Retrieval enhancement generation method for feedback type refined enhancement retrieval

Abstract

The invention discloses a feedback type retrieval enhancement generation method for refining and enhancing retrieval, which can give the confidence degree of answers while answering questions, and further process the answers lacking confidence. According to the method, the reliability of answers generated based on retrieval content is measured by introducing a confidence coefficient evaluation mechanism based on a retrieval strategy of first prediction and then matching, when the confidence coefficient is lower than a preset threshold value, it is indicated that a current retrieval result is insufficient to support accurate answering, at the moment, refined and enhanced retrieval is activated, refined questions are generated for original questions, and the accuracy of the questions is improved. Therefore, the understanding and retrieval effects on questions are enhanced, and the answering accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of retrieval enhancement generation (RAG), and in particular relates to a retrieval enhancement generation method for feedback-type refinement enhancement retrieval. Background Art

[0002] Large language models (LLMs) can suffer from problems such as hallucinations, outdated information, and a lack of in-depth domain knowledge when answering questions. Therefore, leveraging the language processing capabilities of LLMs and integrating them with external knowledge to compensate for their shortcomings and improve the accuracy of their answers has become a key issue in the field of natural language processing.

[0003] A knowledge graph is a directed graph consisting of nodes (representing entities) and edges (representing relationships between entities). In RAG (retrieval-augmented generation) technology, knowledge graphs are one of the most commonly used data sources due to their structured data. The application of RAG on knowledge graphs is demonstrated through the KGQA (knowledge graph question answering) task. Previous research has focused on obtaining as much relevant information as possible during the retrieval phase to solve the problem, while neglecting to evaluate the credibility of the answers after the retrieval and further processing unreliable results.

[0004] Therefore, how to overcome the above-mentioned defects in the existing technology and develop a more advanced search enhancement generation method has become a more urgent technical issue in this industry. Summary of the Invention

[0005] To achieve the above-mentioned objectives, the present invention aims to provide a retrieval enhancement generation method for feedback-based refinement-enhanced retrieval, which guides large language models (LLMs) to answer questions while giving the confidence (i.e., degree of confidence) of the answers to the questions. For questions with lower confidence in the answers, refined questions will be further generated for them, and then the refined questions will be retrieved and answered, and this process will be iteratively executed until the confidence of the answer is higher than a predetermined threshold or the maximum number of iterations is reached.

[0006] Specifically, the feedback-based refinement enhanced search generation method provided by the present invention includes the following steps:

[0007] Step 1: parse the question Q, extract key entities e and construct a key entity set E = {e1, e2, ..., e n}, where n is the total number of key entities e extracted from question Q.

[0008] It can be understood that the above question Q can be a question given by a user.

[0009] Step 2: For the above key entity set E={e1,e2,...,e n Each entity e in i , guiding large language models (LLMs) to predict the potential relationship r associated with it p (e i ), and construct the potential relationship set R p (e i )={r p1 (e i ),r p2 (e i ),...,r pm (e i )}, where i∈n, m is the potential relationship r p (e i ) total number.

[0010] Step 3: For each key entity e i ∈E query the knowledge graph and get the key entity e i All known relationships directly related to r r (e i ), and construct a known relationship set R r (e i )={r r1 (e i ),r r2 (e i ),...,r rl (e i )}, where l is the known relationship r r (e i ) total number.

[0011] It can be understood that the above knowledge graph is a directed graph consisting of nodes (representing entities) and edges (representing relationships between entities).

[0012] Step 4: For each key entity e i ∈E, calculate each pair of potential relations r px (e i )∈R p (e i ) and the known relationship r ry (e i )∈R r (e i ) between the similarity score sim(r px (e i ),r ry (e i )), where x∈m, y∈l.

[0013] It is understandable that the similarity score can be calculated in various applicable ways, for example:

[0014]

[0015] in, and Respectively, the relationship r px (e i )∈R p (e i ) and r ry (e i )∈R r (e i ) is a vector representation of .

[0016] It can be understood that the purpose of the above step 4 is to ensure that the selected relationship can accurately reflect the intention of the original question.

[0017] Step 5: For each key entity e i ∈E, the similarity score sim(r px (e i ),r ry (e i )) Sort from high to low and select the first k similarity scores sim(r px (e i ),r ry (e i )) corresponding to the known relations (i.e., the top k most relevant candidate relations) to construct the final relation set R for subsequent processing f (e i )={r f1 (e i ),r f2 (e i ),...,r fk (e i )}.

[0018] Based on the above steps, it is obvious that three types of relationships are set up in the present invention: one is the relationship in the retrieved knowledge graph (i.e., known relationship), the second is the relationship predicted by the large language model (LLMs) for question Q (i.e., potential relationship), and the third is the relationship extracted from the known relationship that can match the potential relationship (i.e., similar relationship with high score), which is called candidate relationship here, that is, the relationship used to construct the final relationship set.

[0019] Step 6: For each key entity e i ∈E, based on the above final relation set R f (e i )={r f1 (ei ),r f2 (e i ),...,r fk (e i )}Construct a complete triple form data T(e i )={e i ,r fj (e i ),t}, where j∈k, t represents the key entity e i Through the relationship fj The target entity to connect to.

[0020] Step 7: Based on the above triple form data T(e i )={e i ,r fj (e i ),t} to design the corresponding prompt template and combine it with the question Q to generate the prompt text.

[0021] Prompt templates are commonly known as templates or structures used to guide large language models (LLMs) to generate specific types of output. They define how large language models (LLMs) should be guided to generate specific content. Prompt templates play a crucial role in the training and application of large language models (LLMs), ensuring that large language models (LLMs) generate output in the expected manner after receiving instructions.

[0022] Step 8. Input the above prompt text into the large language model (LLMs) to guide the large language model (LLMs) to generate an answer A for question Q and a confidence score C(A) to reflect the level of confidence in the answer it gives. The confidence score C(A) is defined as: C(A)∈[0.0, 1.0].

[0023] It can be understood that C(A)=1.0 may represent absolute certainty, while C(A)=0.0 may represent complete uncertainty.

[0024] It can be understood that, in the present invention, the above prompt text can not only guide the large language model (LLMs) to answer the question, but also guide the large language model (LLMs) to give a confidence score C(A).

[0025] Step 9: Decide whether to accept the answer A generated by the large language model (LLM) based on a preset confidence threshold θ. If the confidence score C(A) > θ, the answer A is considered credible and the current answer A is directly returned to the user. Conversely, if the confidence score C(A) ≤ θ and the number of iterations is less than the maximum allowed number of iterations M, indicating that the current answer is not reliable, the following step 10 is triggered to perform a refined and enhanced search to obtain a higher-quality answer. If the number of iterations is equal to or greater than the maximum allowed number of iterations M, the current answer A is directly returned to the user.

[0026] The confidence threshold θ is directly related to the specific large language model (LLM) used. If different large language models (LLMs) are used, the confidence threshold θ may be different because the confidence levels of answers given by different large language models (LLMs) vary. ChatGPT, deepseek, and other large language models suitable for the present invention are examples.

[0027] The maximum allowed number of iterations M may be a natural number equal to or greater than 2.

[0028] Through this confidence scoring mechanism, the generation method of the present invention effectively combines external knowledge and internal reasoning capabilities, ensuring that the generated content is both rich and accurate. Furthermore, the confidence scoring mechanism provides a self-verification method, helping to identify and improve responses that may be biased or lack sufficient supporting evidence.

[0029] When the confidence score C(A) of the given answer A is ≤ θ, it indicates that the current question Q may be ambiguous or missing key information, which affects the model's ability to correctly understand the user's intent. To address this issue, this paper introduces an iterative refinement-enhanced retrieval mechanism, which gradually clarifies the actual requirements of question Q by generating a series of refined questions q.

[0030] Step 10: Perform refinement and enhanced retrieval to guide the large language model (LLMs) to further generate a refined question set DQ = {q1,q2,...,q h}, and based on each refined question q i ∈DQ executes the above steps 1-6 to construct the corresponding triple form data T′(e i )={e i ′,r fj ′(e i ),t′}.

[0031] Preferably, a prompt template can be constructed based on all questions Q with confidence C(A)≤θ and input to a large language model (LLMs) (such asFigure 5 As shown in Figure 2), the large language model (LLM) is guided to generate a set of refined questions DQ = {q1,q2,...,q h The construction of the detailed question set DQ can follow the following principles:

[0032] 1. For multi-hop problems that require reasoning across multiple entities to reach a conclusion, DQ can be constructed into a series of decomposed sub-problems.

[0033] 2. For single-hop questions caused by lexical polysemy, DQ can be designed as a set of questions to eliminate ambiguity.

[0034] It is understandable that once a more specific and clear set of detailed problems DQ = {q1,q2,...,q h}, the search and answer process will be restarted.

[0035] Step 11: Use the refined problem set DQ = {q1,q2,...,q h} generated triple form data T′(e i )={e i ′,r fj ′(e i ),t′} to design the corresponding prompt template, and combine it with the refined question set DQ to generate the prompt text.

[0036] Step 12: Increase the number of iterations by 1 and repeat steps 8-9 above.

[0037] Each step of the above generation method of the present invention may be continuously iterated until one of the following termination conditions is met:

[0038] 1. The confidence score C(A) of the answer A provided by the large language model (LLM) is greater than θ, and the answer is considered to be credible;

[0039] 2. The maximum allowed number of iterations M is reached, and the loop ends even if the ideal confidence score is not reached.

[0040] Finally, no matter under which condition the iteration is stopped, the last answer A is regarded as the final result and returned to the user.

[0041] Furthermore, in order to implement the above-mentioned retrieval enhancement generation method, the present invention constructs a corresponding model, which at least includes a retriever, a response feedback module (RFM) and a refinement enhanced retrieval module (RERM).

[0042] The retriever is mainly responsible for retrieving the knowledge required to support answering questions, and realizing the question parsing and corresponding retrieval functions involved in each step of the above method. For example, in step 1 of the above method, the retriever can be used to parse question Q.

[0043] The answer feedback module (RFM) is mainly used to answer questions based on the retrieved knowledge and give the confidence of the answer. It determines whether the current question enters the refinement and enhancement retrieval module based on whether the confidence exceeds the predetermined threshold. It is responsible for integrating the information retrieved from the knowledge graph and generating responses to the user's questions Q through large language models (LLMs). In order to improve the accuracy and reliability of the answers, the module also introduces a confidence evaluation mechanism. For example, steps 8 and 9 of the above method can be implemented using the answer feedback module.

[0044] The Refinement Enhanced Retrieval Module (RERM) is mainly used to generate a refined question set DQ = {q1,q2,...,q h}, and further implement the retrieval and answering process for the refined questions. For example, step 10 of the above method can be implemented using the answer feedback module.

[0045] Furthermore, the present invention also provides an electronic device, which includes a processor and a memory; the memory is used to store computer programs; and the processor is used to implement the steps of the above method when executing the program stored in the memory.

[0046] Furthermore, the present invention also provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the steps of the above method are implemented.

[0047] Compared to existing technologies, the enhanced retrieval generation method provided by the present invention can simultaneously provide confidence in the answer while answering the question, and further process answers that lack confidence. Based on a retrieval strategy that predicts first and then matches, this method introduces a confidence assessment mechanism to measure the reliability of the answers generated based on the search content. When the confidence level falls below a predetermined threshold, indicating that the current search results are insufficient to support an accurate answer, enhanced refinement retrieval is activated, generating refined questions based on the original question to enhance understanding of the question and retrieval results, thereby improving the accuracy of the answer. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The above-mentioned and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings. In the drawings, unless otherwise specified, the same reference numerals throughout the multiple drawings represent the same or similar components or elements. These drawings are not necessarily drawn to scale. It should be understood that these drawings only depict some embodiments disclosed in accordance with the present invention and are used to provide further understanding of the present invention. They constitute part of this application and should not be regarded as limiting the scope of the present invention. Among them:

[0049] Figure 1 This is an overall flow chart of the search enhancement generation method provided by the present invention;

[0050] Figure 2 This is a model structure diagram corresponding to the search enhancement generation method provided by the present invention;

[0051] Figure 3 A prompt template for guiding the large language model to answer questions and give confidence scores in the present invention;

[0052] Figure 4 A prompt template for guiding a large language model to generate a set of refined questions in the present invention;

[0053] Figure 5 A prompt template designed for a set of triple-form data generated based on a refined question set in the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and should not be construed as limiting the present invention.

[0055] The present invention provides a retrieval enhancement generation method for feedback type refinement enhancement retrieval. Figure 1-2 Specifically described, the method includes the following steps:

[0056] Step 1: parse the question Q, extract key entities e and construct a key entity set E = {e1, e2, ..., e n}, where n is the total number of key entities e extracted from question Q.

[0057] It can be understood that the above question Q may be a question asked by a user. In this embodiment, the question Q is, for example, “who was Stephen R. Covey?”.

[0058] Step 2: For the above key entity set E={e1,e2,...,e nEach entity e in i , guiding large language models (LLMs) to predict the potential relationship r associated with it p (e i ), and construct the potential relationship set R p (e i )={r p1 (e i ),r p2 (e i ),...,r pm (e i )}, where i∈n, m is the potential relationship r p (e i ) total number.

[0059] In this embodiment, for the above question Q “who was Stephen R. Covey?”, by predicting the entities and relationships of the question, it can be obtained that the entity is “Stephen R. Covey”, and the potential relationships may be “is_a”, “born_on” or “education”, etc.

[0060] Step 3: For each key entity e i ∈E query the knowledge graph and get the key entity e i All known relationships directly related to r r (e i ), and construct a known relationship set R r (e i )={r r1 (e i ),r r2 (e i ),...,r rl (e i )}, where l is the known relationship r r (e i ) total number.

[0061] It can be understood that in this embodiment, the required known relationship is obtained by matching the predicted entity with the corresponding relationship in the knowledge graph and the predicted potential relationship.

[0062] Step 4: For each key entity e i ∈E, calculate each pair of potential relations r px (e i )∈R p (e i ) and the known relationship r ry (e i )∈R r (e i) between the similarity score sim(r px (e i ),r ry (e i )), where x∈m, y∈l.

[0063] In this embodiment, the similarity score is calculated using the following method:

[0064]

[0065] in, and Respectively, the relationship r px (e i )∈R p (e i ) and r ry (e i )∈R r (e i ) is a vector representation of .

[0066] Step 5: For each key entity e i ∈E, the similarity score sim(r px (e i ),r ry (e i )) Sort from high to low and select the first k similarity scores sim(r px (e i ),r ry (e i )) corresponding to the known relations (i.e., the top k most relevant candidate relations) to construct the final relation set R for subsequent processing f (e i )={r f1 (e i ),r f2 (e i ),...,r fk (e i )}.

[0067] Step 6: For each key entity e i ∈E, based on the above final relation set R f (e i )={r f1 (e i ),r f2 (e i ),...,r fk (e i )}Construct a complete triple form data T(e i )={e i ,r fj(e i ),t}, where j∈k, t represents the key entity e i Through the relationship fj The target entity to connect to.

[0068] Step 7: Based on the above triple form data T(e i )={e i ,r fj (e i ),t} is used to design the corresponding prompt template T1, and is combined with the question Q to generate the prompt text.

[0069] In this embodiment, the specific style of the prompt template T1 is as shown in the attached Figure 3 shown.

[0070] Step 8. Input the above prompt text into the large language model (LLMs) to guide the large language model (LLMs) to generate an answer A for question Q and a confidence score C(A) to reflect the level of confidence in the answer it gives. The confidence score C(A) is defined as: C(A)∈[0.0, 1.0].

[0071] In this embodiment, C(A)=1.0 may represent absolute certainty, while C(A)=0.0 may represent complete uncertainty.

[0072] In this embodiment, the prompt text can not only guide the large language model (LLMs) to answer the question Q, but also guide the large language model (LLMs) to give a confidence score C(A). Figure 3 As shown, the “Give the probability” in the prompt template T1 used to generate the prompt text is used to guide the large language model (LLMs) to give a confidence score C(A).

[0073] Step 9: Decide whether to accept the answer A generated by the large language model (LLM) based on a preset confidence threshold θ. If the confidence score C(A) > θ, the answer A is considered credible and the current answer A is returned directly to the user. If the confidence score C(A) ≤ θ and the number of iterations is less than the maximum allowed number of iterations M, indicating that the current answer is not reliable, the following step 10 is triggered to perform a refined enhanced search to obtain a higher-quality answer. If the number of iterations is equal to or greater than the maximum allowed number of iterations M, the current answer A is returned directly to the user.

[0074] In this embodiment, the confidence threshold θ can be set to 0.8; the above-mentioned maximum allowed number of iterations M can be set to 2.

[0075] In this embodiment, if the confidence score C(A) ≤ θ, the predicted relationship cannot answer the question "Who was Stephen R. Covey?" Submitting the retrieved results to the Large Language Model (LLM) to answer the question will yield an unreliable answer A, which the LLM will assign a low confidence score. If the confidence score C(A) ≤ θ, the refinement and enhancement retrieval process is triggered.

[0076] Step 10: Perform refinement and enhanced retrieval to guide the large language model (LLMs) to further generate a refined question set DQ = {q1,q2,...,q h}, and based on each refined question q i ∈DQ executes the above steps 1-6 to construct the corresponding triple form data T′(e i )={e i ′,r fj ′(e i ),t′}.

[0077] In this embodiment, a prompt template T2 can be constructed based on all questions Q with confidence scores C(A)≤θ, and input to the large language model (LLMs), thereby guiding the large language model (LLMs) to generate a set of refined questions DQ = {q1, q2, ..., q h}. The specific style of the above prompt template T2 is as shown in the attached Figure 4 shown.

[0078] Once a more specific and clear set of detailed problems DQ = {q1,q2,...,q h}, the search and answer process will be restarted.

[0079] In this implementation, for the above question Q “who was Stephen R. Covey?”, multiple refined questions q can be generated, such as: “What was Stephen R. Covey's profession?”, “Which book did Stephen R. Covey author that became famous?”, “In which field did Stephen R. Covey make significant contributions?”, etc.

[0080] Step 11: Use the refined problem set DQ = {q1,q2,...,q h} generated triple form data T′(ei )={e i ′,r fj ′(e i ),t′}, the corresponding prompt template T3 is designed and combined with the refined question set DQ to generate the prompt text.

[0081] In this embodiment, the specific style of the prompt template T3 is as shown in the attached Figure 5 shown.

[0082] Step 12: Increase the number of iterations by 1 and repeat steps 8-9 above.

[0083] In this implementation, prediction and retrieval are also performed for the above-mentioned refined questions, and the entity "Stephen R. Covey" and the relationship "profession" can be predicted. By comparing the relationships between the entities in the knowledge graph, we can obtain knowledge such as (Stephen R. Covey, profession, Author), (Stephen R. Covey, profession, Writer), and (Stephen R. Covey, profession, Motivational speaker), which are more meaningful for answering the question. This knowledge is then given to large language models (LLMs) to answer the original question Q, so that an answer with as high a confidence level as possible can be obtained.

[0084] Furthermore, in order to implement the above-mentioned retrieval enhancement generation method, the present invention constructs a corresponding model, which at least includes a retriever, a response feedback module (RFM) and a refinement enhanced retrieval module (RERM).

[0085] The retriever is mainly responsible for retrieving the knowledge required to support answering questions, and realizing the question parsing and corresponding retrieval functions involved in each step of the above method. For example, in step 1 of the above method, the retriever can be used to parse question Q.

[0086] The answer feedback module (RFM) is mainly used to answer questions based on the retrieved knowledge and give the confidence of the answer. It determines whether the current question enters the refinement and enhancement retrieval module based on whether the confidence exceeds the predetermined threshold. It is responsible for integrating the information retrieved from the knowledge graph and generating responses to the user's questions Q through large language models (LLMs). In order to improve the accuracy and reliability of the answers, the module also introduces a confidence evaluation mechanism. For example, steps 8 and 9 of the above method can be implemented using the answer feedback module.

[0087] The Refinement Enhanced Retrieval Module (RERM) is mainly used to generate a refined question set DQ = {q1,q2,...,q h}, and further implement the retrieval and answering process for the refined questions. For example, step 10 of the above method can be implemented using the answer feedback module.

[0088] In another embodiment, the present invention further provides an electronic device comprising a processor and a memory; the memory is used to store computer programs; and the processor is used to implement the steps of the above method when executing the program stored in the memory.

[0089] In another embodiment, the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0090] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and purpose of the present invention.

Claims

1. A retrieval enhancement generation method for feedback-based refinement-enhanced retrieval, characterized in that: The method specifically comprises the following steps: Step 1: parse the question Q, extract key entities e and construct a key entity set E = {e1, e2, ..., e n }, where n is the total number of key entities e extracted from the question Q; Step 2: For each entity e in the key entity set E i , guiding the large language model to predict the potential relationship r related to it p (e i ), and construct the potential relationship set R p (e i )={r p1 (e i ),r p2 (e i ),...,r pm (e i )}, where i∈n, m is the potential relationship r p (e i ) Step 3: For each key entity e i ∈E query the knowledge graph and get the key entity e i All known relationships directly related to r r (e i ), and construct a known relationship set R r (e i )={r r1 (e i ),r r2 (e i ),...,r rl (e i )}, where l is the known relationship r r (e i ) Step 4: For each key entity e i ∈E, calculate each pair of potential relations r px (e i )∈R p (e i ) and the known relationship r ry (e i )∈R r (e i ) between the similarity score sim(r px (e i ),r ry (e i )), where x∈m, y∈l; Step 5: For each key entity e i ∈E, the similarity score sim(r px (e i ),r ry (e i )) Sort from high to low and select the first k similarity scores sim(r px (e i ),r ry (e i )) The known relations corresponding to the final relation set R are constructed f (e i )={r f1 (e i ),r f2 (e i ),...,r fk (e i )}; Step 6: For each key entity e i ∈E, based on the above final relation set R f (e i ) constructs a complete triple form data T(e i )={e i ,r fj (e i ),t}, where j∈k, t represents the key entity e i Through the relationship fj The target entity to be connected; Step 7: Based on the above triple form data T(e i )={e i ,r fj (e i ),t} construct a set T(E), design a corresponding prompt template based on the set T(E), and then combine it with the above question Q to generate the corresponding prompt text; Step 8: Input the prompt text into the large language model to guide the large language model to generate an answer A for the question Q and generate a confidence score C(A) that reflects its confidence level in the answer A. The confidence score C(A) is defined as: C(A)∈[0.0,1.0]; Step 9: Decide whether to accept the answer A generated by the large language model based on the preset confidence threshold θ. Specifically: If the confidence score C(A)>θ, then accept the above answer A; If the confidence score C(A)≤θ and the number of iterations is less than the maximum allowed number of iterations M, the following step 10 is triggered to perform a refined enhanced retrieval; If the number of iterations is equal to or greater than the maximum allowed number of iterations M, then accept the above answer A; Step 10: Perform refinement and enhancement retrieval to guide the large language model to further generate a set of refined questions DQ = {q1,q2,...,q h }, and based on each refined question q i ∈DQ executes the above steps 1-6 to construct the corresponding triple form data T′(e i )={e i ′,r fj ′(e i ),t′}; Step 11: Use the triple form data T′(e i )={e i ′,r fj ′(e i ),t′} construct a set T′(E), design a corresponding prompt template based on the set T′(E), and then combine it with the refined question set DQ to generate the corresponding prompt text; Step 12: Increase the number of iterations by 1 and repeat steps 8-9 above.

2. The search enhancement generation method according to claim 1, characterized in that: In step S1, the question Q is a question given by the user.

3. The search enhancement generation method according to claim 1, characterized in that: In step S4, the similarity score is calculated using the following formula: in, and Respectively, the relationship r px (e i )∈R p (e i ) and r ry (e i )∈R r (e i ) is a vector representation of .

4. The search enhancement generation method according to claim 1, characterized in that: In step S8, in C(A)∈[0.0, 1.0], C(A)=1.0 indicates absolute certainty, while C(A)=0.0 indicates complete uncertainty.

5. The search enhancement generation method according to any one of claims 1 to 4, characterized in that: In step S9, after answer A is accepted, answer A is returned to the user.

6. The search enhancement generation method according to any one of claims 1 to 4, characterized in that: In step S9, the preset confidence threshold θ is 0.

8.

7. The search enhancement generation method according to any one of claims 1 to 4, characterized in that: In step S9, the maximum allowed number of iterations M is a natural number equal to or greater than 2.

8. An electronic device, characterized in that: The electronic device includes a processor and a memory; the memory is used to store computer programs; the processor is used to implement the steps of the retrieval enhancement generation method described in any one of claims 1 to 7 when executing the program stored in the memory.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the search enhancement generation method according to any one of claims 1 to 7 are implemented.

Images