge storage by the same model and showing a relationship.

FIG. 7 is a diagram illustrating knowledge storage by the same model, and is a diagram for explaining the principle of a method of interpreting the meaning of a compound noun phrase.

FIG. 8 is a diagram illustrating knowledge storage by the model, and is a diagram illustrating a syntax analysis method.

FIG. 9 is a diagram illustrating knowledge storage by the model, and is a diagram illustrating a syntax analysis example.

FIG. 10 is a diagram showing knowledge storage by the model, and showing an example of syntax analysis.

FIG. 11 is a diagram illustrating knowledge storage by the model, and is a diagram illustrating a syntax analysis example.

FIG. 12 is a diagram showing knowledge storage by the same model, showing a preposition.

FIG. 13 is a diagram showing knowledge storage by the same model, showing a relation descriptor.

FIG. 14 is a diagram showing knowledge storage according to the same model, and is a hardware configuration diagram of a system according to an embodiment of the present invention.

[Explanation of symbols]

503: Vocabulary analysis unit, 504: Syntax analysis unit, 506: Compound noun analysis unit, 511: Structured concept file, 512 ...
Vocabulary dictionary, 514 ... syntax rule grammar file, 221 ... concept name table, 231 ... subsumption relation table, 241 ...
General relation definition table, 251,... Relation table, 601
... a preposition table, 611 ... a relation descriptor table.

Continuing from the front page (72) Inventor Itiko Kiuchi 1-280 Higashi Koikekubo, Kokubunji-shi, Tokyo Inside the Central Research Laboratory, Hitachi, Ltd. (56) References JP-A-59-223883 (JP, A) JP-A-61-220027 ( JP, A) JP-A-58-56071 (JP, A) JP-A-61-150592 (JP, A) JP-A-62-86429 (JP, A) JP-A-63-213066 (JP, A) Hiromichi Fujisawa In addition, "Intelligent filing system using concept network", Hitachi Review, vol. 69, no. 3 (March 1987) pp. 29-36 Hiromichi Fujisawa, Junichi Higashino, Atsushi Hatakeyama, "Conceptual Network Semantic Retrieval Method-Intelligent Filing (Part 5)-", Information Processing Society of Japan 31st ) National Convention Lecture Papers (II), pp. 1087-1088 (September 9, 1985) (58) Field surveyed (Int. Cl. ⁶ , DB name) G06F 17/30 G06F 17/28

Claims (3)

(57) [Claims] 1. An input means for inputting a natural language sentence including a plurality of words each having a concept, and a plurality of words having no relationship described therein, and a concept of the noun as a node, A memory that stores a knowledge base that expresses a relationship between concepts as a link and a concept network that forms a concept tree with nodes and links, and performs a semantic analysis process of the natural language sentence with reference to the knowledge base And an output device for outputting a processing result. When the noun composed of the plurality of words does not exist as one concept in the concept network, the concept network is searched for the noun, Obtaining a structured concept definition expression represented by a relational expression between the concepts of a plurality of words constituting the noun by search, and further storing in the memory,
A database device, wherein the semantic analysis process is performed by referring to the structured concept definition expression together with the knowledge base. 2. The knowledge base includes a concept name table storing names of the concepts, a subsumption relation table storing the subsumption relations between the concepts, and a definition of relations other than the subsumption relations existing between the concepts. 2. The database device according to claim 1, comprising a general relation definition table for storing, and a relation table for storing existence of a relation between the concepts. 3. The semantic analysis process includes a vocabulary analysis process for analyzing attributes of each word of the noun, a grammar file storing syntax rules, the knowledge base, and an analysis result of the vocabulary analysis process. 3. The database device according to claim 2, further comprising a syntax analysis process for analyzing the syntax of the noun.