Employee knowledge question answering system and method based on large language model and enterprise knowledge base

Abstract

The invention discloses an employee knowledge question-answering system and method based on a large language model and an enterprise knowledge base. Comprising a question collection module, a question preprocessing module, a matching answer module, a new question storage module and an incentive mechanism module, and the question collection module is used for identifying chat records and entering a question queue according to question time; the question preprocessing module obtains previous business questions and generates a knowledge base after preprocessing the previous business questions; and the matching answering module is based on a knowledge base, uses a large language model to understand user input, matches questions and outputs the questions. A unified knowledge base obtaining system is established, questions directly reach employees, the uploading and issuing process is solved, rapid questions and answers are provided for the employees in multiple modes such as a large language model, waiting is avoided, and meanwhile expert knowledge is precipitated through question and answer storage.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to an employee knowledge question-answering system and method based on a large language model and an enterprise knowledge base. Background Art

[0002] At present, enterprises have established corporate WeChat groups and corporate instant messaging service groups as important tools and venues for timely consultation and resolution of technical problems encountered. In the course of business operations, grassroots employees are unable to receive timely guidance on system understanding, system operation and technical issues, which is mainly manifested in the gaps in uploading and issuing notification requirements, the mismatch between employee questions and answers, and the inability to effectively consolidate and pass on expert knowledge and experience. Therefore, it is urgent to further open up channels for resolving employee problems, further optimize customer service experience, and improve customer service quality.

[0003] In view of this, the present invention proposes an employee knowledge question-answering system and method based on a large language model and an enterprise knowledge base to solve the above problems. Summary of the invention

[0004] The purpose of the present invention is to provide an employee knowledge question-answering system and method based on a large language model and an enterprise knowledge base.

[0005] In order to achieve the above object, one of the technical solutions adopted by the present invention is:

[0006] Provided is an employee knowledge question-and-answer system based on a large language model and an enterprise knowledge base, comprising a question collection module, a question preprocessing module, a matching answer module, a new question storage module and an incentive mechanism module, wherein: the question collection module is used to identify chat records and enter the question queue according to the question time; the question preprocessing module obtains past business questions for preprocessing and generates a knowledge base; the matching answer module is based on the knowledge base and uses a large language model to understand user input, match questions and output.

[0007] In a feasible implementation, an employee knowledge question and answer system based on a large language model and an enterprise knowledge base also includes a new question storage module, which is used to automatically record new questions entering the question queue and valid answers in the queue into the question and answer system.

[0008] In a feasible implementation, an employee knowledge question and answer system based on a large language model and an enterprise knowledge base also includes an incentive mechanism module, which is used to introduce a competitive incentive mechanism. The chat robot counts and publishes the number and speed of experts answering questions in real time, and compares them with the best historical records.

[0009] In a feasible implementation, the matching answer module is based on the knowledge base and uses a large language model to understand user input, match questions and output them. If a question with a similarity of more than 60% is matched, a question and answer link is generated and pushed to the questioner immediately. If no match is found, it is automatically recorded in the knowledge base model to generate a question link, and the question link is sent through the group @ function to remind experts to review the question and provide guidance.

[0010] The second technical solution adopted by the present invention is:

[0011] Provided is an employee knowledge question-and-answer method based on a large language model and an enterprise knowledge base, comprising the following steps: S1, real-time monitoring of the question dynamics of an enterprise WeChat group and a company instant messaging service group, identifying employee questions through specific keywords, and entering the question queue according to the question time; S2, a question preprocessing module determines the correlation between questions in the question queue and classifies them, forms a similar question group and builds a question-and-answer engine; S3, matches similar questions from an existing knowledge base, and if a question with a similarity of more than 60% is matched, generates a question-and-answer link and pushes it to the questioner in real time; if no question is matched, it is automatically recorded in the knowledge base model to generate a question link, and the question link is sent through the group @ function for answering and automatically enters a new question storage module.

[0012] In a feasible implementation, the employee knowledge question-answering method based on a large language model and an enterprise knowledge base also includes the following steps: S4, the chat robot counts and publishes the number and speed of experts answering questions in real time, compares them with the best historical records, and stimulates the enthusiasm and participation of experts.

[0013] In a feasible implementation, the question-answering engine in S2 includes the establishment of ambiguous word text and classification text, collecting past business questions to establish a knowledge base and knowledge corpus, using TF-IDF to segment questions into phrases, counting the number of keywords, establishing a weight relationship text between keywords and the entire knowledge base, preliminarily implementing text segmentation, and then performing manual intervention to adjust the stop word weight to zero, increase the weight of proper nouns related to the company's business system and business, and establish an ambiguous word dictionary text;

[0014] The business and system-related keywords counted in the ambiguous word text stage are classified according to the company's existing technology and system classification to establish classification text;

[0015] Establish a classification expert association database, classify experts in the industry according to business categories, and associate categories with expert accounts;

[0016] ChatGLM model training first uses unsupervised learning to let the model learn language structure, semantics, and contextual dependencies on large-scale text data. Then, the collected business question-and-answer knowledge base is used as a fine-tuning dataset, and supervised learning is used to fine-tune the model. Based on the ambiguous word text and classified text established previously, weight pruning technology is used to remove weights in the model that contribute less to the generation effect, thereby reducing the complexity of the model.

[0017] Beneficial effects of the present invention:

[0018] The program of the present invention monitors group chat records in real time, and identifies the chat records as employee questions based on language features such as "Excuse me" and other semantic logics for seeking opinions or "what" and "maybe" and other questionable words. The program also sets a set of link production rules, which can bind question and answer links through classification and serial numbers, and use the rules to interact with the knowledge base system to realize WeChat group link push. The present invention establishes a unified knowledge base acquisition system, so that questions go directly to employees, solves the upload and download process, and uses a large language model and other methods to provide employees with quick questions and answers to avoid waiting, while precipitating expert knowledge through question and answer storage. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The figure is a flowchart of the method of the present invention. DETAILED DESCRIPTION

[0020] See also Figure 1 The embodiment of the present application provides an employee knowledge question and answer system based on a large language model and an enterprise knowledge base, including a question collection module, a question preprocessing module, a matching answer module, a new question storage module and an incentive mechanism module, wherein: the question collection module is used to identify chat records and enter the question queue according to the question time; the question preprocessing module obtains past business questions and generates a knowledge base after preprocessing; the matching answer module is based on the knowledge base and uses a large language model to understand user input, match questions and output.

[0021] In this embodiment, an employee knowledge question and answer system based on a large language model and an enterprise knowledge base also includes a new question storage module, which is used to automatically record new questions entering the question queue and valid answers in the queue into the question and answer system.

[0022] In this embodiment, an employee knowledge question and answer system based on a large language model and an enterprise knowledge base also includes an incentive mechanism module, which is used to introduce a competitive incentive mechanism. The chat robot counts and publishes the number and speed of experts answering questions in real time, and compares them with the best historical records.

[0023] The matching answer module described in this embodiment is based on the knowledge base, uses a large language model to understand user input, matches questions and outputs them. If a question with a similarity of more than 60% is matched, a question and answer link is generated and pushed to the questioner immediately. If no match is found, it is automatically recorded in the knowledge base model to generate a question link, and the question link is sent through the group @ function to remind experts to check the questions and provide guidance.

[0024] A method for employee knowledge question and answer based on a large language model and an enterprise knowledge base, comprising the following steps: S1, real-time monitoring of the question dynamics of an enterprise WeChat group and a company instant messaging service group, identifying employee questions through specific keywords, and entering the question queue according to the question time; S2, a question preprocessing module determines the correlation between questions in the question queue and classifies them, forms a similar question group and builds a question and answer engine; S3, matches similar questions from an existing knowledge base, and if a question with a similarity of more than 60% is matched, a question and answer link is generated and pushed to the questioner in real time. If no question is matched, it is automatically recorded in the knowledge base model to generate a question link, and the question link is sent through the group @ function for answering and automatically enters the new question storage module.

[0025] The employee knowledge question-and-answer method based on a large language model and an enterprise knowledge base in this embodiment also includes the following steps: S4, the chat robot counts and publishes the number and speed of experts answering questions in real time, compares them with the best historical records, and stimulates the enthusiasm and participation of experts.

[0026] The question-answering engine in S2 of this embodiment includes the establishment of ambiguous word text and classification text, collecting past business questions to establish a knowledge base and knowledge corpus, using TF-IDF to segment questions into phrases, counting the number of keywords, establishing a weight relationship text between keywords and the entire knowledge base, and preliminarily implementing text segmentation. Then, manual intervention is performed to adjust the stop word weight to zero, increase the weight of proper nouns related to the company's business system and business, and establish an ambiguous word dictionary text;

[0027] The business and system-related keywords counted in the ambiguous word text stage are classified according to the company's existing technology and system classification to establish classification text;

[0028] Establish a classification expert association database, classify experts in the industry according to business categories, and associate categories with expert accounts;

[0029] ChatGLM model training first uses unsupervised learning to let the model learn language structure, semantics, and contextual dependencies on large-scale text data. Then, the collected business question-and-answer knowledge base is used as a fine-tuning dataset, and supervised learning is used to fine-tune the model. Based on the ambiguous word text and classified text established previously, weight pruning technology is used to remove weights in the model that contribute less to the generation effect, thereby reducing the complexity of the model.

[0030] The specific implementation process of the present invention is as follows:

[0031] Real-time monitoring of questioning dynamics in corporate WeChat groups, company instant messaging service groups, etc., identify chat records, and identify employee questions through keywords such as "ask", "please ask", and "please advise", and enter the question queue. If a new question is identified, it will enter the next queue; split the questions into phrases, perform text segmentation, establish an ambiguous word dictionary, manually intervene in the segmentation results, and collect past business problems to establish a knowledge base; match questions with a similarity of more than 60% from the historical knowledge base, generate question and answer links, and push them to the questioner in real time. Unmatched questions are identified as new questions, and answers are generated through a large language model to provide ideas for solving the problems. New questions are classified according to question keywords, automatically recorded in the knowledge base system, and question links are generated. Questions are accurately matched to experts in their fields based on classification, and question links are sent through the group @ function to remind experts to review the questions and provide guidance.

[0032] for example:

[0033] An employee raises a question in the group: "Does any colleague know how to open a bank acceptance bill? The last step is to send the voucher to the xx system, but the xx system has no pending tasks. How can I solve this problem?" The program detects "Excuse me" in real time and identifies the chat record as an employee's question. This embodiment uses a knowledge question and answer collection program to monitor group chat records in real time. According to language features such as "Excuse me" and other semantic logics for seeking opinions or "what" and "what" and other question words, the chat record is identified as an employee's question, and link production rules are set. Question and answer links can be bound by classification and serial number, and the rules are used to interact with the knowledge base system to realize WeChat group link push;

[0034] After identifying the question, the ambiguous word dictionary established in the early stage identifies "opening bank acceptance bills", "sending vouchers", "xx system" and other business and system keywords involved in this question, as well as the main information of the question "no pending tasks" after ignoring the stop words, and matches the question with the business keyword "opening bank acceptance bills" and classifies it as "bill business";

[0035] The key information and classifications identified in the previous steps are quickly matched with the knowledge base established in the early stage, and all questions with a similarity of more than 60% are listed, and links are automatically sent to the group and @ed by the employee asking the question;

[0036] If no relevant knowledge is matched in the knowledge base, the question will be identified as a new question and recorded in the system. If the employee needs to raise a new question based on the unsolved problem in the knowledge base, the new question will be answered by the Q&A robot through the large language model, and a Q&A link will be generated and @ed to the employee in the group to provide him with ideas for the question; at the same time, the "bill business" expert will be matched according to the business classification in the previous step, and @ed to the expert in the group to reply in time;

[0037] All bank employees and experts can reply to the Q&A link. After the employee or expert answers the question, the link will be shared in the group and @ the questioner will be prompted to view the answer. At the same time, the number and speed of the answerer's answers will be published and compared with the best historical record to stimulate the enthusiasm and participation of experts;

[0038] The answers from domain experts are directly incorporated into the knowledge base, and the answers from the large language model and other employees are incorporated into the knowledge base if adopted by the questioner or liked by domain experts.

[0039] The present invention establishes a unified knowledge base acquisition system, so that questions can be directly sent to employees, solving the upload and download process, and using a large language model and other methods to provide employees with quick questions and answers to avoid waiting. At the same time, expert knowledge is accumulated by storing questions and answers in the database.

[0040] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An employee knowledge question-answering system based on a large language model and an enterprise knowledge base, characterized in that: It includes a question collection module, a question preprocessing module, a matching answer module, a new question storage module and an incentive mechanism module, wherein: the question collection module is used to identify chat records and enter the question queue according to the question time; the question preprocessing module obtains past business questions and generates a knowledge base after preprocessing; the matching answer module is based on the knowledge base and uses a large language model to understand user input, match questions and output.

2. According to claim 1, an employee knowledge question-answering system based on a large language model and an enterprise knowledge base is characterized in that: It also includes a new question storage module, which is used to store new questions entering the question queue and automatically record valid answers to the question and answer system.

3. The employee knowledge question-answering system based on a large language model and an enterprise knowledge base according to claim 1 or 2, characterized in that: It also includes an incentive mechanism module, which is used to introduce a competitive incentive mechanism. The chat robot counts and publishes the number and speed of experts answering questions in real time and compares them with the best historical records.

4. The employee knowledge question-answering system based on a large language model and an enterprise knowledge base according to claim 3 is characterized in that: The matching answer module is based on the knowledge base and uses a large language model to understand user input, match questions and output them. If a question with a similarity of more than 60% is matched, a question and answer link is generated and pushed to the questioner immediately. If no match is found, it is automatically recorded in the knowledge base model to generate a question link, and the question link is sent through the group @ function to remind experts to check the questions and provide guidance.

5. An employee knowledge question-answering method based on a large language model and an enterprise knowledge base, characterized in that: The following steps are involved: S1, real-time monitoring of the dynamics of questions in the enterprise WeChat group and the company's instant messaging service group, identifying employee questions through specific keywords, and entering the question queue according to the question time; S2, the question preprocessing module determines the correlation between the questions in the question queue and classifies them to form a knowledge base for matching similar question groups; S3, matches similar questions from the existing knowledge base. If a question with a similarity of more than 60% is matched, a question and answer link is generated and pushed to the questioner in real time. If no match is found, it is automatically recorded in the knowledge base model to generate a question link. The question link is sent through the group @ function for answering and automatically enters the new question storage module. The new question is answered through the large language model.

6. The employee knowledge question-answering method based on a large language model and an enterprise knowledge base according to claim 5 is characterized in that: The following steps are also included: S4, the chatbot counts and publishes in real time the number and speed of experts answering questions, compares them with the best historical records, and stimulates the enthusiasm and participation of experts.

7. The employee knowledge question-answering method based on a large language model and an enterprise knowledge base according to claim 6 is characterized in that: The establishment of the knowledge base in S2 includes the establishment of ambiguous word text, classification text, and classification expert association library. Past business problems are collected to establish a knowledge base and knowledge corpus. TF-IDF is used to segment the problems into phrases, the number of keywords is counted, and the weight relationship text between keywords and the entire knowledge base is established. The text segmentation is initially realized, and then manual intervention is performed to adjust the stop word weight to zero, increase the weight of the proper nouns related to the system and business of the company's business, and establish an ambiguous word dictionary text; the business and system-related keywords counted in the ambiguous word text stage are classified according to the company's existing technology and system, and the classification text is established; Establish a classification expert association database, divide in-line experts according to business categories, and associate categories with expert accounts.

8. The employee knowledge question-answering method based on a large language model and an enterprise knowledge base according to claim 5, characterized in that: The large language model adopts the ChatGLM model. First, unsupervised learning is used to allow the model to learn language structure, semantics, and contextual dependencies on large-scale text data. Then, the collected business question and answer knowledge base is used as a fine-tuning dataset, and supervised learning is used to fine-tune the model. Based on the ambiguous word text and classified text established previously, weight pruning technology is used to remove weights in the model that contribute less to the generation effect.

9. The employee knowledge question-answering method based on a large language model and an enterprise knowledge base according to claim 5, characterized in that: The relevance in S2 is achieved by identifying keywords in the question based on the ambiguous word dictionary established in the early stage, and classifying the question according to the keyword matching.

Images