Machine question and answer large model illusion question relieving method based on domain experts
By employing expert domain identification and evidence verification methods, and utilizing a domain expert large language model to identify claiming experts, extract key entities, and retrieve external evidence, the LLM illusion problem is solved, and the accuracy and generalization ability of claims are improved.
Patent Information
- Application Number
- CN202510992055.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-18
- Publication Date
- 2025-11-07
AI Technical Summary
Large Language Models (LLMs) struggle to effectively mitigate the problem of generating erroneous reasoning (illusions) in the real world, especially in applications requiring factual accuracy. Existing methods may impair generalization ability across knowledge domains and are unreliable.
By identifying experts in the domain, extracting entities, retrieving evidence, and verifying correctness, the system utilizes a domain expert large language model to identify the experts involved in the statement, extract key entities, retrieve external evidence, compare the correctness of the statement, and provide modification suggestions.
It effectively mitigates the illusion problem of LLM, improves the accuracy of statements and cross-domain generalization ability, and leverages the expertise of domain experts to verify reasoning judgments.
Smart Images
Figure CN120911599A_ABST
Abstract
Description
[TECHNICAL FIELD]
[0001] The present application relates to a method for reducing hallucination problems of a large model of machine question answering based on domain experts. [BACKGROUND]
[0002] Because of the strong reasoning and generation capabilities of large language models (LLMs), LLMs are increasingly valued in the field of artificial intelligence. However, because of the limited knowledge stored by LLMs, sometimes incorrect reasoning is generated, commonly known as "hallucination". Fact hallucination poses a serious challenge to the application of LLMs, especially in the real world where factual accuracy is of great importance. Therefore, research on how to alleviate LLM hallucination is a hot topic in the field of artificial intelligence.
[0003] Some methods for alleviating LLM hallucination include using search engines to retrieve evidence, designing a trust score for LLMs to judge whether it is trustworthy based on the trust score, using the prompt method to let LLMs request more information before answering questions, and using training-based methods to perform linear probing on the hidden layers of LLMs, etc. to evaluate the accuracy of the generated text. However, these methods all require a set for training and hyperparameter tuning, which may compromise the generalization ability across knowledge domains. These methods usually rely on the assumption of "self-reflection", but confirmation bias and hallucination make "self-reflection" unreliable. [SUMMARY]
[0004] The present application overcomes the shortcomings of the prior art and provides a method for reducing hallucination problems of a large model of machine question answering based on domain experts.
[0005] To achieve the above-mentioned purpose, the following technical solutions are adopted in the present application:
[0006] A method for reducing hallucination problems of a large model of machine question answering based on domain experts, characterized in that it comprises
[0007] S1, expert domain identification, generating an initial reasoning judgment statement based on a large language model, and identifying the expert involved in the statement through a domain expert large language model;
[0008] S2, obtaining final evidence, identifying key entities in the statement through a domain expert large language model, and using the key entities as query external evidence, and selecting external evidence useful for analyzing the statement as final evidence;
[0009] S3, verifying correctness, comparing the statement and the final evidence through a domain expert large language model, analyzing the correctness of the statement, and giving modification suggestions for the statement.
[0010] The hallucination problem alleviating method for a machine question answering large model based on a domain expert according to the method has the characteristics that the expert discriminator large language model in the domain expert large language model identifies the expert involved in the statement in S1.
[0011] The hallucination problem alleviating method for a machine question answering large model based on a domain expert according to the method has the characteristics that the expert discriminator large language model in the domain expert large language model identifies the expert involved in the statement in S1.
[0012] The hallucination problem alleviating method for a machine question answering large model based on a domain expert according to the method has the characteristics that S2 includes
[0013] S21, entity recognition, the expert entity extraction large language model in the domain expert large language model identifies the key entity in the statement;
[0014] S22, evidence retrieval, the domain expert large language model takes the key entity as a query to retrieve external evidence;
[0015] S23, evidence filtering, the evidence filtering large language model in the domain expert large language model selects external evidence useful for analyzing the statement as final evidence.
[0016] The hallucination problem alleviating method for a machine question answering large model based on a domain expert according to the method has the characteristics that the expert entity extraction large language model in the domain expert large language model identifies the key entity in the statement in S21.
[0017] The hallucination problem alleviating method for a machine question answering large model based on a domain expert according to the method has the characteristics that the domain expert large language model uses the Bing search engine to retrieve external evidence in S22.
[0018] The hallucination problem alleviating method for a machine question answering large model based on a domain expert according to the method has the characteristics that the newspaper library in python is used to obtain the text content of the retrieved external evidence in S22.
[0019] The hallucination problem alleviating method for a machine question answering large model based on a domain expert according to the method has the characteristics that the evidence filtering large language model in S23 sets the prompt of useful evidence, and the evidence filtering large language model selects external evidence useful for analyzing the statement according to the prompt.
[0020] The hallucination problem alleviating method for a machine question answering large model based on a domain expert according to the method has the characteristics that the statement confirmation large language model in the domain expert large language model compares the statement and the final evidence, analyzes the correctness of the statement, and gives the statement modification suggestion in S3.
[0021] The beneficial effects of the present application are:
[0022] The present application generates the declaration basis of the reasoning judgment by a large language model, judges the domain to which the extracted key information belongs by a large language model, identifies the entity of the key information by a domain expert large language model, searches external evidence according to the entity as a query, selects the evidence related to the reasoning task by the evidence filtering large language model, and finally compares the useful external evidence searched with the declaration basis of the reasoning judgment generated by the large language model before to verify its correctness, thereby realizing the function of reducing the large model illusion based on the domain expert. [BRIEF DESCRIPTION OF DRAWINGS]
[0023] Figure 1 The schematic diagram of the present application is shown in the figure. [DETAILED DESCRIPTION]
[0024] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings.
[0025] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative positional relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications will also change accordingly. In addition, the descriptions involving "preferred", "suboptimal" and the like in the present application are only for descriptive purposes, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features with "preferred", "suboptimal" can be explicitly or implicitly included at least one of the features.
[0026] As shown in Figure 1 A method for reducing the large model illusion of machine question answering based on a domain expert, comprising
[0027] S1, expert domain identification, generating the declaration basis of the initial reasoning judgment by a large language model, and identifying the expert involved in the declaration by a domain expert large language model;
[0028] Specifically, the expert discriminator large language model in the domain expert large language model identifies the expert involved in the declaration in S1; the prompt for expert identification of the expert discriminator large language model is set in S1, and the expert discriminator large language model identifies the expert involved in the declaration according to the prompt. The learning template of the set prompt is "Please indentify the expert area of the given {claim}." The declaration claim is the basis generated by the large language model initially.
[0029] S2, obtaining final evidence, identifying key entities in the statement through the domain expert large language model, and retrieving external evidence as queries with the key entities, and selecting external evidence useful for analyzing the statement as final evidence; specifically, S2 includes
[0030] S21, entity recognition, identifying key entities in the statement through the expert entity extraction large language model in the domain expert large language model; wherein the prompt for the expert entity extraction large language model to identify key entities in the statement is set in S21, and the expert entity extraction large language model identifies key entities in the statement according to the prompt. Wherein the prompt learning template is "Supose you are an {Expert}, Please use your professional knowledge to extract entities which is helpful for verifying the correctness of the claim.", the case utilizes the information extraction capability of large language model, and uses the domain expert large language model to detect entities, showing strong generalization ability.
[0031] S22, retrieving evidence, the domain expert large language model retrieves external evidence as queries with the key entities; wherein the domain expert large language model uses bing search engine to retrieve external evidence in S22; the newspaper library in python is used to obtain the text content of the retrieved external evidence in S22.
[0032] S23, evidence filtering, selecting external evidence useful for analyzing the statement as final evidence through the evidence filtering large language model in the domain expert large language model. Wherein the prompt for the evidence filtering large language model to select useful evidence is set in S23, and the evidence filtering large language model selects external evidence useful for analyzing the statement according to the prompt. Wherein the prompt learning template is "You are given evidence regarding to the {Claim}. Please extract key evidence which is conductive to verify the correctness of the {Claim}".
[0033] S3, verifying correctness, comparing the statement and the final evidence through the domain expert large language model, analyzing the correctness of the statement, and giving statement modification suggestions.
[0034] In the case, the domain expert has professional knowledge in the field, is good at identifying entities in the field, and verifies whether the reasoning of the large language model is correct according to the retrieved information, therefore, based on the domain expert, the large model illusion technology can be reduced.
[0035] The above merely describes the preferred embodiments of the present application, and is not intended to limit the patent scope of the present application. Any equivalent structural variations or direct or indirect applications in other related technical fields, which are made based on the contents of the present application specification and drawings, are included in the patent protection scope of the present application.
Claims
1. A field expert-based machine question and answer large model hallucination question mitigation method, characterized by: Comprising S1, expert field identification, generating the initial inference judgment statement basis through a large language model, and identifying the expert involved in the statement through a domain expert large language model; S2, obtaining final evidence, identifying key entities in the statement through a domain expert large language model, and retrieving external evidence using key entities as queries, and selecting external evidence useful for analyzing the statement as final evidence; S3, verifying correctness, comparing the statement and the final evidence through a domain expert large language model, analyzing the correctness of the statement, and giving statement modification suggestions.
2. The method of claim 1, wherein the method is based on a domain expert. In S1, the expert discriminator large language model in the domain expert large language model identifies the expert involved in the statement.
3. The method of claim 2, wherein the method is characterized by: In S1, the expert discriminator large language model is set to identify the expert according to the prompt.
4. The method of claim 1, wherein the method is based on a domain expert. S2 comprises S21, entity recognition, identifying key entities in the statement through the expert entity extraction large language model in the domain expert large language model; S22, retrieving evidence, using key entities as queries to retrieve external evidence through the domain expert large language model; S23, evidence filtering, selecting external evidence useful for analyzing the statement as final evidence through the evidence filtering large language model in the domain expert large language model.
5. The method of claim 4, wherein the method is characterized by: In S21, the expert entity extraction large language model is set to identify key entities in the statement according to the prompt.
6. The method of claim 4, wherein the method is characterized by: In S22, the domain expert large language model uses the Bing search engine to retrieve external evidence.
7. The method of claim 6, wherein the method is based on a domain expert. In S22, the newspaper library in python is used to obtain the text content of the retrieved external evidence.
8. The method of claim 6, wherein the method is characterized by: In S23, the evidence filtering large language model is set to select useful evidence according to the prompt.
9. The method of claim 1, wherein the method is based on a domain expert. In S3, the statement confirmation large language model in the domain expert large language model compares the statement and the final evidence, analyzes the correctness of the statement, and gives statement modification suggestions.