CN115526319A - Method for realizing expert system based on multi-source knowledge - Google Patents

Abstract

An implementation method of an expert system based on multi-source knowledge, by constructing a knowledge base, parsing the knowledge from different sources and converting it into a format that the knowledge base can recognize, then collecting question input and performing domain judgment on the input question, and Use the domain method to solve the problem; then conduct reasoning observation on the input problem, and according to the similar matching, query the matching rules or functions, according to the rules and functions obtained from the similar matching, further complete matching, generate reasoning results, and reasoning results Add to the knowledge base, delete redundant, high-probability useless candidate inference results through inference shrinkage, and output the final result of inference engine inference. The invention greatly enhances the versatility of the knowledge base; greatly broadens the source of knowledge; and greatly enhances the ability of knowledge reasoning.

Description

Realization Method of Expert System Based on Multi-source Knowledge

technical field

The present invention relates to a technology in the field of expert systems, in particular to a method for realizing an expert system based on multi-source knowledge.

Background technique

An expert system is an intelligent computer program system, which contains a large amount of expert knowledge and experience in a certain field. It can apply artificial intelligence technology and computer technology to reason and judge according to the knowledge and experience in the system, simulating human In short, an expert system is a computer program system that simulates human experts to solve domain problems.

An expert system usually consists of six parts: human-computer interaction interface, knowledge base, inference engine, interpreter, comprehensive database, and knowledge acquisition. Among them, the separation of knowledge base and inference engine is unique. The architecture of an expert system varies with the type, function, and scale of the expert system.

Since the first expert system DENDRAL came out at Stanford University in the United States in 1965, after 50 years of development, various expert systems have spread across various professional fields, involving industry, agriculture, military and various departments of the national economy and many aspects of social life. aspect.

Early expert systems such as the DENDRAL system (1968, Feigenbaum et al., Stanford University) were used to infer chemical molecular structures, and the MYCSYMA system (1971, Massachusetts Institute of Technology) was used for mathematical operations. It is characterized by a high degree of specialization and a strong ability to solve special problems, but its structure and functions are incomplete, its portability is poor, and it lacks an explanation function.

Later, the expert system entered a mature stage, such as the MYCIN system (Stanford University) for the diagnosis of blood infectious diseases, the PROSPECTOR system (Stanford Research Institute) for prospecting, and the CASNET system (Rutger University) for glaucoma diagnosis and treatment. , AM system (1981, Stanford University) is used to simulate human beings for generalization, abstraction and inductive reasoning, and discovers some concepts and theorems of number theory, and HEARSAY system (Carnegie Mellon University) is used for speech recognition expert systems. The characteristics of these systems are single-disciplinary professional expert systems; the system has complete structure, comprehensive functions, and good portability; it has reasoning and explanation functions and good transparency; it uses heuristic reasoning and imprecise reasoning; Network representation of knowledge; Human-Computer Interaction in Qualified English.

From the 1980s to the present, the expert system has entered a further development period, and XCON (DEC Corporation, Carnegie-Mellon University) has appeared to formulate hardware configuration solutions for the VAX computer system, as well as expert system development tools, such as the skeleton system ( EMYCIN, KAS, EXPERT, etc.), general-purpose knowledge expression language (OPS5, etc.), expert system development environment (AGE, etc.).

At present, my country has also developed an expert system for fertilization (Hefei Institute of Intelligent Machinery, Chinese Academy of Sciences), an expert system for finding water in new structures (Nanjing University), an expert system for exploration and oil and gas resource evaluation (Jilin University), an expert system for clothing tailoring and Expert systems for calico pattern design (Zhejiang University), Guan Youbo Liver Disease Diagnosis Expert System (Beijing College of Traditional Chinese Medicine) and other expert systems.

Although expert systems have been widely used in various fields and have received good results, their range of problem solving is often limited, mainly because of: ① lack of knowledge; ② inappropriate methods to solve problems. Moreover, most of the expert systems are established for a specific field, and once they go beyond this specific field, the system may no longer be able to operate effectively.

Contents of the invention

Aiming at the above-mentioned deficiencies in the prior art, the present invention proposes an implementation method of an expert system based on multi-source knowledge, which greatly enhances the versatility of the knowledge base, greatly broadens the knowledge sources, and greatly enhances the knowledge reasoning ability.

The present invention is achieved through the following technical solutions:

The invention relates to a realization method of an expert system based on multi-source knowledge, by constructing a knowledge base, analyzing and converting the knowledge from different sources into a format that can be recognized by the knowledge base, and then collecting and inputting questions and performing field analysis on the input questions Judgment, and use the domain method to solve the problem; then conduct reasoning and observation on the input problem, and query the matching rules or functions according to the similar matching, and further perform a complete matching, generate reasoning results, and Add the reasoning results to the knowledge base, delete redundant and useless candidate reasoning results through reasoning shrinkage, and output the final result of reasoning machine reasoning.

The present invention relates to a system for implementing the above method, including: a knowledge parser, a knowledge base, a reasoning engine, a user interface, and an output generator, wherein: the knowledge parser performs corresponding parsing processing according to different types of input knowledge, And transform the analysis results into the knowledge base; the knowledge base uses a knowledge model with general-purpose knowledge expression ability to store knowledge; the reasoning engine performs reasoning according to the questions input by the user, and obtains the reasoning results (the reasoning may not be successful failure result); the user interface is responsible for interacting with the user, and the interactive mode is the command line; the output generator outputs the object specified by the user into the format specified by the user according to the instruction input by the user.

technical effect

The present invention uses knowledge map + rules + functions as basic elements to realize a general-purpose knowledge base, uses various existing knowledge extraction and knowledge analysis technologies to analyze knowledge from various sources, and integrates the analysis results into the general-purpose knowledge base; The reasoning of the reasoning machine is divided into simple reasoning and complex reasoning. Simple reasoning uses knowledge map matching technology to find solutions to problems, and complex reasoning uses the cyclic process of reasoning observation-reasoning expansion-reasoning contraction to try to find solutions to problems.

Compared with the prior art, the present invention supports multiple knowledge sources, solves the problem that the previous expert system can only resolve a single knowledge source, and enhances the ease of use and scalability of the expert system; the use of complex reasoning breaks the In the past, the reasoning machine of the expert system was limited by the domain, so that some problems that were not covered by the algorithm pre-designed by the expert system could be automatically solved by the reasoning machine.

Description of drawings

Figure 1 is a flowchart of multi-source knowledge analysis;

Fig. 2 is a flow chart of the core of the inference engine;

Fig. 3 is an expert system architecture diagram;

Fig. 4 is a logical item structure diagram;

Figure 5 is a flowchart of the use of the expert system.

detailed description

As shown in Figure 1, this embodiment relates to an implementation method of an expert system based on multi-source knowledge, comprising the following steps:

Step 1) Build Knowledge Base

The knowledge base includes: a main knowledge map, a rule library, and a function library, wherein: the main knowledge map is a graph composed of nodes and relationships, and is used to describe a large number of triplet relationships (A, Relation, B); the rules The library consists of a large number of rules, each of which is a proposition, which may or may not have parameters. A rule without parameters is in the form of IFATHEN B, and a rule with parameters is in the form of Foreach(x, y, z): IFATHEN B, where A and B are each a logical AND item (see Figure 4). Each logical AND item consists of at least one small knowledge graph and zero or several logical OR items, and each logical OR item consists of at least one logical AND item; the function library consists of a large number of functions. Each function consists of input parameters, local variables, and a sequence of instructions. Each instruction is a control instruction (such as return, conditional execution), a basic instruction (such as addition, subtraction, multiplication and division), and a call to other functions. The default execution mode of the function is to execute the function body in an infinite loop, and the return instruction jumps out of the execution of the function body. In addition, each function and each instruction is allowed to set the start property and end property, both of which are a logical AND item, used to describe the start state and end state of the corresponding function or instruction.

Step 2) Analyzing the knowledge from the knowledge map, graph database, and relational database and converting it into a format that can be recognized by the knowledge base, which is convenient for the expert system to analyze and process, specifically including:

2.1) For the data of the knowledge map and the graph database, directly analyze the data, read out all the triple relations (A, Relation, B), and then add these triple relations to the knowledge map of the knowledge base.

2.2) For the data of the relational database, the data is analyzed first, and all the relational records Ri(e1, e2,...,eN) are read out. For a record of the form R(A), convert it to a triple relation of the form (A,R,null). For a record of the form R(A,B), convert it to a triplet relation of the form (A,R,B). For a longer record R(e1,e2,...,eN), convert it to a set of the form (e1,R,e2), (e1,R,e3),...,(e1,R,eN) triplet relationship. Then add all the above-converted triplet relationships to the knowledge map of the knowledge base.

Step 3) parse the knowledge from XML and JSON and convert it into a format that can be recognized by the knowledge base, which is convenient for the expert system to analyze and process, specifically including:

3.1) For the XML format, first use DOM (Document Object Model) technology to parse the XML file into a DOM tree, and then identify the nodes of the DOM tree: the knowledge map node (<kg>) in the DOM tree according to the knowledge map in step 2.1 The parsing method parses into a triplet relationship and adds it to the knowledge map of the knowledge base. The rule node (<rule>) in the DOM tree is parsed into a rule in the knowledge base according to the mark of its child node, and the function node in the DOM tree is (<func>) resolves to functions in the knowledge base according to the tokens of its child nodes.

3.2) For the JSON format, first use the finite state machine to perform lexical analysis on the JSON file, and then perform top-down grammatical analysis on the result obtained by the lexical analysis to obtain the JSON object tree. Then identify the nodes of the object tree: parse the knowledge map object ({kg:...}) in the object tree into a triplet relationship according to the analysis method of the knowledge map in step 2.1 and add it to the knowledge map of the knowledge base, and add the object tree The rule object ({rule:…}) is parsed into rules in the knowledge base according to the key-value pairs of its sub-objects, and the function object ({func:…}) in the object tree is parsed according to the key-value pairs of its sub-objects into Functions in the knowledge base.

Step 4) parse the knowledge from Lisp and Prolog and convert it into a format that can be recognized by the knowledge base, which is convenient for the expert system to analyze and process, specifically including:

4.1) For the Lisp format, first use the finite state machine to perform lexical analysis on the Lisp file, and then perform top-down grammatical analysis on the result obtained from the lexical analysis to obtain the Lisp abstract syntax tree. Then the Lisp abstract syntax tree is parsed, and the Lisp functions are converted into functions in the function library of the knowledge base of the present invention.

4.2) For the Prolog format, first use the finite state machine to perform lexical analysis on the Prolog file, and then perform top-down grammatical analysis on the result obtained by the lexical analysis to obtain the Prolog abstract syntax tree. Then the Prolog abstract syntax tree is analyzed, and the Prolog rules are converted into rules in the rule base of the knowledge base of the present invention.

Step 5) collection problem input: a problem is to prove a proposition (its form is consistent with the rule in the rule base of the knowledge base of the present invention), that is, a proof problem; for a given group of input parameters and a logical AND item, It is a query problem to find a number of existing entities (Entities) in the knowledge base that match the parameters one by one, and satisfy the input logical AND items to be inferred by the current knowledge base after parameter replacement. Give an example to illustrate a proof problem. The problem is Foreach(x)IF x is an integer THEN x is a rational number. When there is enough knowledge (knowledge graph, rules, functions) in the knowledge base, it should be able to draw a conclusion—the proposition is established. To illustrate a query problem, the problem is Select(a,b):(a,capital_is,b), when there is appropriate knowledge in the knowledge base, several solutions should be found.

For example: (a, b) = (China, Beijing), namely (China, capital_is, Beijing). Specifically, how to solve the input problem requires the following steps.

Step 6) Domain judgment on the input problem: According to the problem input by the user, traverse the rule base and function library in the knowledge base to determine whether the problem is solvable by a domain-specific method. The judgment method is to completely match the input question according to the conditional clause of the rule (IF-clause, which is a logical AND item) or the initial nature of the function (also a logical AND item) (that is, replace the logical AND item with an existing entity Parameter entity in , and then calculate whether the parameter logical AND item is included in the existing logical AND item). When the matching is successful, it means that a certain rule or a certain function in the knowledge base is the solution to the problem, that is, the solution to the problem can be obtained by directly using the corresponding rule or directly running the corresponding function. This situation is called simple inference. When no match can be found, it enters the circular reasoning process of the reasoning machine, which is called complex reasoning.

Step 7) Use domain methods to solve the problem: According to the result of step 6, the problem input by the user is handed over to the corresponding matching domain-specific method to solve, that is, use the corresponding rules to deduce or run the corresponding functions.

For example: when the user requests to solve a quadratic equation of one variable QuadEq(3,4,5), the rule library and function library of the matching knowledge base are traversed, and when the function SolveQuadEq(a,b,c) for solving the quadratic equation of one variable is found , call this function to solve the input quadratic equation, that is, call SolveQuadEq(3,4,5).

Step 8) Perform reasoning observation on the input question, and query matching rules or functions according to similar matching. Since step 2 determines that there is no rule or method in the knowledge base that can completely match the input question, the reasoning machine can continue to try reasoning in order to make progress.

For example: if there is no rule or function for solving a quadratic equation in one variable in the knowledge base, then it can also be solved by transposition and formula; when the coefficient of the quadratic term is observed to be 0, then the quadratic term can be eliminated, and then the unary equation can be searched The solution of a linear equation.

Step 9) According to the rules and functions obtained by similar matching in step 4, further perform complete matching, generate inference results, and add the inference results to the knowledge base.

For example: when solving a quadratic equation with one variable can be formulated, this step is to completely match the corresponding rules or functions of the formula, and then implement the corresponding formula conversion for the formula, and add the new formula after the formula to the knowledge base. Since each parameter is independent, the scale of complete matching may be very large (N1*N2*...Nm, Ni is the number of candidate entities for the i-th parameter), and the matching time and the number of matching results are both combination-level growth Therefore, it is necessary to manually set the threshold. The thresholds set by the present invention are the maximum number of parameters for a single rule/function (6), the maximum number of candidate entities for a single parameter (100), and the maximum number of successful matching results for a single match (100).

Step 10) Delete redundant and highly probable useless inference results through inference shrinkage, so as to prevent the inference engine from quickly exhausting system resources and stagnant inference progress. This step performs similar matching on the generated entities and relationships. When the relationships and relationship objects owned by several entities (threshold value of the present invention is 10) are similar, then these entities are likely to be redundant, and only one of them can be retained. One. For the generated rules, the IF clauses and THEN clauses are completely matched in pairs. When a rule is completely a subset of another rule, the weaker rule is deleted. For the generated functions, the start and end properties are completely matched in pairs. When the properties of a function are completely covered by another function, the weaker function is deleted.

Step 11) output the final result of the reasoning machine reasoning, that is, output the result calculated by using rules or calling functions in step 3, the result output by observing in step 4 that the problem has been solved, or the observation in step 4 that the system time resource or space resource has been blocked Exhausted and can no longer free up system resources through speculative shrinking, thus outputting error messages. The output mode of the final result can be any input format of knowledge source, such as knowledge map, XML, Lisp language, etc.

This embodiment relates to the expert system of the above method, including: a knowledge parser, a knowledge base, an inference engine, a user interface, and an output generator, wherein: the knowledge parser performs corresponding parsing processing according to different types of input knowledge, and Transform the analysis results into the knowledge base; the knowledge base uses a knowledge model with general knowledge expression ability to store knowledge; the reasoning engine performs reasoning according to the questions input by the user, and obtains the reasoning results; the user interface is responsible for interacting with users. The method is the command line; the output generator outputs the object specified by the user into the format specified by the user according to the instruction input by the user.

After specific practical experiments, running the above-mentioned expert system under the Windows platform can successfully read and analyze knowledge in formats such as knowledge graphs, relational databases, graph databases, XML, JSON, Lisp, Prolog, etc., and add them to the knowledge base, and the knowledge base outputs The results show that the content of the knowledge base is in line with expectations; the user's questions are input in JSON format, and it can be found that the results of simple query problems are in line with expectations (query success or failure), and some of the complex query problems can also run successfully. Since the threshold is set small, complex query problems have a very high failure rate as the problem size increases.

Compared with the prior art, the method greatly enhances the versatility of the knowledge base; greatly broadens the source of knowledge; greatly enhances the ability of knowledge reasoning.

The above specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The scope of protection of the present invention is subject to the claims and is not limited by the above specific implementation. Each implementation within the scope is bound by the invention.

Claims (4)

1. An implementation method of an expert system based on multi-source knowledge, characterized in that by constructing a knowledge base and analyzing and converting the knowledge from different sources into a format that the knowledge base can recognize, then collecting question input and inputting The problem is judged in the domain, and the domain method is used to solve the problem; then the input problem is reasoned and observed, and according to the similar matching, query matching rules or functions, and the rules and functions obtained from the similar matching, further complete matching and generative reasoning Results, and add the reasoning results to the knowledge base, delete redundant and useless candidate reasoning results through reasoning shrinkage, and output the final result of reasoning machine reasoning. 2. The method for realizing an expert system based on multi-source knowledge according to claim 1, wherein said knowledge base comprises: a master knowledge graph, a rule base, and a function library, wherein: the master knowledge graph is a node composed of A graph composed of and relations is used to describe a large number of triple relations (A, Relation, B); the rule base is composed of a large number of rules, and each rule is a proposition, where: A and B are each a logical AND item, Each logical AND item consists of at least one small knowledge graph and zero or several logical OR items, and each logical OR item consists of at least one logical AND item; the function library consists of a large number of functions; each function consists of input parameters, local variables Consists of a series of instructions, each instruction is a control instruction, a basic instruction, or a call to other functions; the default execution mode of the function is to execute the function body in an infinite loop, and the return instruction jumps out of the execution of the function body; each function, each Each instruction allows to set the start property and end property, which are both a logical AND item, which is used to describe the start state and end state of the corresponding function or instruction. 3. the method for realizing the expert system based on multi-source knowledge according to claim 1 or 2, is characterized in that, specifically comprises: Step 1) Build the knowledge base: Step 2) Parse and convert the knowledge from the knowledge map, graph database, and relational database into a format that the knowledge base can recognize, specifically including: 2.1) For the data of the knowledge map and the graph database, directly analyze the data, read out all the triplet relationships (A, Relation, B), and then add these triplet relationships to the knowledge map of the knowledge base; 2.2) For the data in the relational database, first analyze the data and read out all the relational records Ri(e1,e2,...,eN), specifically: a) For the record of R(A), convert it into a triple relation of (A, R, null); b) For the record of R(A,B), convert it into a triple relation of (A,R,B); c) For a longer record R(e1,e2,...,eN), transform it into a set of (e1,R,e2), (e1,R,e3),...,(e1,R,eN) The triplet relationship; Then add all the above-converted triplet relationships to the knowledge map of the knowledge base; Step 3) parse the knowledge from XML and JSON and convert it into a format that the knowledge base can recognize, specifically including: i) For the XML format, first use DOM technology to parse the XML file into a DOM tree, and then identify the nodes of the DOM tree: parse the knowledge map node (<kg>) in the DOM tree into three parts according to the parsing method of the knowledge map in step 2.1 The tuple relationship is added to the knowledge map of the knowledge base, the rule node (<rule>) in the DOM tree is parsed into a rule in the knowledge base according to the mark of its child node, and the function node (<func>) in the DOM tree is Parsed into functions in the knowledge base according to the tags of its child nodes; ii) For the JSON format, first use the finite state machine to perform lexical analysis on the JSON file, and then perform top-down grammatical analysis on the results obtained from the lexical analysis to obtain the JSON object tree; then identify the nodes of the object tree: the object tree in the The knowledge graph object ({kg:…}) is parsed into a triplet relationship according to the analysis method of the knowledge graph in step 2.1 and added to the knowledge graph of the knowledge base, and the rule object ({rule:…}) in the object tree is The key-value pairs of sub-objects are parsed into rules in the knowledge base, and the function objects ({func:...}) in the object tree are parsed into functions in the knowledge base according to the key-value pairs of their sub-objects; Step 4) parse the knowledge from Lisp and Prolog and convert it into a format that the knowledge base can recognize, specifically including: α) For the Lisp format, first use the finite state machine to perform lexical analysis on the Lisp file, and then perform top-down grammatical analysis on the result obtained from the lexical analysis to obtain the Lisp abstract syntax tree; then analyze the Lisp abstract syntax tree, and convert the Lisp The function translates to a function in the function library of the knowledge base of the present invention; β) For the Prolog format, first use the finite state machine to perform lexical analysis on the Prolog file, and then perform top-down grammatical analysis on the result obtained from the lexical analysis to obtain the Prolog abstract syntax tree; then analyze the Prolog abstract syntax tree, and convert the Prolog The rules are transformed into rules in the rule base of the knowledge base of the present invention; Step 5) collecting question input; Step 6) Domain judgment on the input problem: According to the problem input by the user, traverse the rule library and function library in the knowledge base to determine whether the problem is solvable by a domain-specific method; the judgment method is based on the conditional clause or function of the rule The initial nature, to fully match the input question, that is, to replace the parameter entity in the logic and item with the existing entity, and then calculate whether the parameter logic and item is included in the existing logic and item; when the match is successful, it indicates that the knowledge base A certain rule or a certain function is the solution to the problem, that is, the corresponding rule can be directly used for derivation, or the corresponding function can be directly run to obtain the solution to the problem. When no match can be found, it will enter the circular reasoning process of the reasoning machine; Step 7) Use the domain method to solve the problem: according to the result of the judgment in step 2, the problem input by the user is handed over to the corresponding matching domain-specific method to solve, that is, use the corresponding rule to deduce, or run the corresponding function; Step 8) Perform inference observation on the input question, and query matching rules or functions according to similar matching; Step 9) According to the rules and functions obtained by similar matching in step 4, further complete matching, generate inference results, and add inference results to the knowledge base; Step 10) delete redundant reasoning results through reasoning contraction, including: 10.1) Perform similar matching on the generated entities and relationships, and when the relationships and their relationship objects owned by several entities are similar, only one of them will be kept; 10.2) For the generated rules, the IF clauses and THEN clauses are completely matched in pairs, and when a rule is completely a subset of another rule, the weaker rule is deleted; 10.3) For the generated functions, the start and end properties are completely matched in pairs, and when the properties of a function are completely covered by another function, the weaker function is deleted; Step 11) output the final result of the reasoning machine reasoning, that is, output the result calculated by using rules or calling functions in step 3, the result output by observing in step 4 that the problem has been solved, or the observation in step 4 that the system time resource or space resource has been blocked Exhausted and can no longer free up system resources through speculative shrinking, thus outputting error messages. 4. according to the realization method of the expert system based on multi-source knowledge described in any one in claim 1-3, it is characterized in that, described expert system comprises: knowledge parser, knowledge base, reasoning engine, user interface and The output generator, wherein: the knowledge parser performs corresponding parsing processing according to different types of input knowledge, and transforms the parsing results into the knowledge base; the knowledge base uses a knowledge model with general knowledge expression ability to store knowledge; reasoning The engine performs inference according to the questions input by the user, and obtains the inference result; the user interface is used to interact with the user through the command line; the output generator outputs the object specified by the user into the format specified by the user according to the instruction input by the user.

Images