JPH04100151A - Expert system string reference method - Google Patents

Abstract

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

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an expert system, and more particularly to a string reference method for an expert system that interacts with a user in the form of natural language.

[Prior Art] A conventional expert system consists of an input/output section, an inference section, and a knowledge base.

The input/output unit inputs factual data regarding the inference target in the form of natural language (hereinafter, the input factual data will be referred to as input event),
The inference results are displayed on the CRT.

The inference section uses an inference rule consisting of an IF section and a then section to search for an inference rule for which the proposition in the IF section holds true, and outputs the proposition in the then section of the inference rule as a conclusion.

Whether or not the proposition in the IF section holds true is determined by referring to factual data. Hereinafter, the propositions describing the IF part and the Then part will be referred to as the event and conclusion of the inference rule, respectively. Even if the inference rule holds the event, if the conclusion does not answer the problem to be solved. , and searches for the next inference rule using that conclusion as an input event. in this way,
An inference in which the conclusion obtained as a result of one inference is used as an input event and is connected to the next inference, and the inference is repeated until the conclusion that answers the problem, that is, the target conclusion is obtained, is called multistage inference, and is the conclusion of a certain inference rule. The proposition that is the input event for the next inference rule is called an intermediate hypothesis.

The knowledge base consists of a rule base and a database. A rule base is usually configured on a magnetic disk and stores inference rules.

The database stores factual data and intermediate hypotheses.

Inference refers to fact data or an intermediate hypothesis in the IF part of each inference rule, and the search results in a problem of searching for an inference rule having an IF part that matches the fact data or intermediate hypothesis.The search is performed using a predetermined search algorithm. The search algorithm is expressed in the form of a search tree with each inference rule as a contact point or a list with each inference rule as an element.

References between factual data or intermediate hypotheses and propositions in the IF section are made as follows.

Suppose that there are two inference rules as follows.

Rule (1): IF Weather == Sunny Then Mobile; Hat Rule (2): IF Weather:・Rain then
Cellphone = Umbrella In this case, the character strings of language elements that describe each inference rule are as follows.

Rule (1) Weather: Clear, Mobile Phone, Hat Rule (2)
: Weather Rain Mobile If all the factual data or the intermediate hypothesis is weather/rain, the inference section first refers to (weather clear) and (weather rain), and then (weather rain).
Refer to (rain) and (weather rain) and get a cell phone/umbrella as a conclusion. At this time, since each kanji character is specified by a 2-byte address, for example, (weather, clear) and (weather, rain) are referenced by referring to a 6-byte code one byte at a time.

[Problem to be Solved by the Invention] In the above-mentioned conventional expert system character string reference method, in order to check whether character strings ABC and character strings ABCD match, it is necessary to check all characters from the beginning. Since each byte has to be compared, not only does it take time to reference strings, but there is also a problem if the reference time varies depending on the length of the string.

The present invention provides that the reference time does not depend on the length of the character string, and
The purpose of this invention is to provide a method for referencing strings in an expert system that allows high-speed referencing.

[Means for Solving the Problems] A character string reference method for an expert system of the present invention is a character string reference method for an expert system that has rule-based knowledge and interacts with a user in the form of natural language, comprising: Strings of language elements that describe inference rules are stored in a predetermined memory in advance, and when initializing the inference rules to an expert system, the strings of language elements that describe inference rules are input to the expert system, and a predetermined search is performed. According to an algorithm, a character string that matches the input character string is searched from a set of character strings stored in the memory, and the first address of the address in the memory of the matching character string is generated as data, and an inference rule An inference rule is written using the above-mentioned start address corresponding to a character string of a language element that describes factual data, and at the time of inference, the character string of a language element that describes factual data is input to the expert system, and according to a predetermined search algorithm, the above-mentioned A language that searches for a character string that matches an input character string from a set of character strings stored in memory, generates the first address of the matching character string in the memory as data, and describes factual data. Fact data is described using the first address corresponding to the character string of the element, and during inference, the character string of the language element that describes the inference rule and fact data is referenced by referring to the first address.

[Operation] In the character string reference method of the expert system of the present invention, character strings of language elements that describe inference rules, that is, character strings of language elements that describe events and conclusions, and factual data, that is, characters of input events. column and the start address of the corresponding string in memory (hereinafter referred to as a token)
correspond to The token is different for each string, and
Since it is uniquely determined for each character string, this token can be used as an identification code for the corresponding character string. Therefore, during inference, reference between the conclusion or intermediate hypothesis and the event of the next inference rule is achieved by referring to the respective tokens.

Since the token is usually a 4-byte code that does not depend on the length of the string, referencing the string is usually completed by referencing only 4 bytes, regardless of the length of the string.

[Embodiments] Next, two embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of an expert system to which the character string referencing method of the expert system of the present invention is applied, and FIG. 2 is a text book showing a character string management format.

The expert system of this embodiment includes an input/output section 1, a reference section 2. Reasoning part 3. It is composed of knowledge base 4. Among these, input/output section 1. The inference unit 3 and the knowledge base 4 are [
It has the same functions and operates in the same way as those described in the section ``Prior Art''.

The reference unit 2 includes a memory in which a set of character strings of language elements describing inference rules is stored. This collection of character strings is managed in the format of a text book as shown in FIG. When character strings of language elements describing inference rules and factual data are input through the input/output unit 1, the reference unit 2
searches for a character string that matches the input character string from the set of character strings stored in the memory according to the search algorithm corresponding to the text book shown in FIG. (4 bytes), that is, 1.-kun is output as data. The token is used as an identification code for the input string.

As a result of searching the printed books on the memory, if there is no character string matching the input character string, a new character string identical to the input character string is added to the memory.

Newly added strings are given addresses to add nodes to already stored strings. When adding new strings to a text book, maximum care is taken to ensure that no two identical strings exist in the text book, ie, the strings are unique. This guarantee is achieved by using B-trees and hash functions.

Further, when the reference unit 2 determines that the input string is a string of a language element that describes an inference rule, the reference unit 2 stores the token of the input string on the rule base where the input string is to be stored. If the input string is determined to be a string of language elements describing factual data, the token of the input string is sent to the address on the database where the input string is to be stored. Transmit. In this way, rulebases and databases have
Inference rules and factual data expressed in tokens are written. Furthermore, since inference rules are expressed as tokens, intermediate hypotheses in multi-stage inference are also expressed as tokens and stored in the database.

Next, the operation of this embodiment will be explained.

First, when initializing the expert system, inference rules are written into the rule base.

When the reference unit 2 receives a character string of a language element of an inference rule via the input/output unit 1, it searches a character string tree and generates a token for the character string. Furthermore, when the reference unit 2 determines that the input character string is a character string of an inference rule, it transfers the token to the address in the rule base where the input character string is to be stored. In this way, inference rules written in tokens are stored in the rule base.

Next, at the time of inference, when the reference unit 2 receives a character string of a linguistic element of fact data via the input/output unit 1, it searches a character string tree and generates a token for the character string. Further, when the reference unit 2 determines that the input character string is factual data, it transfers the token to the address in the database where the input character string is to be stored. The inference section uses the database,
It reads fact data written in tokens, reads inference rules written in tokens from the rule base, and performs inference.

As described above, when the expert system guarantees the uniqueness of a string, the start address of the string is uniquely determined. The inference unit 3 uses the start address, that is, the token, as a unique identification code for the character string within the expert system, and proceeds with the inference by referring to the token. As a result, no matter how long the string is, the reasoning unit 3 can complete the reference to the string by simply comparing 4-byte tokens.

In the above embodiment, a case has been described in which the reference unit 2 manages a set of character strings in memory using the management format of this structure, but as another embodiment, it is also possible to manage character strings using a management format of a list structure. can. What is important is the uniqueness of the strings, not the management format of the string set. This is because if the uniqueness of character strings is not guaranteed, a match between character strings may be overlooked.

[Effects of the Invention] As explained above, the present invention makes each character string describing an inference rule and fact data correspond to the start address of the character string in memory, and associates the start address with the identification code of the character string. By referencing a character string as , there is an effect that the character string can be referenced regardless of the length of the character string and faster than in the past.

[Brief explanation of drawings]

FIG. 1 is a block diagram of an embodiment of an expert system to which the expert system character string reference method of the present invention is applied, and FIG. 2 is a character string tree showing a character string management format. 1... Input/output device, 2... Reference section, 3... Reasoning section, 4... Knowledge base.

Claims (1)

[Claims] 1. In a string reference method for an expert system that has rule-based knowledge and interacts with a user in the form of natural language, a string of language elements describing inference rules is stored in a predetermined memory in advance. When initializing the inference rules in the expert system, the strings of language elements that describe the inference rules are input to the expert system, and according to a predetermined search algorithm, the strings are extracted from the set of strings stored in the memory. , finds a string that matches the input string, generates the starting address of the matching string in the memory as data, and generates the starting address corresponding to the string of the language element that describes the inference rule. At the time of inference, input strings of language elements that describe factual data into the expert system, and select input strings from the set of strings stored in the memory according to a predetermined search algorithm. Find a character string that matches, generate the start address of the matching character string in the memory as data, and write the fact data using the start address corresponding to the character string of the language element that describes the fact data. A method for referring to strings in an expert system, characterized in that, during inference, a string of a language element that describes an inference rule and fact data is referenced by referring to the start address.