Knowledge reasoning and evidence generating method based on problem decomposition

Abstract

The invention provides a knowledge reasoning and evidence generating method based on problem decomposition, and belongs to the technical field of problem decomposition. The knowledge reasoning and evidence generation method based on question decomposition comprises question analysis, knowledge base and data, reasoning engine, evidence generation and answer production, learning and feedback, and comprises natural language processing, question classification and question understanding, the question analysis uses an NLP technology to analyze and understand questions proposed by users, and the NLP technology is used to analyze and understand the questions proposed by the users; the method comprises the following steps: determining the field or type of a question, decomposing the complex question into a series of simpler sub-questions by using a comparator module to complete relation analysis to generate a final answer, and outputting the answering process of each sub-question as evidence to display the reasoning logic of the model to answer the question, so as to realize the high consistency of the evidence and a decision; the inconsistency of the questions and the answers is ensured, and the accuracy of the answers is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of problem decomposition, and in particular to a knowledge reasoning and evidence generation method based on problem decomposition. Background Art

[0002] With the popularization of human-computer intelligent interaction, human-computer language intelligent interaction has gradually entered people's daily life and production. When a problem arises, it is necessary to decompose the problem, produce evidence for the decomposed sub-problems, and synthesize the answers. However, inconsistencies between questions and answers often occur, which reduces the accuracy of the answers. Summary of the invention

[0003] In order to remedy the above deficiencies, the present invention provides a knowledge reasoning and evidence generation method based on problem decomposition, aiming to improve the problem of frequent inconsistency between questions and answers, which reduces the accuracy of answers.

[0004] The present invention is achieved in that: The present invention provides a knowledge reasoning and evidence generation method based on problem decomposition, comprising the following steps: S1: Question parsing, including natural language processing, question classification and question understanding, uses NLP technology to analyze and understand the questions raised by users, determines the field or type to which the questions belong, and decomposes complex questions into a series of simpler sub-questions; S2: Knowledge base and data, building or leveraging existing knowledge bases, using relational databases and knowledge graphs, and leveraging existing texts, articles, and web content; S3: Reasoning engine, including logical reasoning, probabilistic reasoning, reasoning from basic principles and machine learning, using logical rules and formal methods to derive new information or conclusions, using probability theory to deal with uncertainty, using predefined rules to deal with problems, and using machine learning models to identify patterns and make predictions; S4: Evidence generation, including evidence extraction, evidence scoring and evidence integration, extracting evidence related to the sub-question from the knowledge base and data source, scoring the extracted evidence to determine its relevance and reliability. The scoring can be based on multiple factors, including the source, time, authority, etc. of the evidence, integrating multiple pieces of evidence into a coherent whole; S5: Answer production, including answer synthesis, answer verification and answer presentation, integrating the answers to the various sub-questions to form a complete answer to the original question, which can be achieved by verifying whether the generated answers are accurate, consistent and coherent, and presenting the final answer to the user in an appropriate form; S6: Learning and Feedback, including user feedback and continuous learning, collecting user feedback to evaluate the quality of answers and system performance, and continuously updating and optimizing the knowledge base and reasoning engine based on user feedback and new data.

[0005] Preferably, the natural language processing in step S1 includes word segmentation, part-of-speech tagging, named entity recognition, dependency syntactic analysis and semantic role labeling. Part-of-speech tagging is to segment a sentence into words or phrases, part-of-speech tagging is to mark the part of speech of each word, including nouns, verbs and adjectives, named entity recognition is to identify named entities in a sentence, including names of people, places and organizations, dependency syntactic analysis is to analyze the dependency relationship between words in a sentence, and semantic role labeling is to identify the relationship between a predicate and its arguments in a sentence.

[0006] Preferably, the problem understanding in step S1 includes syntactic and semantic analysis, template matching, and rules and heuristics. Syntactic and semantic analysis is to identify key components and sub-problems in the problem. Template matching is to use predefined templates to match the structure of the problem and decompose it into sub-problems. Rules and heuristics are to use rules and heuristic methods to identify and decompose problems.

[0007] Preferably, the knowledge base and data in step S2 include a knowledge base and a data source, wherein the establishment of the knowledge base includes knowledge acquisition, knowledge representation, knowledge organization and knowledge updating. Knowledge acquisition can be done by experts manually inputting knowledge, or by using natural language processing technology to automatically extract knowledge from text, or by importing data from existing databases or data sources.

[0008] Preferably, the logical reasoning in step S3 includes propositional logic, predicate logic and modal logic, where propositional logic processes simple truth propositions, predicate logic is used to process more complex statements and reasoning, and modal logic is used to process possibility and necessity.

[0009] Preferably, the probabilistic reasoning in step S3 includes a Bayesian network, a Markov model and a factor graph, wherein the Bayesian network represents the conditional dependency between variables, the Markov model is used to describe the transition probability between states, and the factor graph is used to represent the factor decomposition of the probability distribution.

[0010] Preferably, the evidence extraction in step S4 includes text mining, data query and knowledge graph query, wherein text mining is to extract relevant information from text, data query is to query relevant information from a database or API, and knowledge graph query is to query relevant information from a knowledge graph.

[0011] Preferably, the evidence integration in step S5 includes natural language generation, template filling, and rules and heuristics, natural language generation is to convert evidence into natural language text, template filling is used to fill in evidence using predefined templates, and rules and heuristics are based on using rules and heuristic methods to integrate evidence.

[0012] Preferably, the answer synthesis in step S5 includes information extraction, information integration, retrieval-based methods, generation-based methods and hybrid methods, extracting relevant information from multiple sources. This can be done through information retrieval and natural language processing technology, integrating the extracted information, removing duplicate and conflicting information, retaining the most relevant and accurate information, extracting answers directly from the retrieved information, finding the answer that best matches the question from the FAQ library, using a generative model, and generating answers. This method can generate more natural and coherent text, combining the two methods of retrieval and generation to improve the accuracy and coherence of the answer, first finding relevant information through retrieval, and then generating the final answer through a generative model.

[0013] Preferably, the continuous learning in step S6 includes machine learning, reinforcement learning and knowledge base updating, where machine learning is to use machine learning models to optimize the reasoning process and evidence generation, reinforcement learning is to learn the optimal strategy through a reward mechanism, and knowledge base updating is to regularly update the knowledge base to reflect the latest information and knowledge.

[0014] The beneficial effect of the present invention is as follows: the knowledge reasoning and evidence generation method based on problem decomposition obtained by the present invention through the above design adopts the method of problem parsing using NLP technology to analyze and understand the questions raised by users, determine the field or type to which the questions belong, decompose complex problems into a series of simpler sub-problems, build a knowledge base, and use the knowledge base to perform logical reasoning, probabilistic reasoning, reasoning based on basic principles and machine learning processing on multiple sub-problems to produce evidence, integrate multiple evidences to form a coherent whole, integrate the answers to each sub-problem to form a complete answer to the original question, which can be achieved by the following method: verifying whether the generated answers are accurate, consistent and coherent, presenting the final answer to the user in an appropriate form, the comparator module completing the relationship analysis to generate the final answer, and outputting the solution process of each sub-problem as evidence to demonstrate the reasoning logic of the model in solving the problem, achieving a high degree of consistency between evidence and decision-making, ensuring the inconsistency between the question and the answer, and improving the accuracy of the answer. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Example

[0016] The present invention provides a knowledge reasoning and evidence generation method based on problem decomposition, comprising the following steps: S1: Question parsing, including natural language processing, question classification and question understanding, uses NLP technology to analyze and understand the questions raised by users, determines the field or type to which the questions belong, and decomposes complex questions into a series of simpler sub-questions; S2: Knowledge base and data, building or leveraging existing knowledge bases, using relational databases and knowledge graphs, and leveraging existing texts, articles, and web content; S3: Reasoning engine, including logical reasoning, probabilistic reasoning, reasoning from basic principles and machine learning, using logical rules and formal methods to derive new information or conclusions, using probability theory to deal with uncertainty, using predefined rules to deal with problems, and using machine learning models to identify patterns and make predictions; S4: Evidence generation, including evidence extraction, evidence scoring and evidence integration, extracting evidence related to the sub-question from the knowledge base and data source, scoring the extracted evidence to determine its relevance and reliability. The scoring can be based on multiple factors, including the source, time, authority, etc. of the evidence, integrating multiple pieces of evidence into a coherent whole; S5: Answer production, including answer synthesis, answer verification and answer presentation, integrating the answers to the various sub-questions to form a complete answer to the original question, which can be achieved by verifying whether the generated answers are accurate, consistent and coherent, and presenting the final answer to the user in an appropriate form; S6: Learning and Feedback, including user feedback and continuous learning, collecting user feedback to evaluate the quality of answers and system performance, and continuously updating and optimizing the knowledge base and reasoning engine based on user feedback and new data.

[0017] In this embodiment, the natural language processing in step S1 includes word segmentation, part-of-speech tagging, named entity recognition, dependency syntactic analysis and semantic role labeling. Part-of-speech tagging is to divide a sentence into words or phrases, part-of-speech tagging is to mark the part of speech of each word, including nouns, verbs and adjectives, named entity recognition is to identify named entities in a sentence, including names of people, places and organizations, dependency syntactic analysis is to analyze the dependency relationship between words in a sentence, and semantic role labeling is to identify the relationship between a predicate and its arguments in a sentence.

[0018] In this embodiment, the problem understanding in step S1 includes syntactic and semantic analysis, template matching, and rules and heuristics. Syntactic and semantic analysis is to identify the key components and sub-problems in the problem. Template matching is to use predefined templates to match the structure of the problem and decompose it into sub-problems. Rules and heuristics are to use rules and heuristic methods to identify and decompose the problem.

[0019] In practical applications, syntactic and semantic analysis are often used together to gain a deeper understanding, for example: Question answering system: understands the structure of the question through syntactic analysis, and then determines the specific meaning of the question through semantic analysis, thereby generating accurate answers; Machine translation: understand the structure of source language sentences through syntactic analysis, understand the meaning of sentences through semantic analysis, and finally generate the translation into the target language; Text summarization: extract key information from text through syntactic and semantic analysis to generate concise summaries; Dialogue system: Understands the user's intent through syntactic and semantic analysis, and generates coherent and relevant responses. In this embodiment, the knowledge base and data in step S2 include a knowledge base and a data source, wherein the establishment of the knowledge base includes knowledge acquisition, knowledge representation, knowledge organization and knowledge update. Knowledge acquisition can be done by experts manually inputting knowledge, or by using natural language processing technology to automatically extract knowledge from text, or by importing data from an existing database or data source. Knowledge representation is a formalized knowledge representation method that defines concepts, relationships and axioms, uses logical rules to represent knowledge, and uses a framework structure to represent knowledge. Knowledge organization consists of classification, indexing and linking. Classification is to organize knowledge according to a certain classification system, indexing is to establish an index for the knowledge base to facilitate quick retrieval, and linking is to establish links between entities in the knowledge base to form a knowledge network.

[0020] Knowledge representation methods: Logical representation: Propositional Logic: deals with simple truth-valued propositions - true or false.

[0021] Predicate logic: Expands propositional logic and introduces individuals, predicates, and quantifiers.

[0022] Modal logic: Modal operators (such as "must" and "may") are introduced to deal with possibility and necessity.

[0023] The framework represents: Frame: A structured representation method that uses frames and slots to represent knowledge. For example, a "Person" frame may contain slots such as "Name", "Age", "Occupation", etc.

[0024] Script: A framework that describes a typical sequence of events, such as the script of "going to a restaurant for dinner". Semantic Network: Semantic network: A graph structure where nodes represent entities or concepts and edges represent relationships between them, e.g. a “cat” node can be connected to a “mammal” node via an “is-a” edge.

[0025] Ontology: Ontology: A formal knowledge representation method that defines concepts, relations, and axioms. Ontology is usually represented using languages ​​such as OWL.

[0026] OWL: A standard language for representing ontologies, based on RDF.

[0027] The rule states: Production rules: Rules in the form of "IF-THEN" that express conditions and actions. For example, "IF body temperature > 3¾ THEN fever".

[0028] Vector representation: Word embedding: Mapping words to a high-dimensional vector space to capture the semantic relationship between words. For example, Word2Vec, GloVe, and FastText.

[0029] Knowledge graph embedding: embedding entities and relations in the knowledge graph into a low-dimensional vector space for reasoning and querying. In this embodiment, the logical reasoning in step S3 includes propositional logic, predicate logic and modal logic. Propositional logic processes simple truth propositions, predicate logic is used to process more complex statements and reasoning, and modal logic is used to process possibility and necessity.

[0030] In this embodiment, the probabilistic reasoning in step S3 includes a Bayesian network, a Markov model and a factor graph. The Bayesian network represents the conditional dependency between variables, the Markov model is used to describe the transition probability between states, and the factor graph is used to represent the factor decomposition of the probability distribution.

[0031] In this embodiment, the evidence extraction in step S4 includes text mining, data query and knowledge graph query. Text mining is to extract relevant information from text, data query is to query relevant information from a database or API, and knowledge graph query is to query relevant information from a knowledge graph.

[0032] In this embodiment, the evidence integration in step S5 includes natural language generation, template filling, and rules and heuristics. Natural language generation is to convert evidence into natural language text, template filling is used to fill in evidence using predefined templates, and rules and heuristics are based on using rules and heuristic methods to integrate evidence.

[0033] In this embodiment, the answer synthesis in step S5 includes information extraction, information integration, retrieval-based methods, generation-based methods and hybrid methods, extracting relevant information from multiple sources. This can be accomplished through information retrieval and natural language processing technology. The extracted information is integrated, duplicate and conflicting information is removed, the most relevant and accurate information is retained, and the answer is directly extracted from the retrieved information. The answer that best matches the question is found from the FAQ library, and the answer is generated using a generative model. This method can generate more natural and coherent text, combining the two methods of retrieval and generation to improve the accuracy and coherence of the answer. Relevant information is first found through retrieval, and then the final answer is generated through a generative model.

[0034] Answer verification includes internal consistency check, user feedback and expert review. Internal consistency check is to check the logical consistency of the answer. User feedback is used to collect user feedback to evaluate the quality of the answer. Expert review is to ask domain experts to review the answer.

[0035] In this embodiment, the continuous learning in step S6 includes machine learning, reinforcement learning and knowledge base updating. Machine learning is to use machine learning models to optimize the reasoning process and evidence generation, reinforcement learning is to learn the optimal strategy through a reward mechanism, and knowledge base updating is to regularly update the knowledge base to reflect the latest information and knowledge.

[0036] Technical advantages: 1. Easy to understand: Rules are usually based on expert knowledge and experience, easy to understand and explain. 2. Flexibility: Different rules can be defined according to specific needs and scenarios. 3. Customizability: Rules can be customized according to specific performance standards to meet different evaluation needs.

[0037] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A knowledge reasoning and evidence generation method based on problem decomposition, characterized in that: The steps include: S1: Question analysis, including natural language processing, question classification and question understanding, uses NLP technology to analyze and understand the questions raised by users, determines the field or type to which the questions belong, and decomposes complex questions into a series of simpler sub-questions; S2: Knowledge base and data, building or leveraging existing knowledge bases, using relational databases and knowledge graphs, and leveraging existing texts, articles, and web content; S3: Reasoning engine, including logical reasoning, probabilistic reasoning, reasoning from basic principles and machine learning, using logical rules and formal methods to derive new information or conclusions, using probability theory to deal with uncertainty, using predefined rules to deal with problems, and using machine learning models to identify patterns and make predictions; S4: Evidence generation, including evidence extraction, evidence scoring and evidence integration, extracting evidence related to the sub-question from the knowledge base and data source, scoring the extracted evidence to determine its relevance and reliability. The scoring can be based on multiple factors, including the source, time, authority, etc. of the evidence, integrating multiple pieces of evidence into a coherent whole; S5: Answer production, including answer synthesis, answer verification and answer presentation, integrating the answers to the various sub-questions to form a complete answer to the original question, which can be achieved by verifying whether the generated answers are accurate, consistent and coherent, and presenting the final answer to the user in an appropriate form; S6: Learning and Feedback, including user feedback and continuous learning, collecting user feedback to evaluate the quality of answers and system performance, and continuously updating and optimizing the knowledge base and reasoning engine based on user feedback and new data.

2. The knowledge reasoning and evidence generation method based on problem decomposition according to claim 1 is characterized in that: The natural language processing in step S1 includes word segmentation, part-of-speech tagging, named entity recognition, dependency syntactic analysis and semantic role labeling. Part-of-speech tagging is to segment a sentence into words or phrases, part-of-speech tagging is to label the part of speech of each word, including nouns, verbs and adjectives, named entity recognition is to identify named entities in a sentence, including names of people, places and organizations, dependency syntactic analysis is to analyze the dependency relationship between words in a sentence, and semantic role labeling is to identify the relationship between a predicate and its arguments in a sentence.

3. The knowledge reasoning and evidence generation method based on problem decomposition according to claim 1 is characterized in that: The problem understanding in step S1 includes syntactic and semantic analysis, template matching, and rules and heuristics. Syntactic and semantic analysis is to identify the key components and sub-problems in the problem. Template matching is to use predefined templates to match the structure of the problem and decompose it into sub-problems. Rules and heuristics are to use rules and heuristic methods to identify and decompose the problem.

4. The knowledge reasoning and evidence generation method based on problem decomposition according to claim 1 is characterized in that: The knowledge base and data in step S2 include a knowledge base and a data source, wherein the establishment of the knowledge base includes knowledge acquisition, knowledge representation, knowledge organization and knowledge update. Knowledge acquisition can be performed by experts manually inputting knowledge, or by using natural language processing technology to automatically extract knowledge from text, or by importing data from existing databases or data sources.

5. The knowledge reasoning and evidence generation method based on problem decomposition according to claim 1 is characterized in that: The logical reasoning in step S3 includes propositional logic, predicate logic and modal logic. Propositional logic processes simple truth propositions, predicate logic is used to process more complex statements and reasoning, and modal logic is used to process possibility and necessity.

6. The knowledge reasoning and evidence generation method based on problem decomposition according to claim 1 is characterized in that: The probabilistic reasoning in step S3 includes a Bayesian network, a Markov model and a factor graph. The Bayesian network represents the conditional dependency relationship between variables, the Markov model is used to describe the transition probability between states, and the factor graph is used to represent the factor decomposition of the probability distribution.

7. The knowledge reasoning and evidence generation method based on problem decomposition according to claim 1 is characterized in that: The evidence extraction in step S4 includes text mining, data query and knowledge graph query. Text mining is to extract relevant information from text, data query is to query relevant information from a database or API, and knowledge graph query is to query relevant information from a knowledge graph.

8. The knowledge reasoning and evidence generation method based on problem decomposition according to claim 1 is characterized in that: The evidence integration in step S5 includes natural language generation, template filling, and rules and heuristics. Natural language generation is to convert evidence into natural language text, template filling is used to fill in evidence using predefined templates, and rules and heuristics are based on using rules and heuristic methods to integrate evidence.

9. The knowledge reasoning and evidence generation method based on problem decomposition according to claim 1 is characterized in that: The answer synthesis in step S5 includes information extraction, information integration, retrieval-based methods, generation-based methods and hybrid methods, extracting relevant information from multiple sources. This can be done through information retrieval and natural language processing technology, integrating the extracted information, removing duplicate and conflicting information, retaining the most relevant and accurate information, extracting answers directly from the retrieved information, finding the answer that best matches the question from the FAQ library, and using a generative model to generate answers. This method can generate more natural and coherent text, combining the two methods of retrieval and generation to improve the accuracy and coherence of the answer, first finding relevant information through retrieval, and then generating the final answer through a generative model.

10. The knowledge reasoning and evidence generation method based on problem decomposition according to claim 1, characterized in that: The continuous learning in step S6 includes machine learning, reinforcement learning and knowledge base updating. Machine learning is to use machine learning models to optimize the reasoning process and evidence generation, reinforcement learning is to learn the optimal strategy through a reward mechanism, and knowledge base updating is to regularly update the knowledge base to reflect the latest information and knowledge.