CN-121434416-B - Hidden knowledge solidifying method, device, equipment and medium

CN121434416BCN 121434416 BCN121434416 BCN 121434416BCN-121434416-B

Abstract

The invention discloses a method, a device, equipment and a medium for solidifying implicit knowledge, which relate to the technical field of artificial intelligence, and the method comprises the steps of obtaining the context of a current task, determining an unprocessed item by using a generating agent, and searching in a preset fact knowledge base and/or a standard knowledge base by the generating agent according to the unprocessed item to generate an effective search record; the method comprises the steps of assembling a first prompt word, calling a large language model to infer the first prompt word to obtain a preliminary result, comparing the preliminary result with a standard example of a current task context to generate a difference report, assembling a second prompt word, calling the large language model to infer the second prompt word, extracting implicit knowledge according to the difference report, and solidifying the implicit knowledge contained in the standard example. The method and the system for mining and extracting the expert knowledge have the advantages that the expert knowledge is mined and extracted, the solidification of the expert knowledge is realized, and the accumulation and multiplexing and the structural management of the enterprise knowledge assets are realized.

Inventors

ZHU XIANGXIAN
LIN XIN
Lou Yilun
WU YAOGUANG
Hu songpo
ZHOU HONG
CHEN QI
ZHOU QI
LI MING

Assignees

宁波普瑞均胜汽车电子有限公司

Dates

Publication Date: 20260505
Application Date: 20251229

Claims (10)

1. A method for curing implicit knowledge, the method comprising: Acquiring a current task context of a user, determining an unprocessed item in the current task context by using a generating agent, searching in a preset fact knowledge base and/or a standard knowledge base by the generating agent according to the unprocessed item, and generating an effective searching record of the unprocessed item, wherein the fact knowledge base is used for storing fact information, the standard knowledge base is used for storing intention information which is derived from explicit knowledge and is statement information, the intention information is derived from solidified implicit knowledge and is instructive and regular information, the unprocessed item consists of a question point and a state point, each ambiguous fact information forms a question point under the condition that the fact information related to the current task context is determined to have an ambiguity, and each different task intention forms a state point under the condition that the current task context is determined to have different task intents relative to a historical task; The effective retrieval record and the current task context are assembled into a first prompt word for generating the intelligent agent, and a large language model embedded in the intelligent agent is called to be generated for reasoning the first prompt word, so that a preliminary result is obtained; comparing the preliminary result with a standard example of the current task context to generate a difference report; And assembling the difference report, the standard example, the effective search record and the current task context into a second prompt word of the cured intelligent agent, calling a large language model embedded in the cured intelligent agent to infer the second prompt word, extracting hidden knowledge according to the difference report, and curing the hidden knowledge contained in the standard example.
2. The method for solidifying implicit knowledge according to claim 1, wherein the step of obtaining the current task context of the user, determining the unprocessed item in the current task context by using the generating agent, and searching the generating agent in a preset fact knowledge base and/or a standard knowledge base according to the unprocessed item, to generate an effective search record of the unprocessed item, comprises the following specific steps: Acquiring a current task context; determining at least one question point and/or at least one status point in the current task context using the generating agent; Generating an intelligent agent, determining search rounds according to the number of the query points and the state points, generating search words of each round of search, and searching according to the search words and by using a fact knowledge base and/or a standard knowledge base to generate an effective search record; determining at least one query point and/or at least one status point in the active retrieval record using the generating agent; Generating search words of each round of search by the agent according to the number of the query points and the state points, and searching by the agent according to the search words and by using the fact knowledge base and/or the standard knowledge base to supplement the effective search records.
3. The method for solidifying implicit knowledge according to claim 2, wherein the generating agent determines the number of search rounds according to the number of the question points and the status points, generates the search word of each search round, and performs the search according to the search word and by using the fact knowledge base and/or the standard knowledge base, and generates the effective search record, specifically comprising: Generating an agent, determining search rounds according to the number of the query points and the state points, and determining the corresponding query points and/or state points of each round of search; Generating a fact search term for inquiring the fact content of the query point according to the query point, and generating an intention search term for reflecting the task intention of the state point according to the state point; determining the search content of each round of search, determining effective content from the search content, and adding the determined effective content to the tail of the existing content; and collecting the effective contents retrieved in all rounds to obtain an effective retrieval record.
4. The method for solidifying implicit knowledge according to claim 1, wherein the comparing the preliminary results with the standard examples of the current task context to generate the difference report comprises: acquiring a standard example of the current task context; Determining a preset comparison item and a preset neglect item; And according to a preset comparison project, utilizing the curing agent to compare the preliminary result with a standard example item by item to obtain a difference report.
5. The method for solidifying hidden knowledge according to claim 1, wherein the assembling the difference report, the standard case, the effective search record and the current task context into the second prompt word of the solidifying agent, and calling the large language model embedded in the solidifying agent to infer the second prompt word, extracting the hidden knowledge according to the difference report, solidifying the hidden knowledge contained in the standard case, specifically comprises: assembling the difference report, the standard example, the effective search record and the current task context into a second prompt word of the cured agent; calling a large language model embedded in the solidification intelligent body to infer a second prompt word to obtain an inference result, inducing each inference content forming the inference result into a preset style, and determining the type of each inference content according to the preset style; generating implicit knowledge of the deletion class under the condition that the type of the reasoning content is determined to be the first type; Generating implicit knowledge of the modification class in case that the type of the inference content is determined to be the second type; In the event that the type of the inferred content is determined to be the third type, implicit knowledge of the added class is generated.
6. The method for solidifying implicit knowledge of claim 5, wherein the method further comprises: And carrying out review processing on the implicit knowledge, and adding the revived implicit knowledge into the canonical knowledge base.
7. The method for solidifying hidden knowledge according to claim 6, wherein the processing for reviewing hidden knowledge and adding the reviewed hidden knowledge to a canonical knowledge base specifically comprises; acquiring all implicit knowledge output by the solidified intelligent agent; Each piece of implicit knowledge is reviewed; under the condition that the implicit knowledge is determined to be unnecessary to adjust, adding the implicit knowledge into a standard knowledge base; Under the condition that the implicit knowledge is determined to need to be adjusted, the implicit knowledge is adjusted, and the adjusted implicit knowledge is added into the standard knowledge base.
8. A device for solidifying implicit knowledge, said device comprising: The mixed retrieval module is used for acquiring the current task context of the user, determining unprocessed items in the current task context by utilizing the generation agent, retrieving the unprocessed items in a preset fact knowledge base and/or a standard knowledge base by the generation agent according to the unprocessed items, and generating effective retrieval records of the unprocessed items; the method comprises the steps that a current task context consists of a demand text of a user and a product background, a fact knowledge base is used for storing fact information, a normative knowledge base is used for storing intention information, the fact information is derived from explicit knowledge and is statement information, the intention information is derived from solidified implicit knowledge and is instructive and regular information, unprocessed items consist of doubtful points and state points, each ambiguous fact information forms a doubtful point under the condition that the fact information related in the current task context is determined to be ambiguous, and each different task intention forms a state point under the condition that the current task context is determined to be different task intentions relative to historical task intentions; the preliminary reasoning module is used for assembling the effective retrieval record and the current task context into a first prompt word for generating the intelligent agent, and calling a large language model embedded in the intelligent agent to reason the first prompt word so as to obtain a preliminary result; The information comparison module is used for comparing the preliminary result with a standard example of the current task context to generate a difference report; And the knowledge solidifying module is used for assembling the difference report, the standard example, the effective search record and the current task context into a second prompting word of the solidified intelligent body, calling a large language model embedded in the solidified intelligent body to infer the second prompting word, extracting hidden knowledge according to the difference report, and solidifying the hidden knowledge contained in the standard example.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the implicit knowledge solidification method of any one of claims 1 to 7 when the program is executed by the processor.
10. A non-transitory computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the implicit knowledge solidification method according to any one of claims 1 to 7.

Description

Hidden knowledge solidifying method, device, equipment and medium Technical Field The invention relates to the technical field of artificial intelligence, in particular to a method, a device, equipment and a medium for solidifying implicit knowledge. Background In knowledge-intensive organizations, expert implicit knowledge is the most valuable core asset. Implicit knowledge generally refers to knowledge that is difficult to express literally, relies on personal experience and intuition, such as design experience, debug skills, review criteria, and the like. The knowledge often exists in the expert's mind, is difficult to systematically record and inherit, and causes enterprises to face many challenges such as knowledge island, uneven working quality, long new employee training period and the like. The traditional knowledge management method, such as writing an operation manual, recording training videos, organizing teachers and students, and training, can transfer knowledge contents to a certain extent, but has low knowledge content transfer efficiency and is difficult to capture core contents in expert practice. In recent years, large language models (Large Language Model, LLM) exhibit strong capabilities in knowledge generation and question-answering, LLM can generate content from existing explicit knowledge, which tends to be more structural and relevant than direct questions. However, the generated content of the LLM still has obvious differences from the output result of the expert in aspects of specialty, consistency, writing style and standardability, and the differences are the direct manifestation of the lack of implicit knowledge. How to transform expert's implicit knowledge into structured assets that can be used by models or even multiplexed by the whole organization remains an important problem to be solved in the industry. Accordingly, there is a need to provide a method that can mine and refine implicit knowledge. Disclosure of Invention In view of the above, the embodiments of the present invention provide a method, an apparatus, a device, and a medium for solidifying implicit knowledge, so as to solve the problem in the prior art that it is difficult to convert the implicit knowledge of an expert into a structured asset that can be used by a model or even multiplexed by the whole organization. According to a first aspect, an embodiment of the present invention provides a method for curing implicit knowledge, where the method includes: Acquiring a current task context of a user, determining an unprocessed item in the current task context by using a generating agent, and searching in a preset fact knowledge base and/or a standard knowledge base by the generating agent according to the unprocessed item to generate an effective searching record of the unprocessed item, wherein the current task context consists of a demand text of the user and a product background, the fact knowledge base is used for storing actual information, the standard knowledge base is used for storing intention information, the actual information is derived from explicit knowledge and is declarative information, and the intention information is derived from solidified implicit knowledge and is instructive and regular information; The effective retrieval record and the current task context are assembled into a first prompt word for generating the intelligent agent, and a large language model embedded in the intelligent agent is called to be generated for reasoning the first prompt word, so that a preliminary result is obtained; comparing the preliminary result with a standard example of the current task context to generate a difference report; And assembling the difference report, the standard example, the effective search record and the current task context into a second prompt word of the cured intelligent agent, calling a large language model embedded in the cured intelligent agent to infer the second prompt word, extracting hidden knowledge according to the difference report, and curing the hidden knowledge contained in the standard example. With reference to the first aspect, in a first implementation manner of the first aspect, the obtaining a current task context of a user, determining, by using a generating agent, an unprocessed item in the current task context, and retrieving, by the generating agent, the unprocessed item in a preset fact knowledge base and/or a canonical knowledge base according to the unprocessed item, to generate an effective retrieval record of the unprocessed item, including: Acquiring a current task context; The unprocessed items are composed of the query points and the status points, each ambiguous factual information forms a query point when the ambiguous point exists in the factual information related in the current task context, and each different task intention forms a status point when the different task intentions exist in the current task context relative to the historical task inte