A large model-based cross-domain multi-turn question answering method and system
By constructing knowledge bases by domain and combining multiple intent recognition methods, the accuracy and robustness of user intent recognition in cross-domain multi-turn question answering systems are solved, achieving efficient and accurate answers in cross-domain multi-turn question answering and simplifying the dialogue process.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA TELECOM DIGITAL INTELLIGENCE TECH CO LTD
- Filing Date
- 2025-08-06
- Publication Date
- 2026-07-24
AI Technical Summary
Existing question-answering systems struggle to achieve cross-domain, multi-turn question-answering and require data annotation to identify user intent.
It adopts a domain-based knowledge base construction approach, combining cascaded conditional judgments of rule-based intent recognition, semantic routing intent recognition, and language big model intent recognition. It performs intent recognition by rewriting user questions, outputs answers using big models, and does not require data annotation.
It achieves accuracy and robustness in cross-domain multi-turn question answering, improves the accuracy and generalization ability of user intent recognition, simplifies the dialogue process, and enhances the operability of the system.
Smart Images
Figure CN120973900B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of multi-turn question answering technology, specifically relating to a cross-domain multi-turn question answering method and system based on a large model. Background Technology
[0002] Question answering systems (QA) are an advanced form of information retrieval system that can answer user questions posed in natural language with accurate and concise natural language. The rise of QA research is primarily driven by the need for fast and accurate information retrieval. QA systems represent a promising and highly regarded research area within artificial intelligence and natural language processing.
[0003] Currently, one type of question-answering system is based on directly calculating the similarity between knowledge base knowledge and user questions, using knowledge graph techniques to enhance the model and help large models perform multi-hop reasoning. However, this approach is geared towards a single knowledge base and uses supervised training or knowledge graph techniques to enhance the model and help large models perform multi-hop reasoning. Another type of question-answering system uses semantic vectorization detection technology to train an intent recognition model for intent recognition and classification. However, this requires data annotation during intent recognition for training the human-computer dialogue model and cannot meet the needs of cross-domain multi-turn question answering. Summary of the Invention
[0004] This invention addresses the shortcomings of existing technologies by providing a cross-domain multi-turn question answering method and system based on a large model, which can accurately identify user intent across knowledge bases without requiring data annotation.
[0005] This invention provides the following technical solution:
[0006] Firstly, a cross-domain, multi-turn question answering method based on a large model is provided, including:
[0007] Several knowledge bases are constructed for different domains. The cards in each knowledge base include a knowledge base description, business rules, and tag information.
[0008] To identify the knowledge base to which the user intent belongs, the user question is rewritten, and based on the rewritten question, the user question, and the cards in the knowledge base, a cascaded conditional judgment is made for rule intent recognition, semantic routing intent recognition, and language big model intent recognition to determine the knowledge base to which the user intent belongs.
[0009] In the second round of question and answer, after summarizing the context of the user's question, a new question is obtained from the summary. This new question is then used as the user's question to identify the knowledge base to which the user's intent belongs.
[0010] In each round of question and answer, based on the user's question and the knowledge base to which the user's intent belongs, the large-scale question-and-answer model is used to output the answer language intent recognition for the user's question.
[0011] Optionally, the knowledge base includes high-density knowledge in the form of knowledge pairs and text blocks segmented from documents. The knowledge in the knowledge base is vectorized by a vectorization model and then stored in a vector knowledge base.
[0012] Optionally, when a user initiates a new round of question-and-answer dialogue, the original question Q1 is rewritten using a large model, resulting in two rewritten questions. and
[0013] Rule intent recognition: for the original problem Q1 and the rewritten problem and The intent is identified by rules. If the intents of at least two questions match the same knowledge base, then the matched knowledge base is the user intent knowledge base; otherwise, semantic routing intent identification is performed.
[0014] Semantic routing intent recognition: The Sentence-Bert model is used to encode the original question, the rewritten question, and the tag information of each knowledge base, and a similarity matching algorithm is used to find the matching probability of each question with each knowledge base; Language large model intent recognition is based on the original question, the rewritten question, and the knowledge base Z. i The matching probability is calculated, the average value is obtained, and it is determined whether the average value reaches a set threshold. If it does, then the knowledge base Z is... i If the current user's question intent is the knowledge base, then perform large-scale language model intent recognition;
[0015] Language Big Model Intent Recognition: Using a language big model to analyze the original problem Q1 and the rewritten problem. and Intent recognition is performed, and the knowledge base to which the user's intent belongs is selected based on the content of the knowledge cards in each knowledge base.
[0016] Optionally, after the second round of question-and-answering, a new question is obtained by summarizing the context of the user's question and using this new question as the user's question to re-identify the knowledge base to which the user's intent belongs. Specifically, this involves using an LLM model to summarize the user's historical inquiry questions and outputting a new question Q. did And the new question Q mid After rewriting, use the new question Q mid The system performs cascaded conditional judgments on rule-based intent recognition, semantic routing intent recognition, and language big model intent recognition to determine the knowledge base to which the user's intent belongs.
[0017] Optionally, during a multi-round conversation, it is determined whether the conditions for ending the multi-round conversation are met. If they are met, the conversation is terminated. The conditions for ending the multi-round conversation include: the user actively ending the conversation after terminating the operation, ending the conversation after a timeout, the user's question being located in different knowledge bases for consecutive rounds, in which case the current round of conversation is forcibly ended, and termination intention recognition. If the user's intention is identified as termination intention, the conversation is ended after answering the question in the last round.
[0018] Optionally, the method for identifying termination intent is as follows: constructing a termination intent knowledge base that includes conversational greetings, conversation termination, and non-business scope knowledge, and identifying whether the knowledge base to which the user intent belongs is a termination intent knowledge base.
[0019] Secondly, a cross-domain multi-turn question answering system based on a large model is provided, including:
[0020] The knowledge base construction module is used to build several knowledge bases by domain. The cards of the knowledge base include knowledge base description, business rules and tag information.
[0021] The domain intent recognition module is used to identify the knowledge base to which the user intent belongs. Specifically, it rewrites the user question and performs cascaded conditional judgments of rule intent recognition, semantic routing intent recognition, and language big model intent recognition based on the rewritten question, the user question, and cards in the knowledge base to determine the knowledge base to which the user intent belongs. In the second round of question answering and beyond, the user question is summarized in context to obtain a new question, which is then used as the user question to identify the knowledge base to which the user intent belongs again.
[0022] The answer module is used in each round of question and answer to output the answer language of the user's question based on the knowledge base to which the user's intent belongs, using the large-scale question and answer model to identify the intent of the large-scale question and answer model.
[0023] Optionally, it also includes a session termination module, used to determine whether the conditions for ending a multi-round session are met during the multi-round session process. If the conditions are met, the session is terminated. The conditions for ending a multi-round session include: the user actively terminates the session after an operation, the session ends after a timeout, the user's question is located in different knowledge bases for consecutive rounds, in which case the current round of dialogue is forcibly ended, and termination intent recognition is performed. If the user's intent is identified as termination intent, the session ends after answering the question in the last round. The termination intent recognition is performed by constructing a termination intent knowledge base that includes conversational greetings, session termination, and non-business scope knowledge, and identifying whether the knowledge base to which the user's intent belongs is a termination intent knowledge base.
[0024] Thirdly, a computer device is provided, including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the steps of the cross-domain multi-turn question-answering method based on a large model as described in any one of the first aspects.
[0025] Fourthly, a computer-readable storage medium is provided for storing a computer program; when the computer program is executed by a processor, it implements the steps of the cross-domain multi-turn question-answering method based on a large model as described in any one of the first aspects.
[0026] Compared with the prior art, the beneficial effects of the present invention are:
[0027] (1) The present invention adopts a business-based database design, which facilitates the management, expansion, updating and maintenance of business domain knowledge; the knowledge base cards include three types of information, which correspond to three intent recognition algorithms, thereby improving the accuracy of intent recognition.
[0028] (2) This invention rewrites the user's original question or historical dialogue question into a large model to enhance the generalization of the user's intent in the dialogue scenario, thereby dealing with the situation of forgetting or confusing the user's question intent and the dynamic change of intent, and realizing the accurate understanding of the user's intent; this method can also generalize to phrases, ellipses, spoken expressions, etc., so that it has better robustness.
[0029] (3) This invention integrates three methods: business rules, semantic routing, and large model to identify user intent in dialogue scenarios without the need to label dialogue data. This solves the problem that labeled intent text is often scarce and difficult to obtain, and that self-labeling is difficult.
[0030] (4) In multi-turn question-and-answer scenarios, the present invention adopts a cyclical architecture pattern, which helps to simplify the dialogue process and enhance the operability of the dialogue system. In addition, the present invention considers user context information in multi-turn dialogues, rather than just the current user question. Its key point is that by rewriting the user's historical questions, it enhances the generalization ability of user intent recognition. Attached Figure Description
[0031] Figure 1 This is a flowchart of the cross-domain multi-turn question answering method based on a large model according to the present invention;
[0032] Figure 2 This is a structural block diagram of the cross-domain multi-turn question-answering system based on a large model according to the present invention. Detailed Implementation
[0033] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be used to limit the scope of protection of the present invention. It should be noted that the term "comprising" and any variations thereof in the specification, claims and the above-mentioned drawings of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or devices.
[0034] Before explaining the embodiments of the present invention, some terms that will appear thereafter will be explained:
[0035] LLM (Large Language Model): A large language model is an artificial intelligence model designed to understand and generate human language. It has hundreds of billions of parameters. Large language models are pre-trained on massive amounts of text data and then fine-tuned through supervised learning and reinforcement learning based on human feedback to acquire capabilities such as text generation, text understanding, and interactive dialogue.
[0036] BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained language model in the field of natural language processing. It is trained on large-scale text data and used for various NLP tasks.
[0037] Sentence-BERT (SBERT) is a variant of BERT that uses Siamese BERT-Networks to generate semantically rich sentence vectors. It is suitable for large-scale semantic similarity calculation and clustering, and significantly improves computational efficiency.
[0038] Example 1
[0039] like Figure 1 As shown, a cross-domain multi-turn question answering method based on a large model includes the following steps:
[0040] Step S1: Construct several knowledge bases by domain. The cards in each knowledge base include a knowledge base description, business rules, and tag information.
[0041] Build a domain knowledge base Z based on business domains i Let i = 1, 2, ..., n. The knowledge base includes high-density knowledge in the form of knowledge pairs, as well as text blocks segmented from documents. The knowledge in the knowledge base is vectorized using a vectorization model and then stored in a vector knowledge base.
[0042] Each business domain's knowledge base includes a detailed knowledge base description, business rules, and tag information. This is defined here as a knowledge base card.
[0043] In some other embodiments, a termination intent knowledge base is constructed, which includes knowledge of conversational greetings, conversation termination, and non-business-related matters.
[0044] The specific settings for business rules refer to existing technologies. Business rules are used for rule intent determination, tags are used for semantic routing intent determination, and the knowledge base card content, composed of knowledge base description, business rules, and tag information, is used for semantic big model intent determination.
[0045] A knowledge base is built based on business domains, and each knowledge base card contains a detailed description of the knowledge base, business rules, and tag information. This facilitates the maintenance of knowledge points in the knowledge base and also assists in intent determination by introducing knowledge base cards.
[0046] Step S2: Identify the knowledge base to which the user intent belongs. Specifically, rewrite the user question, and based on the rewritten question, the user question, and the cards in the knowledge base, perform cascaded conditional judgments of rule intent recognition, semantic routing intent recognition, and language big model intent recognition to determine the knowledge base to which the user intent belongs.
[0047] Step S21: User Question Rewriting. When a user initiates a new round of question-and-answer dialogue, the original user question Q1 is rewritten using a semantic big data model, resulting in two rewritten questions. and Alternatively, the semantic large model can be an LLM (Large Language Model) or a GPT series (OpenAI) model.
[0048] It is worth noting that if the user question contains an image, the existing image big model is used to convert the image into text to represent the user question, and then the user question is rewritten. For specific image big models, refer to existing technologies, such as Tongyi Qianwen-VL-Chat.
[0049] Step S22: Rule intent recognition.
[0050] Regarding the original problem Q1 and the rewritten problem and The intent is identified by rules. If at least two questions hit the same knowledge base, then the hit knowledge base is the user intent knowledge base; otherwise, semantic routing intent identification is performed.
[0051] Specifically, this involves both the original problem Q1 and the rewritten problem. and Keyword extraction is performed. If the extracted keywords for a given question match a rule in a knowledge base, then that question matches that knowledge base. A voting method is used to select the correct keywords for Q1, Q2, Q3, Q4, Q5, Q6, Q7, Q8, Q9, Q1 ...2, Q3, Q4, Q5, Q6, Q7, Q8 The intention is used to determine the rule, Ω(Q1)→Z i , Only when at least two intentions Figure 1 Only if the condition is met will the rule be applied. Otherwise, the semantic routing intent identification in step S23 will be performed.
[0052]
[0053] Step S23: Semantic routing intent identification.
[0054] The Sentence-Bert model is used to encode the original question, the rewritten question, and the label information of each knowledge base, and a similarity matching algorithm is used to find the matching probability between each question and each knowledge base.
[0055] If the user intent skips the rules, semantic routing intent identification is performed. This is a typical short text classification. In actual intelligent question-answering and intelligent customer service scenarios, questions and answers have been found to have the following characteristics: sparsity (few words, lack of context), non-standardization (concise text expression, abbreviations, non-standard terms), many proper nouns (domain terms, platform names, tool names), and imbalance of business categories. Therefore, the Sentence-Bert model is adopted for semantic routing intent identification.
[0056] The semantic routing discrimination output is the probability P of being assigned to each domain knowledge base. i , i = 1, 2, ..., n, P i The value range is [0,1]. This relates to the original user problem Q1 and its rewritten form. and Semantic routing determines whether the knowledge base Z is located. i The probability of that, i.e.:
[0057] Step S24: Based on the original problem, rewrite the problem and knowledge base Z i The matching probability is calculated, the average value is obtained, and it is determined whether the average value reaches a set threshold. If it does, then the knowledge base Z is... i If the current user's question intent is the knowledge base, then perform large-scale language model intent recognition.
[0058]
[0059] If each question falls into knowledge base Z i The average probability P i 1 If the value is ≥0.8, then the knowledge base Zi A knowledge base for the current user's problem intent.
[0060] Step S25: Language large model intent recognition.
[0061] Using a large language model to analyze the original problem Q1 and its rewritten version and Intent recognition is performed, and the knowledge base to which the user's intent belongs is selected based on the content of the knowledge cards in each knowledge base.
[0062] The language big data model can directly reference existing technologies, such as the Qwen2.5-32B model. After the language big data model identifies the intent, it selects the knowledge base that the user's question points to: Q1→Z. i .
[0063] By rewriting the problem, the robustness and generalization ability of intent recognition are improved, and challenges such as sparsity, non-standardization, and the presence of many proper nouns are addressed. By integrating business rules, semantic routing, and a large language model, rapid determination of intent is achieved, improving recognition accuracy.
[0064] Step S3: After the second round of question and answer, summarize the context of the user's question to obtain a new question, and use the new question as the user's question to identify the knowledge base to which the user's intent belongs again.
[0065] Specifically, using the LLM model, the user's historical consultation questions are summarized, and a new question Q is output after the summary. mid ,Right now Where m ≥ 2. And the new problem Q... mid After rewriting, use the new question Q mid The system performs cascaded conditional judgments on rule-based intent recognition, semantic routing intent recognition, and language big model intent recognition to determine the knowledge base to which the user's intent belongs.
[0066] By analyzing multiple dialogues within this round, a large model is used to summarize the user's true intent in the conversation, thus achieving intent extraction. Subsequently, the question is rewritten to identify the content of intent jumps in multi-round dialogues within the large model.
[0067] Step S4: In each round of question and answer, based on the user's question and the knowledge base to which the user's intent belongs, the large-scale question-and-answer model is used to output the answer to the user's question.
[0068] Specifically, the large-scale response model can refer to existing technologies. The conversational system performs semantic and keyword retrieval, sorting, addition, and deletion of knowledge pairs, knowledge blocks, and other content within a specific knowledge base based on the user's question, until the large-scale model responds based on the retrieved content and the user's question.
[0069] In some other embodiments, during a multi-turn session, it is determined whether the conditions for ending the multi-turn session are met; if so, the session is terminated.
[0070] The conditions for ending a multi-turn session include:
[0071] (1) The user terminates the operation and ends the process manually;
[0072] (2) End after timeout; for example, 10s.
[0073] (3) If the user's inquiry is located in different knowledge bases in consecutive rounds, the current round of dialogue will be forcibly ended.
[0074] (4) Perform termination intent recognition. If the user's intent is identified as termination intent, the session ends after answering the last round of questions. That is, the last round of questions is rewritten, and then the operation in step S2 is executed.
[0075] The method for identifying termination intent is to construct a termination intent knowledge base that includes conversational greetings, conversation termination, and non-business-related knowledge, and to identify whether the knowledge base to which the user's intent belongs is a termination intent knowledge base. If so, it is a termination intent.
[0076] A loop architecture is adopted to identify the intent to query the knowledge base, the intent to engage in multi-turn dialogues, the intent to terminate a conversation, and the intent to switch knowledge bases. By rewriting the first question, the last question, and the questions summarized from historical dialogues, the robustness and generalization ability of intent recognition are improved.
[0077] Example 2
[0078] like Figure 2 As shown, a cross-domain multi-turn question answering system based on a large model is provided, including:
[0079] The knowledge base construction module is used to build several knowledge bases by domain. The cards of the knowledge base include knowledge base description, business rules and tag information.
[0080] The domain intent recognition module is used to identify the knowledge base to which the user intent belongs. Specifically, it rewrites the user question and performs cascaded conditional judgments of rule intent recognition, semantic routing intent recognition, and language big model intent recognition based on the rewritten question, the user question, and cards in the knowledge base to determine the knowledge base to which the user intent belongs. In the second round of question answering and beyond, the user question is summarized in context to obtain a new question, which is then used as the user question to identify the knowledge base to which the user intent belongs again.
[0081] The answer module is used in each round of question and answer to output an answer to the user's question based on the knowledge base to which the user's intent belongs, using a large-scale question and answer model.
[0082] In this embodiment, a cross-domain multi-turn question-answering system based on a large model further includes a session termination module, used to determine whether the conditions for ending a multi-turn session are met during the multi-turn session. If the conditions are met, the session is terminated. The conditions for ending a multi-turn session include: the user actively terminates the session after an operation, the session ends after a timeout, the user's question is located in different knowledge bases for consecutive rounds (in which case the current round of dialogue is forcibly ended), and termination intent recognition. If the user's intent is identified as termination intent, the session ends after answering the question in the last round. The termination intent recognition is performed by constructing a termination intent knowledge base that includes conversational greetings, session termination, and non-business-related knowledge, and identifying whether the knowledge base to which the user's intent belongs is a termination intent knowledge base.
[0083] For more detailed information on the above methods, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.
[0084] Example 3
[0085] The present invention provides a computer device, including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the steps of the above-described cross-domain multi-turn question-answering method based on a large model.
[0086] For more detailed information on the above methods, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.
[0087] Example 4
[0088] The present invention provides a computer-readable storage medium for storing a computer program; when the computer program is executed by a processor, it implements the steps of the above-described cross-domain multi-turn question-answering method based on a large model.
[0089] For more detailed information on the above methods, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.
[0090] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The systems, devices, and storage media disclosed in the embodiments are described simply because they correspond to the methods disclosed in the embodiments; relevant details can be found in the method section.
[0091] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.
[0092] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.
Claims
1. A cross-domain multi-turn question answering method based on a large model, characterized in that, include: Several knowledge bases are constructed for different domains. The cards in each knowledge base include a knowledge base description, business rules, and tag information. To identify the knowledge base to which the user intent belongs, the user question is rewritten, and based on the rewritten question, the user question, and the cards in the knowledge base, a cascaded conditional judgment is made for rule intent recognition, semantic routing intent recognition, and language big model intent recognition to determine the knowledge base to which the user intent belongs. In the second round of question and answer, after summarizing the context of the user's question, a new question is obtained from the summary. This new question is then used as the user's question to identify the knowledge base to which the user's intent belongs. In each round of question and answer, based on the user's question and the knowledge base to which the user's intent belongs, the large-scale question-and-answer model is used to output the answer to the user's question. When a user initiates a new round of question-and-answer dialogue, a large model is used to analyze the user's original question. Rewrite the code to obtain two rewrite problems. and ; Rule intent recognition: for the original problem and rewriting issues and The intent is identified by rules. If the intents of at least two questions match the same knowledge base, then the matched knowledge base is the user intent knowledge base; otherwise, semantic routing intent identification is performed. Semantic routing intent recognition: The Sentence-Bert model is used to encode the original question, the rewritten question and the tag information of each knowledge base, and the similarity matching algorithm is used to find the matching probability of each question and each knowledge base respectively. Based on the original question, the rewritten question, and the knowledge base. The matching probability is calculated, the average value is obtained, and it is determined whether the average value reaches a set threshold. If it does, the knowledge base... If the current user's question intent is the knowledge base, then perform large-scale language model intent recognition; Language Big Model Intent Recognition: Using a language big model to analyze the original problem and rewriting issues and Intent recognition is performed, and the knowledge base to which the user's intent belongs is selected based on the content of the knowledge cards in each knowledge base.
2. The cross-domain multi-turn question answering method based on a large model according to claim 1, characterized in that, The knowledge base includes high-density knowledge in the form of knowledge pairs and text blocks segmented from documents. The knowledge in the knowledge base is vectorized by a vectorization model and then stored in a vector knowledge base.
3. The cross-domain multi-turn question answering method based on a large model according to claim 1, characterized in that, In the second round of question-and-answer sessions and beyond, after summarizing the context of the user's questions to obtain new, summarized questions, these new questions are used as the user's questions to re-identify the knowledge base to which the user's intent belongs. Specifically, this involves using an LLM model to summarize the user's historical inquiry questions and outputting new, summarized questions. New problems After rewriting, utilize the new problem The system performs cascaded conditional judgments on rule-based intent recognition, semantic routing intent recognition, and language big model intent recognition to determine the knowledge base to which the user's intent belongs.
4. The cross-domain multi-turn question answering method based on a large model according to claim 1, characterized in that, During a multi-round conversation, it is determined whether the conditions for ending the multi-round conversation are met. If they are met, the conversation is terminated. The conditions for ending a multi-round conversation include: the user actively ending the conversation after terminating the operation, ending the conversation after a timeout, the user's question being located in different knowledge bases for consecutive rounds, in which case the current round of conversation is forcibly ended, and termination intention recognition. If the user's intention is identified as termination intention, the conversation is ended after answering the question in the last round.
5. The cross-domain multi-turn question answering method based on a large model according to claim 4, characterized in that, The method for identifying termination intent is as follows: construct a termination intent knowledge base that includes conversation greetings, conversation termination, and non-business scope knowledge, and identify whether the knowledge base to which the user's intent belongs is a termination intent knowledge base.
6. A cross-domain multi-turn question-answering system based on a large model, characterized in that, include: The knowledge base construction module is used to build several knowledge bases by domain. The cards of the knowledge base include knowledge base description, business rules and tag information. The domain intent recognition module is used to identify the knowledge base to which the user intent belongs. Specifically, it rewrites the user question and performs cascaded conditional judgments of rule intent recognition, semantic routing intent recognition, and language big model intent recognition based on the rewritten question, the user question, and cards in the knowledge base to determine the knowledge base to which the user intent belongs. In the second round of question and answer, after summarizing the context of the user's question, a new question is obtained from the summary. This new question is then used as the user's question to identify the knowledge base to which the user's intent belongs. The language big model intent recognition and answer module is used to output an answer to the user's question in each round of question and answer, based on the user's question and the knowledge base to which the user's intent belongs, using the big model of answering. When a user initiates a new round of question-and-answer dialogue, a large model is used to analyze the user's original question. Rewrite the code to obtain two rewrite problems. and ; Rule intent recognition: for the original problem and rewriting issues and The intent is identified by rules. If the intents of at least two questions match the same knowledge base, then the matched knowledge base is the user intent knowledge base; otherwise, semantic routing intent identification is performed. Semantic routing intent recognition: The Sentence-Bert model is used to encode the original question, the rewritten question and the tag information of each knowledge base, and the similarity matching algorithm is used to find the matching probability of each question and each knowledge base respectively. Based on the original question, the rewritten question, and the knowledge base. The matching probability is calculated, the average value is obtained, and it is determined whether the average value reaches a set threshold. If it does, the knowledge base... If the current user's question intent is the knowledge base, then perform large-scale language model intent recognition; Language Big Model Intent Recognition: Using a language big model to analyze the original problem and rewriting issues and Intent recognition is performed, and the knowledge base to which the user's intent belongs is selected based on the content of the knowledge cards in each knowledge base.
7. The cross-domain multi-turn question-answering system based on a large model according to claim 6, characterized in that, It also includes a session termination module, used to determine whether the conditions for ending a multi-round session are met during the process of a multi-round session. If the conditions are met, the session is terminated. The conditions for ending a multi-round session include: the user actively terminates the session after an operation, the session ends after a timeout, the user's question is located in different knowledge bases for consecutive rounds, in which case the current round of dialogue is forcibly ended, and termination intent recognition is performed. If the user's intent is identified as termination intent, the session ends after answering the question in the last round. The termination intent recognition is performed by constructing a termination intent knowledge base that includes conversational greetings, session termination, and non-business scope knowledge, and identifying whether the knowledge base to which the user's intent belongs is a termination intent knowledge base.
8. A computer device, characterized in that, It includes a processor and a memory; wherein, when the processor executes the computer program stored in the memory, it implements the steps of the cross-domain multi-turn question answering method based on a large model as described in any one of claims 1-5.
9. A computer-readable storage medium, characterized in that, Used to store computer programs; when the computer programs are executed by a processor, they implement the steps of the cross-domain multi-turn question answering method based on a large model as described in any one of claims 1-5.