Dialogue supplementary retrieval enhancement generation method for long document set

Abstract

A dialogue supplementary retrieval enhancement generation method for a long document set is characterized by comprising the following steps: 1, constructing a dialogue supplementary retrieval enhancement generation system; 2, the man-machine interaction module obtains initial query of a user; 3, a retrieval query module retrieves the initial information according to the initial query of the user; 4, a decision demand score calculation module calculates the decision demand score of each target dimension according to the current query and the current information, convergence judgment is carried out, and when the convergence condition is not met, the step 5 is executed; otherwise, entering 8; 5, a sub-query generation module generates a sub-query according to the highest score dimension; 6, the man-machine interaction module collects user feedback information; 7, generating a new query by the retrieval query module, retrieving new information, and returning to the step 4; and 8, the decision demand score calculation module transmits all information collected by the current query to the answer generation module, and the answer generation module generates a final answer according to all the current information. The method has the effect that the capability of processing complex problems of the system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a conversation-supplemented retrieval enhancement generation method for long document sets. Background Art

[0002] As an efficient form of content organization, long document sets demonstrate significant advantages in multi-document question answering tasks. Existing retrieval-enhanced generative methods based on long document sets, such as RR-KR1, have achieved excellent results in multi-document and multi-knowledge question answering tasks, but their limitations are becoming increasingly apparent. Traditional RAG methods often struggle to accurately capture user intent, especially when users ask questions with limited information and unclear intent (e.g., "What procedures does Li Hua need to open an eco-farm?").

[0003] The shortcomings of existing RAG methods include: 1) they rely on one-time information retrieval, lack the ability to dynamically adjust during the conversation process, and are difficult to adapt to changing needs; 2) for short and fuzzy queries, it is difficult to obtain sufficient context to generate accurate answers; 3) with the popularization of question-answering systems, users have higher requirements for response quality and accuracy, and traditional methods are difficult to meet the needs of complex scenarios.

[0004] Disadvantages of existing technologies: Existing retrieval-enhanced generation (RAG) methods are difficult to provide satisfactory answers in complex scenarios such as information scarcity and unclear user intent. Summary of the Invention

[0005] The present invention provides a conversation-supplemented retrieval enhancement generation method for long document sets, which can automatically identify missing information in the conversation process and fill in the necessary information with the help of user interaction, thereby improving the system's ability to handle complex problems.

[0006] To achieve the above-mentioned object, the present invention provides a conversational supplementary retrieval enhancement generation method for a long document set, the key of which is to include the following steps:

[0007] Step 1: Constructing a dialogue-supplemented search enhancement generation system, wherein the dialogue-supplemented search enhancement generation system comprises a human-computer interaction module, a search query module, a decision requirement score calculation module, and a subquery generation module connected in sequence, wherein the subquery generation module is further connected to the human-computer interaction module, the decision requirement score calculation module is further connected to the answer generation module, and the answer generation module is further connected to the human-computer interaction module;

[0008] Step 2: The human-computer interaction module obtains the user's initial query Q0 and passes it to the search query module;

[0009] Step 3: The retrieval query module retrieves relevant preliminary information I0 from the long document set of the background knowledge base or other external knowledge base KB according to the user's initial query Q0, which is expressed as:

[0010] I0=Retrieve(KB,Q0)

[0011] Among them, KB represents the knowledge base, and Retrieve is the retrieval function;

[0012] The retrieval query module passes the user's initial query Q0 and preliminary information I0 to the decision-making requirement score calculation module;

[0013] The decision requirement score calculation module updates the query state Q=Q0 and updates the information state I=I0;

[0014] Step 4: The decision need score calculation module adopts a decision strategy based on the 5W2H framework to generate seven 5W2H target dimensions w of the current information. i , and calculate each 5W2H target dimension w i The decision demand score D(w i ), then calculate the sum of the decision-making demand scores of all current dimensions, and compare the calculated sum of the decision-making demand scores with the system convergence threshold ε;

[0015] When the sum of the decision demand scores is greater than the system convergence threshold ε, the decision demand score calculation module calculates the current decision demand score D(w i ) is passed to the subquery generation module; go to step 5; otherwise, go to step 8;

[0016] Step 5: The sub-query generation module generates the decision requirement scores D(w i ), determine the 5W2H target dimension w that has the highest decision-making need score max , then according to the 5W2H target dimension w max Generate subquery q1 and pass the subquery q1 to the human-computer interaction module;

[0017] Step 6: The human-computer interaction module presents the sub-query q1 to the user, collects user feedback information F1, and then sends the user feedback information F1 to the search query module;

[0018] Step 7: The retrieval query module combines the user's initial query Q0 and the feedback information F1 to form a new query Q1. The retrieval query module retrieves relevant new information I1 from the knowledge base based on the new query Q1 and passes the new query Q1 and the new information I1 to the decision requirement score calculation module;

[0019] The decision requirement score calculation module updates the query state Q=Q1, updates the information state I=I1, and enters step 4;

[0020] Step 8: The decision requirement score calculation module passes all the information collected from the current query to the answer generation module. The answer generation module generates a final answer R based on all the current information and presents the final answer R to the user through the human-computer interaction module.

[0021] The present invention establishes a feedback mechanism for human-computer interaction and automatically generates sub-queries based on the highest value of the demand score to guide users to supplement missing or ambiguous information in a targeted manner. It also uses the 5W2H decision-making strategy to systematically identify information gaps and combines the associated knowledge network CKN to perform conflict detection and missing information identification, significantly improving the ability to understand and answer complex and ambiguous user queries. The present invention gradually narrows the information gap through iterative optimization and ultimately generates more accurate and complete answers.

[0022] By incorporating a dynamic information supplementation mechanism and human-computer interaction feedback, the present invention achieves dynamic and accurate information supplementation in conversations, effectively improving the accuracy of complex questions and answers.

[0023] As an example: in step 4, the 5W2H framework includes seven key dimensions, namely, seven 5W2H target dimensions w i , namely: nature of event, cause and reason, time, place, involved entities, method and process, and quantitative aspects; these dimensions together form a comprehensive multidimensional analysis framework for detailed analysis and supplementary information needs.

[0024] The 5W2H decision-making strategy dimension table is shown in Table 1:

[0025]

[0026] This invention introduces a 5W2H decision-making strategy. When search results are insufficient to answer a user's question, the system uses this strategy to identify information gaps and proactively generate subqueries to guide the user to provide more specific information (such as time, place, person, reason, method, quantity, etc.). Through this iterative interaction, the system can more accurately understand the user's intent and obtain the key information necessary to answer the question.

[0027] As an example: in step 4, the decision requirement score calculation module detects each 5W2H target dimension w through the associated knowledge network CKN according to the current query state Q and the current information state I. i Conflict signal CKN in conflict and missing information signal CKN missing .

[0028] As a preference: each 5W2H target dimension w i The decision demand score D(w i )The calculation expression is as follows:

[0029] D(w i )=f(Q,I,w i ,CKN conflict ,CKN missing )

[0030] Among them, f is the prompt function based on the large language model.

[0031] The decision requirement score D(w i ) is used to evaluate the corresponding dimension w in the current information state i The necessity and urgency of the information.

[0032] The decision requirement score calculation process relies on a prompt function f based on a large language model. This prompt function comprehensively considers the current query, the currently retrieved information, the target dimension, and the conflict and missing information signals provided by CKN, so as to accurately calculate the decision requirement score of the corresponding target dimension.

[0033] As a preference: in step 5, the sub-query generation module generates the decision requirement scores D(w i ), determine the 5W2H target dimension w with the highest decision-making need score max , the expression is as follows:

[0034]

[0035] According to the 5W2H target dimension w max Generate subquery q1, the expression is as follows:

[0036] q1=Generate(w max )

[0037] Among them, Generate is the generation function and arg max is the maximum value function.

[0038] For the target dimension w max ,The system generates a subquery q1 with a clear target, which is used to request the user to supplement the information of this dimension.

[0039] Beneficial effects of the present invention:

[0040] 1. Iteratively optimize user queries to gradually narrow information gaps and ensure the system obtains the most valuable supplementary information in each conversation;

[0041] 2. By applying the systematic decision-making strategy of the 5W2H framework, the information retrieval process can be guided more effectively, ensuring that the needs of all key dimensions are met.

[0042] 3. Through the deep integration of the Correlated Knowledge Network (CKN), it can efficiently detect conflicts and identify missing information, thereby significantly improving the accuracy of judgment on query intent and information status.

[0043] 4. With the help of the human-computer interaction feedback mechanism HCIFM, dynamic adjustment of queries is achieved and the gradual improvement of information is promoted.

[0044] 5. This invention effectively enhances the ability of question-answering systems to handle complex and ambiguous questions, improves the user experience, and meets the demand for high-quality, high-precision information services in areas such as intelligent question-answering and intelligent assistants. This invention provides important technical support for building smarter and more flexible conversational question-answering systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the present invention;

[0046] Figure 2 A block diagram of the system structure for conversational supplementary retrieval enhancement. DETAILED DESCRIPTION

[0047] The present invention will be further described in detail below with reference to the accompanying drawings and specific examples. The following examples or drawings are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0048] like Figure 1 A conversation-supplemented retrieval enhancement generation method for a long document set is shown, comprising the following steps:

[0049] Step 1: Construct a dialogue-supplemented search enhancement generation system. The dialogue-supplemented search enhancement generation system is provided with a human-computer interaction module, a search query module, a decision requirement score calculation module, and a sub-query generation module connected in sequence. The sub-query generation module is also connected to the human-computer interaction module. The decision requirement score calculation module is also connected to the answer generation module. The answer generation module is connected to the human-computer interaction module. Figure 2 As shown;

[0050] Step 2: The human-computer interaction module obtains the user's initial query Q0 and passes it to the search query module;

[0051] Step 3: The retrieval query module retrieves relevant preliminary information I0 from the knowledge base according to the user's initial query Q0, and then passes the user's initial query Q0 and preliminary information I0 to the decision requirement score calculation module;

[0052] The decision requirement score calculation module updates the query state Q=Q0 and updates the information state I=I0;

[0053] Step 4: The decision need score calculation module adopts a decision strategy based on the 5W2H framework to generate seven 5W2H target dimensions w of the current information. i , and calculate each 5W2H target dimension w i The decision demand score D(w i ), then calculate the sum of the decision-making demand scores of all current dimensions, and compare the calculated sum of the decision-making demand scores with the system convergence threshold ε;

[0054] When the sum of the decision demand scores is greater than the system convergence threshold ε, the decision demand score calculation module calculates the current decision demand score D(w i ) is passed to the subquery generation module; go to step 5; otherwise, go to step 8;

[0055] Step 5: The sub-query generation module generates the decision requirement scores D(w i ), determine the 5W2H target dimension w that has the highest decision-making need score max , then according to the 5W2H target dimension w max Generate subquery q1 and pass the subquery q1 to the human-computer interaction module;

[0056] Step 6: The human-computer interaction module presents the sub-query q1 to the user, collects user feedback information F1, and then sends the user feedback information F1 to the search query module;

[0057] Step 7: The retrieval query module combines the user's initial query Q0 and the feedback information F1 to form a new query Q1. The retrieval query module retrieves relevant new information I1 from the knowledge base based on the new query Q1 and passes the new query Q1 and the new information I1 to the decision requirement score calculation module;

[0058] The decision requirement score calculation module updates the query state Q=Q1, updates the information state I=I1, and enters step 4;

[0059] Step 8: The decision requirement score calculation module passes all the information collected from the current query to the answer generation module. The answer generation module generates a final answer R based on all the current information and presents the final answer R to the user through the human-computer interaction module.

[0060] The 5W2H framework contains seven key dimensions, namely the seven 5W2H target dimensions w i , respectively: the nature of the event, the cause and reason, the time, the place, the participating entities, the method and process, and the quantitative aspect;

[0061] The 5W2H decision-making strategy dimension table is shown in Table 1:

[0062]

[0063] In step 4, the decision requirement score calculation module detects each 5W2H target dimension w through the associated knowledge network CKN according to the current query state Q and the current information state I. i Conflict signal CKN in conflict and missing information signal CKN missing .

[0064] The associated knowledge network CKN is crucial in step 4. For example, if the query is "Li Hua opens a farmhouse", CKN detects that the "farmhouse" entity has multiple associated "location" information (such as "Dianjiang County" and "Linshui County"), then CKN will be generated. conflict If CKN finds that the "farmhouse" entity is usually associated with the "business entity (Who)" information, but it is not clear in the current query Q and information I whether "Li Hua" is the business operator, then SHCKN is generated. missing signal, improve D(Who) score.

[0065] Each 5W2H target dimension w i The decision demand score D(w i )The calculation expression is as follows:

[0066] D(w i )=f(Q,I,w i ,CKN conflict ,CKN missing ).

[0067] In step 5, the sub-query generation module generates the decision requirement scores D(w i ), determine the 5W2H target dimension w with the highest decision-making need score max , the expression is as follows:

[0068]

[0069] According to the 5W2H target dimension w max Generate subquery q1, the expression is as follows:

[0070] q1=Generate(w max )

[0071] Among them, Generate is the generation function and arg max is the maximum value function.

[0072] For example, if w maxIf it is "Where", the generated Q1 might be: "Is your homestay located in Dianjiang County or Linshui County?"

[0073] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A conversational supplementary retrieval enhancement generation method for long document sets, characterized by: The following steps are involved: Step 1: Constructing a dialogue-supplemented search enhancement generation system, wherein the dialogue-supplemented search enhancement generation system comprises a human-computer interaction module, a search query module, a decision requirement score calculation module, and a subquery generation module connected in sequence, wherein the subquery generation module is further connected to the human-computer interaction module, the decision requirement score calculation module is further connected to the answer generation module, and the answer generation module is further connected to the human-computer interaction module; Step 2: The human-computer interaction module obtains the user's initial query Q0 and passes it to the search query module; Step 3: The retrieval query module retrieves relevant preliminary information I0 from the knowledge base according to the user's initial query Q0, and then passes the user's initial query Q0 and preliminary information I0 to the decision requirement score calculation module; The decision requirement score calculation module updates the query state Q=Q0 and updates the information state I=I0; Step 4: The decision need score calculation module adopts a decision strategy based on the 5W2H framework to generate seven 5W2H target dimensions w of the current information. i , and calculate each 5W2H target dimension w i The decision demand score D(w i ), then calculate the sum of the decision-making demand scores of all current dimensions, and compare the calculated sum of the decision-making demand scores with the system convergence threshold ε; When the sum of the decision demand scores is greater than the system convergence threshold ε, the decision demand score calculation module calculates the current decision demand score D(w i ) is passed to the subquery generation module; go to step 5; otherwise, go to step 8; Step 5: The sub-query generation module generates the decision requirement scores D(w i ), determine the 5W2H target dimension w that has the highest decision-making need score max , then according to the 5W2H target dimension w max Generate subquery q1 and pass the subquery q1 to the human-computer interaction module; Step 6: The human-computer interaction module presents the sub-query q1 to the user, collects user feedback information F1, and then sends the user feedback information F1 to the search query module; Step 7: The retrieval query module combines the user's initial query Q0 and the feedback information F1 to form a new query Q1. The retrieval query module retrieves relevant new information I1 from the knowledge base based on the new query Q1 and passes the new query Q1 and the new information I1 to the decision requirement score calculation module; The decision requirement score calculation module updates the query state Q=Q1, updates the information state I=I1, and enters step 4; Step 8: The decision requirement score calculation module passes all the information collected from the current query to the answer generation module. The answer generation module generates a final answer R based on all the current information and presents the final answer R to the user through the human-computer interaction module.

2. The method for enhancing the generation of conversational supplementary retrieval for long document sets according to claim 1, characterized in that: In step 4, the 5W2H framework includes seven 5W2H target dimensions w i , respectively: the nature of the event, the cause and reason, the time, the place, the participating entities, the method and process, and the quantitative aspect; The 5W2H decision-making strategy dimension table is shown in Table 1:

3. The method for enhancing the generation of conversational supplementary retrieval for long document sets according to claim 1, characterized in that: In step 4, the decision requirement score calculation module detects each 5W2H target dimension w through the associated knowledge network CKN according to the current query state Q and the current information state I. i Conflict signal CKN in conflict and missing information signal CKN missing .

4. The method for enhanced generation of conversational supplementary retrieval for long document sets according to claim 2 or 3, characterized in that: Each 5W2H target dimension w i The decision demand score D(w i )The calculation expression is as follows: D(in i )=f(Q,I,w i ,CKN conflict ,CKN missing )。 5. The method for enhanced generation of conversational supplementary retrieval for long document sets according to claim 1 or 4, characterized in that: In step 5, the sub-query generation module generates the decision requirement scores D(w i ), determine the 5W2H target dimension w with the highest decision-making need score max , the expression is as follows: According to the 5W2H target dimension w max Generate subquery q1, the expression is as follows: q1=Generate(w max ) Among them, Generate is the generation function and arg max is the maximum value function.

Images