Intelligent question and answer method, device and product

Abstract

The embodiment of the invention provides an intelligent question and answer method, device and product, and is applied to the technical field of insurance. The method comprises the following steps: in response to a received consultation service of a user, processing the consultation service by using an intelligent agent to obtain a reply result of the consultation service; wherein the intelligent agent is an intelligent agent which is trained or retrieved and enhanced by utilizing scene elements extracted by an expert seat dialogue log and strategy elements corresponding to the scene elements. The intelligent agent can determine the scene elements of the consultation service and determine the corresponding strategy elements according to the scene elements after the intelligent agent is trained or enhanced by using the scene elements extracted by the expert seat dialogue log and the corresponding strategy elements, and therefore the strategy elements are obtained on the basis of the business knowledge of the expert seat. In this way, the accuracy of the reply result can be improved, and the reply result conforms to the expectation of the user.

Description

Technical Field

[0001] This application relates to the insurance field, and more particularly to an intelligent question-answering method, device, and product. Background Technology

[0002] Intelligent agents (AI agents) refer to software requirements that, based on artificial intelligence technology, simulate human interaction logic and autonomously handle inquiries. They are widely used in customer service scenarios in fields such as insurance, finance, and e-commerce. Intelligent agents are used to replace or assist human agents in handling customer reception, business inquiries, and dispute resolution, thereby improving service efficiency. In existing technologies, intelligent agents often use large language models to directly process inquiries, obtaining corresponding responses based on the model's understanding capabilities. However, this method of directly processing inquiries using large language models suffers from poor accuracy in the responses. Summary of the Invention

[0003] This application provides an intelligent question-answering method, apparatus, and product to improve the accuracy of AI Agent's answers.

[0004] In a first aspect, embodiments of this application provide an intelligent question-answering method, the method comprising:

[0005] In response to receiving a user's inquiry, the system uses a smart agent to process the inquiry and obtain a response to the inquiry.

[0006] The intelligent agent is an enhanced intelligent agent trained or retrieved using scene elements extracted from expert agent dialogue logs and the corresponding policy elements.

[0007] Optionally, the method further includes:

[0008] The target dialogue log is obtained by removing the target data from the expert agent dialogue log; the target data includes system breakpoints, data that exceeds the model's capabilities, and data missing from the business knowledge base.

[0009] The target dialogue log is processed using a large language model to obtain the scene elements and the policy elements.

[0010] Optionally, processing the target dialogue log using a large language model to obtain the scene elements and the policy elements includes:

[0011] The target dialogue log is semantically segmented to obtain n semantically complete segments, where n is a positive integer.

[0012] The n semantically complete fragments are processed using a large language model to obtain the scene elements and the strategy elements.

[0013] Optionally, the target dialogue log is semantically segmented to obtain n semantically complete segments, including:

[0014] The target dialogue log is segmented according to a preset sliding window to obtain m initial window segments, where m is a positive integer.

[0015] Obtain local summary information for m initial window segments;

[0016] By using the summary chain, the local summary information of the m initial window segments is processed to obtain n semantically complete segments.

[0017] Optionally, the step of processing the n semantically complete fragments using a large language model to obtain the scene elements and the policy elements includes:

[0018] Using a lightweight classification model or pre-defined rules, identify key decision-making inflection points from n semantically complete fragments; the key decision-making inflection points include changes in customer emotion, the customer's first mention, and the agent's intention to respond.

[0019] The local context corresponding to the key decision-making turning point is processed using the large language model to obtain the scene elements and the strategy elements.

[0020] Optionally, processing the target dialogue log using a large language model to obtain the scene elements and the policy elements includes:

[0021] Using the large language model, the target dialogue log is processed by simulating multiple preset angles to obtain scene element fragments and strategy element fragments for each preset angle;

[0022] Among them, the multiple preset angles include the role of a top customer service representative, a composite role of risk compliance officer and marketing expert, the role of a customer psychology analyst, and / or the role of a strategy structuring expert.

[0023] By integrating scene element fragments and strategy element fragments from multiple preset angles, the scene elements and strategy elements are obtained.

[0024] Optionally, the method further includes:

[0025] Obtain a strategy knowledge base that includes the scene elements and the strategy elements;

[0026] The intelligent agent is trained using supervised fine-tuning SFT using the policy knowledge base to obtain an intelligent agent that internalizes the policy knowledge base.

[0027] Optionally, before processing the consultation service using a smart agent, the method further includes:

[0028] Obtain a strategy knowledge base that includes the scene elements and the strategy elements;

[0029] The consultation service is subjected to intent recognition, and policy elements irrelevant to the intent are filtered out; the consultation service and the current dialogue context are converted into vectors, and similarity retrieval is performed in the vector knowledge base to obtain the top k matching policy elements; where K is a positive integer;

[0030] The process of using intelligent agents to process the consultation service includes:

[0031] The intelligent agent is used to process the matched strategy elements and generate a response result.

[0032] Secondly, embodiments of this application provide an intelligent question-answering device, characterized in that the device comprises:

[0033] A response unit is configured to respond to a received consultation service by processing the consultation service using an intelligent agent to obtain a response result for the consultation service; wherein the intelligent agent is an enhanced intelligent agent.

[0034] The enhancement unit is used to train or retrieve enhanced intelligent agents using scene elements extracted from expert agent dialogue logs and the corresponding policy elements of the scene elements.

[0035] Thirdly, embodiments of this application provide an electronic device, including:

[0036] Memory, used to store computer programs;

[0037] A processor is used to execute computer programs to implement the methods described in any of the first aspects.

[0038] Fourthly, embodiments of this application provide a computer program that, when run on a computer, causes the computer to perform the method in any of the possible implementations of any of the above aspects.

[0039] Fifthly, embodiments of this application provide a computer-readable storage medium storing a computer program (also referred to as code or instructions) that, when run on a computer, causes the computer to perform the method in any of the possible implementations of any of the above aspects.

[0040] Sixthly, embodiments of this application provide a chip system including one or more processors for calling and executing instructions stored in memory, causing the methods in any of the above aspects or possible implementations to be executed. The chip system may be composed of chips or may include chips and other discrete devices.

[0041] This application provides an intelligent question-answering method, apparatus, and product. The method includes, in response to receiving a user's inquiry service, processing the inquiry service using an intelligent agent to obtain a response result; wherein the intelligent agent is an intelligent agent trained or enhanced using scene elements extracted from expert agent dialogue logs and corresponding strategy elements. By training or enhancing the intelligent agent using scene elements extracted from expert agent dialogue logs and corresponding strategy elements, the intelligent agent can determine the scene elements of the inquiry service and, based on these scene elements, determine the corresponding strategy elements. These strategy elements are obtained based on the expert agent's business knowledge; therefore, this approach can improve the accuracy of the response result and ensure that the response result meets the user's expectations. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart illustrating an intelligent question-answering method provided in this application embodiment;

[0044] Figure 2 This is a schematic diagram of the structure of an intelligent question-answering device provided in an embodiment of this application. Detailed Implementation

[0045] To enable those skilled in the art to better understand the present application, the technical solutions in this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] Current AI agents used for consultation services suffer from the following drawbacks: First, it's difficult to scalably transfer the business knowledge of top-performing agents; second, the massive amount of dialogue log data lacks structured extraction methods, making it difficult for AI agents to effectively elicit users' implicit needs and corresponding response strategies. This results in poor accuracy in the AI ​​agent's responses to consultation services.

[0047] In view of this, embodiments of this application provide an intelligent response method to address at least one of the aforementioned deficiencies, thereby improving the response accuracy of the AI ​​Agent and ensuring that the response results meet user expectations. In this method, the intelligent agent is an enhanced intelligent agent trained or retrieved using scenario elements extracted from expert agent dialogue logs and corresponding policy elements. By training or enhancing the intelligent agent using scenario elements extracted from expert agent dialogue logs and corresponding policy elements, the intelligent agent can determine the scenario elements of the consultation service. Based on these scenario elements, it determines the corresponding policy elements, which are obtained based on the expert agent's business knowledge. Therefore, this approach can improve the accuracy of the response results and ensure that the response results meet user expectations.

[0048] The intelligent question-answering method provided in this application can be applied to an AI Agent. This AI Agent can include a software system, which can be a standalone system or integrated into other software systems as a plugin, component, functional module, service, microservice, or mini-program. The software system can be provided to users as a software package for self-deployment. Alternatively, it can be provided as a cloud service, allowing users to subscribe and utilize the corresponding question-answering capabilities and methods. In some possible implementations, the AI ​​Agent can include a hardware system. This hardware system can include a cluster of computing devices with question-answering capabilities. When the cluster runs, it executes the question-answering method of this application. When executing a specific method, the AI ​​Agent can be an AI Agent trained using scene elements extracted from expert agent dialogue logs and corresponding strategy elements. This AI Agent can simulate the expert agent's response logic to obtain the corresponding response result. Specifically, the AI ​​Agent can receive user-inputted consultation services, process the consultation services using a large language model, and obtain the corresponding response result.

[0049] The intelligent question-answering method provided in the embodiments of this application will be described below with reference to the accompanying drawings.

[0050] Appendix Figure 1 A flowchart of an intelligent question-answering method provided in this application embodiment, the method including S10~S20:

[0051] S10: Consultation service that accepts user input.

[0052] In one example, the smart agent includes a dialog interface where users can input their inquiries. For instance, an inquiry might be, "I'm sick, I've already filed a claim and received compensation, can I still purchase insurance?"

[0053] It's understandable that consultation services often involve implicit needs and confusion regarding technical terminology. For example, a consultation asking "Can I still buy insurance after making a medical claim?" might implicitly reveal concerns such as "worrying about my health condition affecting my insurance application" or "wanting to know whether to renew or reapply for insurance." Directly processing these consultations using a large language model might result in insufficient accuracy in the responses.

[0054] S20: In response to receiving a consultation service, the intelligent agent processes the consultation service to obtain a response result for the consultation service.

[0055] The intelligent agent is an enhanced intelligent agent trained or retrieved using scene elements extracted from expert agent dialogue logs and the corresponding policy elements. It can be understood that this intelligent agent can simulate the expert agent and respond to inquiries; therefore, responses obtained using this intelligent agent have higher accuracy.

[0056] Scene elements describe the dialogue scenario, including but not limited to: customer status elements, dialogue stage elements, and external environment elements. Customer status elements include, but are not limited to, explicit intentions and implicit needs. Implicit needs include, but are not limited to: budget sensitivity, emotional state, knowledge level, relationship stage, and identity characteristics. Dialogue stage elements include, but are not limited to, icebreaking, needs assessment, product / service explanation, objection handling, closing, after-sales consultation, claims assistance, complaint handling, and small talk. External environment elements include, but are not limited to, product type and timeliness.

[0057] Strategic elements correspond to scenario elements; once the scenario elements are determined, the strategic elements are also determined. Strategic elements refer to the response strategies under the scenario elements. In the embodiments of this application, strategic elements include, but are not limited to: understanding and thinking, key information points, content to be avoided, expression methods and tone, core objectives, coping skills, success indicators, information priority, recommended dialogue structure, and potential risk warnings.

[0058] Expert agent dialogue logs refer to the complete dialogue records generated by expert agents during business interactions with customers. They are the core data source for extracting scenario elements and corresponding strategy elements. It should be noted that "expert agent" refers to the entity to be simulated; for example, an expert agent is one whose monthly conversion rate is in the top 15%, customer satisfaction is greater than or equal to a preset threshold, and there are no compliance or violation records for the quarter.

[0059] It is understandable that although expert agent dialogue logs contain high-quality response experience, they still contain invalid information. This invalid information is called target data. This target data can affect extraction accuracy. In one example, firstly, target data can be removed from the expert agent dialogue logs to obtain target dialogue logs; the target data includes system breakpoints, data beyond the model's capabilities, and data missing from the business knowledge base. Then, the target dialogue logs are processed using a large language model to obtain the scenario elements and the strategy elements, thereby eliminating the impact of target data on extraction accuracy.

[0060] System breakpoints refer to non-business interaction-related prompts generated during the dialogue, including but not limited to system timeouts, product unknowns, and query failures. Data exceeding the model's capabilities refers to data generated during the dialogue that is irrelevant to answering inquiries, including but not limited to agent status and non-interactive proactive marketing. Data lacking in the business knowledge base refers to data beyond the coverage of the business knowledge base, including but not limited to cross-domain customer questions and new knowledge not yet entered into the business knowledge base.

[0061] In this embodiment of the application, a large language model can be used to process the expert agent dialogue logs to extract scene elements and the corresponding strategy elements.

[0062] It's understandable that expert agent dialogue logs are lengthy. Directly processing these logs might lead to the "illusion problem," resulting in issues like missing fields, inconsistent descriptions, and omissions of key strategies in the extracted scene and strategy elements. Therefore, in some examples, the target dialogue log can be semantically segmented into n semantically complete fragments (n being a positive integer). Then, a large language model is used to process these n semantically complete fragments to obtain scene and strategy elements. By segmenting the logs according to semantic completeness using a large language model, the illusion problem caused by long texts can be reduced, improving extraction accuracy.

[0063] In some examples, to improve the completeness of semantic fragments, the target dialogue log can be segmented according to a preset sliding window to obtain m initial window fragments, where m is a positive integer; local summary information of the m initial window fragments is obtained; and the local summary information of the m initial window fragments is processed using a summary chain to obtain n semantically complete fragments. The size of the preset sliding window is adjusted accordingly.

[0064] For example: If the preset sliding window size is 5 rounds of dialogue and the step size is 2 rounds, the target dialogue log is initially segmented to obtain m initial window fragments. For each initial window fragment, a large language model is called to process it and generate local summary information. For example, the specific content of the initial window fragment is:

[0065] Customer: "If I make a claim under my critical illness insurance policy, can I renew the policy?"

[0066] Customer service representative: "Don't worry, this product guarantees renewal for 20 years. Even if you've made a claim, it will be renewed as long as it's within the coverage period."

[0067] Customer: "Will the premium increase when I renew my policy? Do I need to apply for it myself?"

[0068] The partial summary information of the initial window fragment obtained is: "Core topic: Customer inquires about the rules for renewal of critical illness insurance claims and premium / application methods; Interaction status: The agent answered the questions about renewal eligibility, but did not answer the questions about premium increases and application methods; Semantic completeness: Incomplete."

[0069] Because sliding windows have overlapping step sizes and the dialogue content of adjacent windows is semantically related, a large language model can be used to process the local summary information of adjacent initial windows to determine whether they belong to the same semantic theme. If so, they are merged to obtain a complete semantic fragment. For example, the local summary information of the previous window (rounds 1-5): Theme: Claim follow-up insurance; Status: Customer asks about premium increase, agent does not answer; Completeness: Incomplete; The local summary information of the next window (rounds 3-7): Theme: Claim follow-up insurance; Status: Agent answers that premiums fluctuate based on age, system automatically deducts fees, customer has no new questions; Completeness: Complete. The large language model is used to determine whether they belong to the same semantic continuation. If they are determined to belong to the same semantic continuation, the two fragments are merged, and the complete semantic fragment is output as: Core theme: Customer fully inquires about the rules for critical illness insurance claim follow-up insurance (eligibility, premium, application method); Interaction status: Agent fully answers, customer has no new questions; Semantic completeness: Complete. This combination of sliding windows and summary chains can efficiently and accurately segment the target dialogue log.

[0070] It's understandable that semantically complete fragments may contain non-critical interactive information, such as confirmatory statements like "Okay, I understand" from a customer. Extracting the entire content of such a fragment would increase invalid elements, leading to redundancy. Therefore, in some examples, a lightweight classification model or pre-defined rules can be used to identify key decision-making inflection points from n semantically complete fragments. Key decision-making inflection points include changes in customer emotion, the customer's first mention of something, and the agent's intention to respond. A large language model is used to process the local context corresponding to the key decision-making inflection points to obtain scenario and strategy elements. This approach allows focus on the core coping experience of expert agents, uncovering the customer's implicit needs, thereby improving extraction quality.

[0071] It should be noted that lightweight classification models refer to machine learning or deep learning models with a smaller parameter size than large language models, low training costs, and fast inference speed. By learning from labeled dialogue data, they can automatically identify key decision-making inflection points with implicit features that are difficult to define by rules, which helps to reduce processing complexity while improving extraction quality.

[0072] In some other examples, a large language model can be used to process the target dialogue log by simulating multiple preset perspectives, and obtain scene element fragments and strategy element fragments for each preset perspective. Among these preset perspectives, there are multiple preset perspectives, including the role of a top customer service representative, a composite role of risk compliance officer and marketing expert, the role of a customer psychology analyst, and / or the role of a strategy structuring expert. By integrating the scene element fragments and strategy element fragments from multiple preset perspectives, scene elements and strategy elements are obtained.

[0073] For example, using a large language model, multiple preset angles are simulated to obtain scene element fragments and corresponding strategy element fragments as follows:

[0074] Role 1 (Gold Medal Agent): You are now the top Gold Medal Agent of the month. Please review why you responded to the customer in this way in this round. Please think about it step by step.

[0075] Role 2 (Risk Compliance Officer & Marketing Expert): From a compliance perspective, what "absolutely unspeakable things" were implied in the agent's responses this time? What things were deliberately avoided (for compliance and marketing purposes)?

[0076] Role 3 (Client Psychological Analyst): What are the client's hidden concerns in this round? How did the agent perceive and resolve them?

[0077] Role 4 (Strategy Structure Expert): Synthesize the thoughts of the above three people and output scene elements and strategy elements strictly according to the following preset format.

[0078] Furthermore, self-consistency checks can be performed by repeatedly extracting the same dialogue 3 to 5 times and using a lightweight consistency scoring model to calculate the consistency score (e.g., residual similarity) of the multiple extraction results. Strategies with scores greater than or equal to the similarity threshold (e.g., 0.85) are used as strategy elements. This approach can further address the illusion problem and improve the accuracy of extracted scene elements and strategy elements.

[0079] In some examples, a policy knowledge base including scene elements and policy elements is obtained; the policy knowledge base is used to perform supervised fine-tuning (SFT) training on the intelligent agent to obtain an intelligent agent that internalizes the policy knowledge base.

[0080] In other examples, a policy knowledge base is used as an add-on to enhance the intelligent agent. Specifically, intent recognition is performed on the consultation service, filtering out policy elements irrelevant to the intent; the consultation service and the current dialogue context are converted into vectors, and a similarity search is performed in the vector knowledge base to obtain the top K matching policy elements; K is a positive integer; the intelligent agent processes the matching policy elements to generate a response.

[0081] Furthermore, embodiments of this application can further improve response accuracy by utilizing verification and negative example injection. Specifically, a portion of conversations between ordinary agents in similar scenarios are randomly selected, and strategy elements and scenario elements are extracted using the same method as described above. The differences between the strategy elements obtained by expert agents and those obtained by ordinary agents are compared; this difference information represents the expert agent's thinking pattern. Then, a large language model generates a response containing this difference, and the large language model learns the difference response, thereby improving the accuracy of the response results.

[0082] Furthermore, in this embodiment, the strategy knowledge base can be updated through a continuous online closed loop. Specifically, after the intelligent agent goes online, it searches and references strategy elements in the strategy knowledge base that are similar to the scenario of the current consultation service to obtain the response result. Then, it collects core indicators such as direct feedback from agents, customer satisfaction, dialogue duration, and conversion rate to back-label which strategies are more effective. Unlabeled strategies in the strategy knowledge base can be re-extracted along with their corresponding scenario elements.

[0083] This application provides an intelligent response method. By training an intelligent agent using scenario elements extracted from expert agent dialogue logs and corresponding strategy elements, or by enhancing the intelligent agent, the intelligent agent can determine the scenario elements of the consultation service and, based on these scenario elements, determine the corresponding strategy elements. These strategy elements are obtained based on the business knowledge of the expert agent. Therefore, this method can improve the accuracy of the response results and make the response results meet the user's expectations.

[0084] In addition, this application also provides an intelligent question-answering device.

[0085] Appendix Figure 2 This is a schematic diagram of the structure of an intelligent question-answering device provided in an embodiment of this application. The device 200 includes:

[0086] The response unit 201 is configured to respond to receiving a consultation service by processing the consultation service using an intelligent agent to obtain a response result for the consultation service; wherein the intelligent agent is an enhanced intelligent agent.

[0087] The enhancement unit 202 is used to train or retrieve enhanced intelligent agents using scene elements extracted from expert agent dialogue logs and the corresponding policy elements of the scene elements.

[0088] Optionally, the device 200 further includes an extraction unit for:

[0089] The target dialogue log is obtained by removing the target data from the expert agent dialogue log; the target data includes system breakpoints, data that exceeds the model's capabilities, and data missing from the business knowledge base.

[0090] The target dialogue log is processed using a large language model to obtain the scene elements and the policy elements.

[0091] Optionally, processing the target dialogue log using a large language model to obtain the scene elements and the policy elements includes:

[0092] The target dialogue log is semantically segmented to obtain n semantically complete segments, where n is a positive integer.

[0093] The n semantically complete fragments are processed using a large language model to obtain the scene elements and the strategy elements.

[0094] Optionally, the target dialogue log is semantically segmented to obtain n semantically complete segments, including:

[0095] The target dialogue log is segmented according to a preset sliding window to obtain m initial window segments, where m is a positive integer.

[0096] Obtain local summary information for m initial window segments;

[0097] By using the summary chain, the local summary information of the m initial window segments is processed to obtain n semantically complete segments.

[0098] Optionally, the step of processing the n semantically complete fragments using a large language model to obtain the scene elements and the policy elements includes:

[0099] Using a lightweight classification model or pre-defined rules, identify key decision-making inflection points from n semantically complete fragments; the key decision-making inflection points include changes in customer emotion, the customer's first mention, and the agent's intention to respond.

[0100] The local context corresponding to the key decision-making turning point is processed using the large language model to obtain the scene elements and the strategy elements.

[0101] Optionally, processing the target dialogue log using a large language model to obtain the scene elements and the policy elements includes:

[0102] Using the large language model, the target dialogue log is processed by simulating multiple preset angles to obtain scene element fragments and strategy element fragments for each preset angle;

[0103] Among them, the multiple preset angles include the role of a top customer service representative, a composite role of risk compliance officer and marketing expert, the role of a customer psychology analyst, and / or the role of a strategy structuring expert.

[0104] By integrating scene element fragments and strategy element fragments from multiple preset angles, the scene elements and strategy elements are obtained.

[0105] Optionally, the enhancement unit 202 is specifically used for:

[0106] Obtain a strategy knowledge base that includes the scene elements and the strategy elements;

[0107] The intelligent agent is trained using supervised fine-tuning SFT using the policy knowledge base to obtain an intelligent agent that internalizes the policy knowledge base.

[0108] Optionally, the agent processing unit 203 is also used for:

[0109] Obtain a strategy knowledge base that includes the scene elements and the strategy elements;

[0110] The consultation service is subjected to intent recognition, and policy elements irrelevant to the intent are filtered out; the consultation service and the current dialogue context are converted into vectors, and similarity retrieval is performed in the vector knowledge base to obtain the top K matching policy elements; where K is a positive integer;

[0111] The process of using intelligent agents to process the consultation service includes:

[0112] The intelligent agent is used to process the matched strategy elements and generate a response result.

[0113] This application provides an intelligent response device. In this device, the intelligent agent is an enhanced intelligent agent trained or retrieved using scene elements extracted from expert agent dialogue logs and corresponding policy elements. By training or enhancing the intelligent agent using scene elements extracted from expert agent dialogue logs and corresponding policy elements, the intelligent agent can resolve the problem of ambiguous intent in consultation services and improve the accuracy of responses.

[0114] According to the method provided in the embodiments of this application, this application also provides a chip system, which includes one or more processors for calling and executing instructions stored in memory, thereby causing the method described in the embodiments of this application to be executed. The chip system may be composed of chips or may include chips and other discrete devices.

[0115] The chip system may include input circuits or interfaces for transmitting information or data, and output circuits or interfaces for receiving information or data.

[0116] According to the method provided in the embodiments of this application, this application also provides a computer program product, which includes: computer program code, which, when run on a computer, causes the computer to execute the various steps or processes executed by the network device or terminal device in any of the foregoing method embodiments.

[0117] According to the method provided in the embodiments of this application, this application also provides a computer-readable storage medium storing program code, which, when run on a computer, causes the computer to execute the various steps or processes executed by the network device or terminal device in any of the foregoing method embodiments.

[0118] The computer-readable storage medium may be the aforementioned volatile memory or non-volatile memory, or it may include both volatile memory and non-volatile memory.

[0119] In the embodiments of this application, the terms and English abbreviations are exemplary examples given for ease of description and should not be construed as limiting the application in any way. This application does not preclude the possibility of defining other terms that can achieve the same or similar functions in existing or future agreements.

[0120] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated.

[0121] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

Claims

1. An intelligent question answering method, characterized by, The method comprises: in response to receiving the user's consultation service, processing the consultation service by using an intelligent agent to obtain a reply result of the consultation service; wherein the intelligent agent is an enhanced intelligent agent trained or retrieved by using scene elements extracted from expert agent dialogue logs and policy elements corresponding to the scene elements.

2. The method of claim 1, wherein, The method further comprises: eliminating target data from the expert agent dialogue logs to obtain target dialogue logs; the target data includes system breakpoints, data beyond the model's capabilities, and missing data in the business knowledge base; processing the target dialogue logs by using a large language model to obtain the scene elements and the policy elements.

3. The method of claim 2, wherein, The processing of the target dialogue logs by using a large language model to obtain the scene elements and the policy elements comprises: cutting the target dialogue logs according to semantics to obtain n semantic complete segments, n being a positive integer; processing n semantic complete segments by using a large language model to obtain the scene elements and the policy elements.

4. The method of claim 3, wherein, The cutting of the target dialogue logs according to semantics to obtain n semantic complete segments comprises: cutting the target dialogue logs according to a preset sliding window to obtain m initial window segments, m being a positive integer; obtaining local summary information of the m initial window segments; processing the local summary information of the m initial window segments by using a summary chain to obtain n semantic complete segments.

5. The method of claim 4, wherein, The processing of n semantic complete segments by using a large language model to obtain the scene elements and the policy elements comprises: identifying key decision turning points from n semantic complete segments by using a lightweight classification model or a preset rule; the key decision turning points include customer emotional changes, customer first mentions, and agent reply intentions; processing local contexts corresponding to the key decision turning points by using the large language model to obtain the scene elements and the policy elements.

6. The method of claim 2, wherein, The processing of the target dialogue logs by using a large language model to obtain the scene elements and the policy elements comprises: processing the target dialogue logs by using the large language model through simulating multiple preset angles to obtain scene element segments and policy element segments for each preset angle; wherein the multiple preset angles include a gold medal agent role, a risk compliance officer and marketing expert composite role, a customer psychological analyst role, and / or a policy structured expert role; integrating scene element segments and policy element segments of multiple preset angles to obtain the scene elements and the policy elements.

7. The method of claim 1, wherein, The method further comprises: obtaining a policy knowledge base including the scene elements and the policy elements; supervising SFT fine-tuning training of the intelligent agent by using the policy knowledge base to obtain an intelligent agent internalizing the policy knowledge base.

8. The method of claim 1, wherein, Before the processing of the consultation service by using an intelligent agent, the method further comprises: obtaining a policy knowledge base including the scene elements and the policy elements; The consultation service is subjected to intention recognition, a policy element irrelevant to the intention is filtered, the consultation service and a current dialogue context are converted into a vector, a similarity search is performed in the vector knowledge base, and k matching policy elements are obtained; k is a positive integer; The consultation service is processed by using the intelligent agent, and the processing includes: The matched policy elements are processed by using the intelligent agent to generate a reply result.

9. An intelligent question answering apparatus, characterized by comprising: The device includes: A response unit is configured to, in response to receiving a consultation service, process the consultation service by using an intelligent agent to obtain a reply result of the consultation service; the intelligent agent is an enhanced intelligent agent; An enhancement unit is configured to train or search the enhanced intelligent agent by using scene elements extracted from expert agent dialogue logs and policy elements corresponding to the scene elements.

10. An electronic device, comprising: The device includes: A memory is configured to store a computer program; A processor is configured to execute the computer program to implement the method according to any one of claims 1 to 8.

Images