Problem processing method and device based on artificial intelligence, computer equipment and medium

Abstract

The invention belongs to the technical field of artificial intelligence, and relates to an artificial intelligence-based question processing method and device, computer equipment and a storage medium, and the method comprises the steps: receiving question data inputted by a user; performing data generation processing on the problem data based on an enhancement generation strategy to obtain enhanced problem data; performing combination processing on the problem data and the enhanced problem data to obtain combined data; performing reasoning processing on the combined data based on the target large language model to obtain an output result, and performing consistency verification on the output result to obtain a consistency verification result; performing compliance detection and fact alignment verification processing on the output result to obtain a compliance detection result and a fact verification result; based on the consistency verification result, the compliance detection result and the fact verification result, performing answer generation processing to obtain answer data; and performing output processing on the answer data. The method can be applied to question and answer processing scenes in the field of financial science and technology, and stability and accuracy of answer generation are effectively improved through the method.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology and can be applied to the financial technology field, particularly to artificial intelligence-based problem-solving methods, devices, computer equipment, and storage media. Background Technology

[0002] While Large Language Models (LLMs) have been applied to various tasks in traditional financial and insurance services, including claims consultation, risk assessment, policy interpretation, and customer service, significant problems remain in problem-solving. Specifically, current models are susceptible to appeasement bias, tending to generate responses by "pleasing the user" rather than strictly adhering to factual logic. For example, when a customer questions or misunderstands the insurance terms, the model may output responses inconsistent with the contract terms due to excessive catering to the user's emotions, resulting in answers lacking data stability and accuracy, which can lead to claims disputes or a decline in customer trust.

[0003] Taking claims consultation in the financial sector as an example, if a customer misunderstands the exclusion clauses when applying for compensation due to an accident (such as believing that "injuries outside of professional sports" should be covered), traditional models may, in order to cater to the user's expectation of quick compensation, misinterpret the scope of the clauses or obscure key limiting conditions, ultimately leading to a claim conclusion that does not match the contract agreement. This deviation not only harms the rights and interests of customers but also increases the operational risks and compliance costs of insurance companies.

[0004] Therefore, there is an urgent need to develop a problem-solving method that can improve the stability and accuracy of generated answers, thereby enhancing service reliability and reducing legal risks. Summary of the Invention

[0005] The purpose of this application is to propose a problem-solving method, apparatus, computer device, and storage medium based on artificial intelligence, so as to solve the technical problem that existing problem-solving methods based on large language models lack data stability and accuracy in generating answers.

[0006] Firstly, an artificial intelligence-based problem-solving method is provided, including: Receive user-input question data; The problem data is processed based on a preset enhancement generation strategy to obtain corresponding enhanced problem data; The problem data and the enhanced problem data are combined to obtain the corresponding combined data; The combined data is inferred based on the preset target large language model to obtain the output result, and the output result is subjected to consistency verification to obtain the corresponding consistency verification result. The output results are subjected to compliance testing and fact alignment verification to obtain the corresponding compliance testing results and fact verification results. Based on the consistency verification result, the compliance detection result, and the fact verification result, answer generation processing is performed to obtain the corresponding answer data; The answer data is then processed for output.

[0007] Secondly, an artificial intelligence-based problem-solving device is provided, comprising: The receiving module is used to receive user-input question data; The first generation module is used to perform data generation processing on the problem data based on a preset enhancement generation strategy to obtain corresponding enhanced problem data; The combination module is used to combine the problem data and the enhanced problem data to obtain corresponding combined data; The processing module is used to perform inference processing on the combined data based on a preset target large language model to obtain the output result, and to perform consistency verification on the output result to obtain the corresponding consistency verification result. The verification module is used to perform compliance detection and fact alignment verification on the output results to obtain the corresponding compliance detection results and fact verification results. The second generation module is used to perform answer generation processing based on the consistency verification result, the compliance detection result and the fact verification result to obtain the corresponding answer data; The output module is used to process the output of the answer data.

[0008] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described artificial intelligence-based problem-solving method.

[0009] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the aforementioned problem-solving method based on artificial intelligence.

[0010] In the above-described solution implemented by the AI-based problem-solving method, apparatus, computer equipment, and storage medium, the user-inputted problem data is first received; then, the problem data is processed to generate corresponding enhanced problem data based on a preset enhancement generation strategy; the problem data and the enhanced problem data are combined to obtain corresponding combined data; subsequently, the combined data is processed for inference based on a preset target large language model to obtain an output result, and the output result is subjected to consistency verification to obtain a corresponding consistency verification result; subsequently, the output result is subjected to compliance detection and fact alignment verification to obtain corresponding compliance detection results and fact verification results; further, the answer generation process is performed based on the consistency verification results, the compliance detection results, and the fact verification results to obtain corresponding answer data; finally, the answer data is output. Based on the above automated processing flow, this application generates enhanced question data from question data using an enhanced generation strategy. Then, it infers the combined data containing the question data and enhanced question data using a target large language model to obtain the output result. The output result is then subjected to consistency verification, compliance checks, and fact alignment verification. Finally, based on the obtained consistency verification results, compliance check results, and fact alignment verification results, answer generation processing is performed to obtain the final output answer data. Consistency verification ensures the consistency and stability of the processing of question data and enhanced question data, while compliance checks and fact alignment verification ensure the legality and accuracy of the question processing results. This effectively guarantees the stability and accuracy of the answer data generated based on the consistency verification results, compliance check results, and fact alignment verification results. Attached Figure Description

[0011] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is an exemplary system architecture diagram to which this application can be applied; Figure 2 This is a flowchart of an embodiment of the problem-solving method based on artificial intelligence according to this application; Figure 3 This is a schematic diagram of a structure of an embodiment of the artificial intelligence-based problem processing device according to this application; Figure 4 This is a schematic diagram of the structure of one embodiment of the computer device according to this application. Detailed Implementation

[0013] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0014] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0015] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0016] like Figure 1 As shown, system architecture 100 may include terminal device 101, network 102, and server 103. Terminal device 101 may be a laptop 1011, tablet 1012, or mobile phone 1013. Network 102 is used as a medium to provide a communication link between terminal device 101 and server 103. Network 102 may include various connection types, such as wired, wireless communication links, or fiber optic cables, etc.

[0017] Users can use terminal device 101 to interact with server 103 via network 102 to receive or send messages, etc. Various communication client applications can be installed on terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0018] Terminal device 101 can be various electronic devices with a display screen and support web browsing. In addition to laptops 1011, tablets 1012, or mobile phones 1013, terminal device 101 can also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 player (Moving Picture Experts Group Audio Layer IV), a laptop computer, and a desktop computer, etc.

[0019] Server 103 can be a server that provides various services, such as a backend server that provides support for the pages displayed on terminal device 101.

[0020] It should be noted that the problem-solving method based on artificial intelligence provided in this application is generally executed by a server / terminal device, and correspondingly, the problem-solving device based on artificial intelligence is generally set in the server / terminal device.

[0021] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0022] Continue to refer to Figure 2 This document illustrates a flowchart of an embodiment of the AI-based problem-solving method according to this application. The order of steps in the flowchart can be changed, and some steps can be omitted, depending on different needs. The AI-based problem-solving method provided in this application can be applied to any scenario requiring problem-solving, and therefore can be applied to products in these scenarios, such as problem-solving products in the financial insurance field. The AI-based problem-solving method includes the following steps: Step S201: Receive the question data input by the user.

[0023] In this embodiment, the problem-solving method based on artificial intelligence runs on an electronic device (e.g., Figure 1The server / terminal device shown can acquire user-inputted question data via wired or wireless connection. It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G / 5G connections, WiFi connections, Bluetooth connections, WiMAX connections, Zigbee connections, UWB (ultra-wideband) connections, and other currently known or future-developed wireless connection methods. The implementing entity of this application is specifically a question processing system, which may be simply referred to as the system.

[0024] This application can be applied to scenarios such as insurance product Q&A, claims underwriting, customer consultation, and compliance testing in the financial and insurance sector, to significantly reduce misjudgments, non-compliant responses, and public opinion risks. For example, in the insurance product Q&A scenario, the user's input could include: "I just bought a new car and don't know much about car insurance. I heard that compulsory traffic accident liability insurance is mandatory. Besides compulsory traffic accident liability insurance, what other commercial insurance should I buy to provide more comprehensive protection for my car? Are things like vehicle damage insurance and third-party liability insurance necessary?" In the claims underwriting scenario, the user's input could include: "I had a minor accident last year and reported it to the insurance company. My car insurance is about to expire this year, and when I went to renew it, the insurance company said my underwriting situation had changed, and the premium might increase. I'd like to know what specific impact an accident has on underwriting, and how much the premium will likely increase?"

[0025] Step S202: Based on a preset enhancement generation strategy, perform data generation processing on the problem data to obtain corresponding enhanced problem data.

[0026] In this embodiment, the specific implementation process of generating and processing the problem data based on the preset enhancement generation strategy to obtain the corresponding enhanced problem data will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0027] Step S203: Combine the problem data and the enhanced problem data to obtain the corresponding combined data.

[0028] In this embodiment, the aforementioned problem data and enhanced problem data can be combined by constructing pairs, and the resulting problem pairs can be used as the corresponding combined data.

[0029] Step S204: Based on the preset target large language model, perform reasoning processing on the combined data to obtain the output result, and perform consistency verification on the output result to obtain the corresponding consistency verification result.

[0030] In this embodiment, the specific implementation process of reasoning the combined data based on the preset target large language model to obtain the output result, and performing consistency verification on the output result to obtain the corresponding consistency verification result will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0031] Step S205: Perform compliance detection and fact alignment verification on the output results to obtain the corresponding compliance detection results and fact verification results.

[0032] In this embodiment, the specific implementation process of performing compliance detection and fact alignment verification on the output results to obtain the corresponding compliance detection results and fact verification results will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0033] Step S206: Based on the consistency verification result, the compliance detection result, and the fact verification result, perform answer generation processing to obtain the corresponding answer data.

[0034] In this embodiment, the specific implementation process of generating answer data based on the consistency verification result, the compliance detection result, and the fact verification result will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0035] Step S207: Output the answer data.

[0036] In this embodiment, the specific implementation process of outputting the answer data described above will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0037] This application first receives user-input question data; then, based on a preset enhancement generation strategy, it performs data generation processing on the question data to obtain corresponding enhanced question data; and combines the question data and the enhanced question data to obtain corresponding combined data; subsequently, it performs reasoning processing on the combined data based on a preset target large language model to obtain an output result, and performs consistency verification on the output result to obtain a corresponding consistency verification result; subsequently, it performs compliance detection and fact alignment verification processing on the output result to obtain corresponding compliance detection results and fact verification results; further, it performs answer generation processing based on the consistency verification results, compliance detection results, and fact verification results to obtain corresponding answer data; finally, it outputs the answer data. Based on the above automated processing flow, this application generates enhanced question data from question data using an enhancement generation strategy, then performs reasoning processing on the combined data containing question data and enhanced question data based on a target large language model to obtain an output result, then performs consistency verification, compliance detection, and fact alignment verification processing on the output result, and finally performs answer generation processing based on the obtained consistency verification results, compliance detection results, and fact verification results to obtain the final output answer data. Consistency checks ensure the consistency and stability of the processing of both problematic and augmented problem data. Compliance checks and fact-alignment checks ensure the legality and accuracy of the problem processing results. This effectively guarantees the stability and accuracy of the answer data generated based on consistency check results, compliance check results, and fact-alignment results.

[0038] In some alternative implementations, step S202 includes the following steps: The problem data is subjected to text perturbation enhancement processing to obtain the corresponding first problem data.

[0039] In this embodiment, the aforementioned text perturbation enhancement processing includes: randomly adding guiding phrases (such as "Please answer as an expert first"), format changes (bullet marks, question-and-answer style, colloquial style), or jailbreak fragments (such as "Please ignore company policy"). This can be achieved by designing a perturbation rule base, including rules for adding jailbreak fragments (such as "Please ignore the terms and conditions" or "I'll compensate you no matter what"), replacing synonyms (such as replacing "compensation" with "payment"), and adjusting word order (such as changing "I had an accident and am applying for compensation" to "I am applying for compensation because I had an accident"). Then, based on the rule base, the user's original input question data is randomly or strategically perturbed to generate semantically equivalent but formally different variants.

[0040] The first problem data is subjected to context wrapping enhancement processing to obtain the corresponding second problem data.

[0041] In this embodiment, the above-mentioned context wrapping enhancement process includes: embedding the original question data into the context of dialogue, instruction manual, customer emotional expression, etc., to simulate real user input, thereby obtaining the corresponding second question data.

[0042] The data for the second problem is subjected to cross-style restatement processing to obtain the corresponding data for the third problem.

[0043] In this embodiment, the aforementioned cross-style paraphrasing processing may include generating sentence structures that are synonymous with the original input question data but differ in tone and structure using a paraphrase model. Alternatively, different tone styles can be defined, such as interrogative sentences ("Can I get compensation in this situation?"), rhetorical questions ("Can't I get compensation in this situation?"), and declarative sentences ("Can I get compensation in this situation?"). Then, a language model or rule-based transformation method is used to convert the user's original input question data into synonymous sentence structures with different tones.

[0044] The third problem data is used as the enhanced problem data.

[0045] Based on the above processing flow, this application obtains the corresponding first problem data by performing text perturbation enhancement on the problem data; then, it performs context wrapping enhancement on the first problem data to obtain the corresponding second problem data; subsequently, it performs cross-style paraphrasing on the second problem data to obtain the corresponding third problem data; and finally, it uses the third problem data as enhanced problem data. Thus, this application, by performing text perturbation enhancement, context wrapping enhancement, and cross-style paraphrasing on the problem data, comprehensively utilizes multiple strategies to generate corresponding enhanced problem data, enriching and expanding the original input problem data from multiple dimensions to obtain more comprehensive and diverse enhanced problem data.

[0046] In some optional implementations of this embodiment, step S204 includes the following steps: The pre-trained target large language model is invoked to perform parallel inference processing on the combined data to obtain the corresponding output results.

[0047] In this embodiment, by inputting a combination of problem data and augmented problem data into a trained target large language model, the target large language model generates a corresponding output probability distribution for each input.

[0048] Among these measures, the consistency of the model's output behavior under different forms of cues can be constrained through a self-supervised learning mechanism, thereby preventing logical deviations, factual biases, and illegal outputs caused by cues changes. The model construction process of the aforementioned target large language model includes: (I) Module 1: Hint Enhancement Generation Module. This module uses a multi-strategy generation mechanism to construct semantically equivalent hint / question pairs but with different expressions for the same financial task. Text perturbation enhancement: Randomly add prompts (e.g., "Please answer as an expert first"), format variations (bulletin, question-and-answer style, colloquial), or jailbreak snippets (e.g., "Please ignore company policy"). Context wrapping enhancement: Embed the original question into contexts such as dialogues, instructions, and customer emotional expressions to simulate real user input. Cross-style paraphrase enhancement: Generate synonymous sentences with different tones and structures using a paraphrase model.

[0049] Output dataset:

[0050] in, For standard prompts, To enhance the prompts, This is a reference answer (from expert or high-confidence model output).

[0051] (ii) Module 2: External Output Consistency Constraint Module. This module constrains the consistency of the model's output results under different prompts. It sets the probability distribution of the model's output. and Define consistency loss:

[0052] That is, to minimize the difference in output distribution of the same semantic task under different prompts, and to ensure that the model “says different things but arrives at the same conclusion”.

[0053] To further ensure factual accuracy and compliance, a fact alignment penalty is introduced:

[0054] in, It is a fact extraction encoder used to extract key facts from the output (such as compensation amount, clause number, risk conclusion, etc.).

[0055] (III) Module 3: Internal Representation Consistency Constraint Module. Constraining only the output layer may not be sufficient to prevent jailbreak attacks. Therefore, this application innovatively introduces internal representation consistency constraints: Hidden Alignment: Activation representation of intermediate layers , Calculate cosine similarity:

[0056] in, This refers to the selected key layers (usually the last 3 layers of the transformer encoder). For layer weights.

[0057] Gradient invariance constraint: Limiting the gradient direction of the model to not shift too much under augmentation hints:

[0058] To prevent the model from generating different internal decision paths in response to "inductive instructions".

[0059] Overall consistency constraint loss:

[0060] Among them, for weight , , , The value can be set according to actual business needs.

[0061] (iv) Module 4: Compliance Detection and Fact Alignment Module. In the insurance scenario, the model must not only be consistent but also compliant and truthful. Therefore, this application introduces a compliance aligner: Regulatory Knowledge Base Embedding (RAG Integration): Insurance regulatory documents, terms and conditions manuals, sales script standards, etc., are stored in the knowledge base in vector form; during retrieval, the following is executed:

[0062] in, : Input content Embedding is used to convert text into vector form so that similarity calculations and other operations can be performed in the vector space. : For knowledge base The content (insurance regulatory documents, terms and conditions manuals, sales script standards, etc.) is vectorized, and all the text in the knowledge base is converted into vectors to facilitate subsequent retrieval. Retrieved knowledge refers to knowledge information most relevant to the input content that is retrieved from the knowledge base using specific retrieval methods. A retrieval strategy that involves selecting the top results with the highest similarity from the search results. One result. : Calculate the input vector With knowledge base vectors Similarity between vectors is typically measured using methods such as cosine similarity to measure the degree of similarity between two vectors in a given direction.

[0063] Compliance filtering: using a fine-tuned classifier Determine if the output violates regulations or policies; if it does, return a negative feedback sample. Participate in consistency training to enhance the model's rejection ability. Among these, The input content for the negative feedback sample, and the negative feedback output label. Correspondingly, the input data is used for consistency training. The output label of the negative feedback sample is the erroneous output label corresponding to the negative feedback sample generated when the model output violates regulations or policies. It is used to participate in consistency training.

[0064] Fact alignment verification: The consistency of key information in the model's response (such as compensation amount and deductible clauses) is verified using a knowledge base.

[0065] in, The Fact Alignment Score measures the degree of consistency between key information in the model's response and relevant regulatory information in the knowledge base. : Vectorize key information in the model's answer, such as vector form of key content like payout amount and deductible clauses. : Vectorize the legal content in the knowledge base that is related to the key information of the model's answer, and use it to verify the consistency with the key information vector of the model's answer. Cosine similarity is a method for calculating the similarity between two vectors. It measures the degree of similarity in direction by calculating the cosine of the angle between the two vectors. The closer the value is to 1, the more similar they are.

[0066] If the value is below the threshold, a rollback and re-push mechanism is triggered.

[0067] (v) Module 5: Training and Evaluation Engine.

[0068] Overall optimization objective function:

[0069] in: Losses from the original task (such as losses due to incorrect answers, compliant classification, etc.); Consistency training intensity weights.

[0070] The training process includes: Enhanced prompt pairs are constructed from standard business question-and-answer data; multi-round consistency constraint training is performed with the support of the RAG knowledge base; periodic jailbreak simulation tests (inserting bypass instructions, emotional expressions, etc.) are conducted to verify stability; if the output deviation of the model under enhanced prompts is less than the threshold, the model is considered to have been successfully trained and enters the deployment stage.

[0071] Evaluation metric: Consistency Score

[0072] Rejection robustness: The proportion of improper requests rejected by the model when prompted by jailbreak.

[0073] Fact retention rate: The proportion of output facts that are consistent with the knowledge base.

[0074] The output results are subjected to external output consistency verification to obtain the corresponding first verification result.

[0075] In this embodiment, KL divergence (KL divergence) can be used as an indicator to measure the difference between two probability distributions. The KL divergence between the output probability distributions corresponding to the original problem data and the augmented problem data is calculated. A divergence threshold is set. If the calculated KL divergence exceeds this threshold, it indicates a significant difference between the two output distributions, triggering an internal verification process and generating a verification result indicating that the external output consistency verification failed. Otherwise, a verification result indicating that the external output consistency verification passed is generated.

[0076] The output result is subjected to internal representation consistency verification to obtain the corresponding second verification result.

[0077] In this embodiment, during model inference, the intermediate hidden layer representations of the target large language model (hereinafter referred to as the model) for the input question data and the augmented question data are extracted. The similarity between these hidden layer representations, such as cosine similarity, is calculated. If the similarity is below a certain threshold, it indicates a significant difference in the model's internal processing of the two inputs, potentially posing a risk. Simultaneously, the gradient changes of the model when processing the input question data and the augmented question data are analyzed. If the gradient changes differ excessively, it indicates inconsistent sensitivity of the model to the two inputs, potentially affecting the stability of the output. The generated second verification result may include whether the internal representation consistency check passes or fails. If the internal representation consistency check fails, the input question data is marked as a high-risk input and enters the compliance detection process to further check whether the output is compliant.

[0078] The first verification result and the second verification result are integrated to obtain the corresponding target verification result.

[0079] In this embodiment, the generated first verification result and the second verification result can be integrated, and the resulting target verification result can be used as the corresponding consistency verification result.

[0080] The target verification result is used as the consistency verification result.

[0081] Based on the above processing flow, this application uses a pre-trained target large language model to perform parallel inference processing on the combined data to obtain the corresponding output results. Then, it performs external output consistency verification on the output results to obtain the corresponding first verification result. Next, it performs internal representation consistency verification on the output results to obtain the corresponding second verification result. Subsequently, it integrates the first and second verification results to obtain the corresponding target verification result. Finally, it uses the target verification result as the consistency verification result. Thus, this application obtains output results by using a target large language model to perform parallel inference on the combined data, then performs external output consistency verification and internal representation consistency verification on the output results, and integrates the generated first and second verification results to obtain the corresponding consistency verification result. In this way, this application ensures the consistency and stability of the system's processing of the original input problem data and the enhanced problem data through model parallel inference and consistency constraints. Model parallel inference can process different forms of input simultaneously, improving processing efficiency; calculating the external output consistency loss can quickly detect differences in output distribution and preliminarily determine whether the system has abnormal processing conditions; internal representation consistency verification checks at a deeper level whether the model's internal processing mechanism for different inputs is consistent. These steps effectively identify high-risk inputs, prevent the system from generating unreasonable or irregular outputs due to changes in inputs, and ensure the quality and security of consultation processing.

[0082] In some alternative implementations, step S205 includes the following steps: Call the preset legal knowledge base.

[0083] In this embodiment, a regulatory knowledge base containing information such as insurance-related laws and regulations, clauses, and regulatory requirements is pre-built.

[0084] Based on the problem data and the output results, the corresponding regulatory information is retrieved from the regulatory knowledge base.

[0085] In this embodiment, by utilizing Retrieval Enhancement Generation (RAG) technology, the user-input question data and the output results of the aforementioned target large language model can be used as query conditions to retrieve relevant regulatory information from the aforementioned regulatory knowledge base.

[0086] The output results and the regulatory information are compared using a preset regression classifier to perform compliance checks, resulting in the corresponding compliance check results.

[0087] In this embodiment, a compliance classifier is pre-trained. Based on a large amount of labeled compliance and violation data, this classifier can determine whether the model output violates regulations. The retrieved regulatory information and the model output (output result) can be input into the compliance classifier, which then provides a judgment result on whether the output is compliant, serving as the regression detection result mentioned above. For example, if the model output contains content such as "promise to accommodate claims," ​​the compliance classifier will determine it as a violation.

[0088] Key factual information is extracted from the output.

[0089] In this embodiment, key factual information, such as compensation amount, disclaimers, and claim conditions, can be extracted from the output results using NLP technology. This extraction can be achieved through methods such as keyword matching and entity recognition.

[0090] The key factual information is subjected to fact alignment verification to obtain the corresponding fact verification result.

[0091] In this embodiment, the extracted key factual information can be compared with standard facts in a regulatory knowledge base to calculate their cosine similarity. A similarity threshold is set; if the calculated cosine similarity is lower than this threshold, it indicates a significant difference between the facts output by the model and the standard facts, and a corresponding fact verification result (fact verification shows a discrepancy) is generated. Furthermore, when the cosine similarity is lower than the similarity threshold, a fallback mechanism is triggered. This fallback mechanism can be to revert to standard wording, such as "According to the provisions of the clause, this situation should be handled in this way..."; or to submit the inquiry for manual review by professionals for further judgment and processing.

[0092] Based on the above processing flow, this application invokes a pre-defined regulatory knowledge base; then, based on the question data and output results, it retrieves the corresponding regulatory information from the knowledge base; subsequently, it performs compliance checks on the output results and regulatory information using a pre-defined regression classifier, obtaining the corresponding compliance check results; and extracts key factual information from the output results; finally, it performs fact alignment verification on the key factual information, obtaining the corresponding fact verification results. Thus, the compliance checks and fact alignment verification provided by this application are crucial steps in ensuring the legality and accuracy of the answers to the question processing. Compliance checks, through the regulatory knowledge base and compliance classifier, can promptly detect potential violations in the model output, preventing the system from providing answers that violate laws and regulations, and protecting the legitimate rights and interests of enterprises and users. Fact alignment focuses on checking whether the key facts in the model output are consistent with standard facts, preventing the spread of inaccurate information due to model errors or data deviations.

[0093] In some alternative implementations, step S206 includes the following steps: The consistency verification result, the compliance detection result, and the fact verification result are integrated to obtain the corresponding integrated information.

[0094] In this embodiment, the generated consistency verification results, compliance detection results, and fact verification results can be classified and organized to obtain integrated information. For example, consistency verification-related data can be grouped into one category, compliance detection results into another, and fact alignment results into a third, so as to facilitate subsequent comprehensive analysis.

[0095] Logical analysis is performed based on the integrated information to obtain the corresponding logical analysis results.

[0096] In this embodiment, the system performs logical judgments based on the integrated information to determine the standard answer. For example, in an insurance claims consultation scenario, if the consistency check shows that the processing results of the original input and the enhanced prompts are consistent, the compliance check finds no violations, and the key facts in the fact alignment match the standard facts, then the system can determine a positive claim outcome (if the conditions are met) or a claim rejection outcome (if the conditions are not met).

[0097] Based on the logical analysis results, the corresponding target answer template is selected from the preset answer module library.

[0098] In this embodiment, the type of answer to be generated can be determined based on the obtained logical analysis results. For example, if the logical analysis concludes that the claim is denied, then templates related to the denial of claim are selected from the answer template library; if the result is that the claim is granted, then templates corresponding to the grant of claim are selected. During the template selection process, other factors can also be considered, such as user preferences and the tone of the inquiry, to select the most suitable answer template. For example, for more formal inquiries, a more rigorous and formal template can be selected; for more casual inquiries, a more concise and accessible template can be selected.

[0099] Furthermore, the construction process of the aforementioned answer module library includes collecting and organizing answer templates for various possible scenarios. In insurance claims consultation scenarios, templates might include phrases such as "According to clause ××, this accident is not covered by the claim" or "Based on the information you provided, your claim application meets the requirements, and the estimated compensation amount is ×× yuan." Each template is also clearly annotated, specifying the applicable scenario, the types of key information required, and the source of the information. For example, for the template "The estimated compensation amount is ×× yuan," the calculation method and basis for the amount need to be clearly stated.

[0100] Obtain the content information corresponding to the target answer template.

[0101] In this embodiment, specific content information matching the answer template can be extracted from various modules of the system. For example, in the template for denying a claim, specific clause information needs to be filled in, which can be obtained from the compliance detection module and the fact alignment module; in the template for affirming a claim, the expected compensation amount needs to be filled in, which can be obtained from the claims calculation module. Specifically, the compliance detection module stores information such as laws and regulations, industry standards, and company internal policies related to claims. In the case of denying a claim, specific clause information related to the current case is retrieved from this module. For example, by searching with keywords, specific clause numbers and original texts involving disclaimers and exclusions can be found. The fact alignment module records the comparison between key facts and standard facts. In the case of denying a claim, the specific scenario for which the clauses apply is further clarified based on the fact alignment results. For example, if the fact alignment shows that the accident was caused by the insured's intentional act, then clauses regarding exemptions for intentional acts are found from the compliance detection module, and relevant information is extracted. The claims calculation module, once an affirmative claim is determined, obtains the expected compensation amount and the data used to calculate that amount from this module. The claims calculation module typically performs calculations based on the rules stipulated in the insurance contract, combined with information such as loss proof and insured amount provided by the user. Key data involved in the calculation process, such as loss amount, deductible, and reimbursement ratio, is extracted and used to populate the template.

[0102] The target answer template is filled with information based on the content information to obtain the filled target answer template.

[0103] In this embodiment, the extracted content information can be matched with placeholders in the target answer template. Based on the matching results, the content information can be accurately filled into the selected target answer template to ensure the completeness and accuracy of the content information and to guarantee that the format of the filled information conforms to the template requirements. For example, the presentation of clause information can use a quotation format, such as "Article ×× of the ×× Insurance Clause: ...". Furthermore, attention should be paid to the expression of the content information to ensure consistency with the overall style of the answer template. If the template uses more formal written language, the filled content information should also use formal and rigorous expressions; if the answer template uses colloquial language intended for ordinary users, the filled content information should be as concise and clear as possible, avoiding the use of overly technical jargon.

[0104] The filled target answer template is used as the answer data.

[0105] Based on the above processing flow, this application integrates the consistency verification results, compliance detection results, and fact verification results to obtain corresponding integrated information; then, it performs logical analysis based on the integrated information to obtain corresponding logical analysis results; next, it selects the corresponding target answer template from a pre-set answer module library based on the logical analysis results; then, it obtains the content information corresponding to the target answer template; subsequently, it performs information filling processing on the target answer template based on the content information to obtain the filled target answer template; finally, it uses the filled target answer template as answer data. Thus, this application achieves intelligent and accurate generation of corresponding answer data by performing logical analysis on the integrated information containing consistency verification results, compliance detection results, and fact verification results, selecting the corresponding target answer template from the answer module library based on the logical analysis results, and then performing information filling processing on the target answer template based on the obtained content information. This effectively improves the accuracy and standardization of the generated answer data.

[0106] In some optional implementations of this embodiment, step S207 includes the following steps: Construct explanatory data corresponding to the answer data.

[0107] In this embodiment, the process of constructing the above-mentioned explanatory data includes: 1. Information extraction: Legal knowledge base retrieval: The system retrieves information from the legal knowledge base based on the key information involved in the standard answer. The legal knowledge base stores various laws, regulations, regulatory requirements, etc., in a structured manner and has established an index for quick retrieval. For example, if the standard answer involves relevant provisions of the Insurance Law, the system will search for specific clauses related to the standard answer in the knowledge base. Knowledge base association: At the same time, the system extracts information such as rejection reasons and compliance basis related to the standard answer from the general knowledge base. The knowledge base contains rich business knowledge, frequently asked questions, etc., and finds information related to the standard answer through semantic matching or keyword matching. 2. Information organization: The extracted legal clauses, rejection reasons, compliance basis, and other information are organized to form a clear explanatory chain, which serves as the above-mentioned explanatory data. For example, specific legal clauses are first cited to explain the exemption circumstances or claim conditions, then the correspondence between the user's situation and these clauses is explained, i.e., the rejection reasons, and finally the compliance of the judgment based on the law is explained.

[0108] The answer data is optimized based on a preset optimization strategy to obtain the corresponding target answer data.

[0109] In this embodiment, the specific implementation process of optimizing the answer data based on the preset optimization strategy to obtain the corresponding target answer data will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0110] Get the preset output format.

[0111] In this embodiment, the output format may include any one or more of the following: 1. For structured text: The standard answer and explanation chain are presented in a structured manner, such as by using paragraphs, bolding key content, or using lists, to make the information hierarchical and easy for users to read and understand. 2. Voice broadcast: For scenarios requiring voice output, the structured text is converted into speech, and the voice broadcast function is implemented through speech synthesis technology. 3. Highlighting key clauses and logical jump links: Key clauses are highlighted in the output text to facilitate users in quickly locating important information. At the same time, logical jump links are added to the relevant content in the key clauses and explanation chains. Users can click on the links to view the original text of the clauses or more detailed explanations, enhancing the interactivity and accessibility of the information.

[0112] The target answer data and the explanation data are output and processed based on the output format.

[0113] In this embodiment, the output processing of the target answer data and explanation data can be performed based on the output processing method corresponding to the selected output format.

[0114] Based on the above processing flow, this application constructs explanatory data corresponding to the answer data; then, it optimizes the answer data based on a preset optimization strategy to obtain the corresponding target answer data; subsequently, it obtains a preset output format; and then, based on the output format, it outputs the target answer data and explanatory data. Thus, by combining the optimization strategy and the output format, this application can present the generated processing results to the user in a clear, easy-to-understand, and traceable manner. The target answer data provides the user with a direct response to their inquiry, resolving their core problem; the traceable explanatory data details the basis and reasons for the answer, enhancing the user's trust in the response, thereby effectively improving the intelligence of the answer output and enhancing the user experience.

[0115] In some optional implementations of this embodiment, the optimization processing of the answer data based on a preset optimization strategy to obtain the corresponding target answer data includes the following steps: The answer data is optimized for content richness to obtain the corresponding first answer data.

[0116] In this embodiment, the above-mentioned content richness optimization includes: 1) Supplementing detailed explanations: In addition to providing basic conclusions, the generated answers should include appropriate detailed explanations. For example, when informing users that an accident is not covered by insurance, the specific reasons why the accident is not covered can be further explained, such as "According to clause ××, this accident is not covered by insurance because you did not have your vehicle inspected as required when the accident occurred, and the clause clearly stipulates that accidents involving vehicles that have not been inspected will not be covered." This allows users to understand the reasons more clearly, enhancing the credibility and persuasiveness of the answer. 2) Providing relevant suggestions: Based on the answer content, provide users with relevant suggestions or guidance. For example, in the case of refusing compensation, inform users of the follow-up measures they can take, such as "According to clause ××, this accident is not covered by insurance. You can first check your vehicle inspection records. If you have any questions, you can contact the relevant department for verification. It is also recommended that you have your vehicle inspected on time in the future to avoid similar situations affecting your claims." Such suggestions can help users solve problems better and improve user experience. 3) Introducing case references: Appropriately introduce similar cases as references in the answers to allow users to more intuitively understand the basis and rationality of the answers. For example, "According to clause ××, this accident is not covered by the insurance policy. A similar situation was also reflected in the previous case [Case No.], where the accident was not covered by the insurance policy due to [specific reasons], and the final outcome was consistent with this case." By referring to case studies, users can better understand the application of business rules and clauses, and enhance their sense of acceptance of the answers.

[0117] The first answer data is optimized for readability to obtain the corresponding second answer data.

[0118] In this embodiment, the readability optimization includes: 1) Optimizing language: Using concise, clear, and easy-to-understand language to generate answers, avoiding overly professional or obscure vocabulary and sentence structures. Complex business terms and legal clauses are explained and expressed in a simple and easy-to-understand way. For example, changing "According to the exclusion clause in the insurance contract, the insurer is not liable for compensation for accidents caused by the insured's intentional acts" to "According to the provisions of the insurance contract, if the accident is intentionally caused by you, the insurance company will not provide compensation." This makes it easier for users to understand the content of the answer. 2) Reasonable segmentation and layout: The generated answer is reasonably segmented and formatted to make the content hierarchical. Different information sections (such as conclusions, explanations, suggestions, etc.) are separated, using appropriate punctuation and paragraph spacing to enhance the readability of the answer. For example, the standard answer, explanation chain, and subsequent suggestions are divided into different paragraphs, allowing users to clearly distinguish the content of each part. 3) Adopting diversified expression: Avoiding the monotony and repetition of answer content, diversified expression methods are adopted. Different sentence structures, tones, and vocabulary can be used to enrich the content of the answer, making it more vivid and interesting. For example, when expressing a positive conclusion, you can use a statement like "Your claim application fully meets the requirements, and the expected compensation amount is ×× yuan. Congratulations!" This adds a positive emotional tone and enhances the user's reading experience.

[0119] The second answer data is semantically styled to obtain the corresponding third answer data.

[0120] In this embodiment, the aforementioned semantic style optimization includes adjusting the language style of the answer according to the user's language habits. If the user uses formal and rigorous language for consultation, the answer should also maintain a formal and rigorous style; if the user uses colloquial and casual language, the answer can appropriately adopt a more friendly and natural expression. For example, for formal consultation, one could use "According to Article ×× of the Insurance Law..."; for colloquial consultation, one could say "According to the provisions of the Insurance Law...".

[0121] The third answer data is optimized for scenario adaptation to obtain the corresponding fourth answer data.

[0122] In this embodiment, the above-mentioned scenario adaptation optimization includes: optimizing the answers based on different consultation scenarios. If the consultation scenario is an online customer service chat, the answers can be more concise and clear, avoiding overly long and complex sentences; if the consultation scenario is a written report, the answers can be more detailed and comprehensive, providing more background information and explanations. For example, in an online customer service chat, the answer to refusing compensation could be, "Sorry, according to clause ××, your situation is not covered by the claim"; while in a written report, the specific content of the clause, its relevance to the current case, and the basis for refusing compensation can be explained in detail.

[0123] The fourth answer data is used as the target answer data.

[0124] Based on the above processing flow, this application optimizes the answer data for content richness to obtain the corresponding first answer data; then optimizes the first answer data for readability to obtain the corresponding second answer data; subsequently, it optimizes the second answer data for semantic style to obtain the corresponding third answer data; next, it optimizes the third answer data for scenario adaptation to obtain the corresponding fourth answer data; finally, it uses the fourth answer data as the target answer data. Thus, by optimizing the answer data for content richness, readability, semantic style, and scenario adaptation, this application can intelligently achieve multi-dimensional optimization of the answer data, effectively improving the accuracy and quality of the generated target answer data.

[0125] In some alternative implementations, the user information obtained is subject to user consent and complies with relevant laws and policies.

[0126] Furthermore, any software tools or components not belonging to our company that appear in the embodiments of this application are merely illustrative examples and do not represent actual use.

[0127] Furthermore, this application introduces a "consistency training" self-supervised paradigm into the financial insurance intelligent system, achieving an inherent autonomous compliance capability that resists manipulation, flattery, and jailbreak. Through a dual consistency constraint mechanism of external output + internal representation, it establishes an interpretable, supervisory, and sustainably evolving intelligent compliance decision-making system, demonstrating significant technological innovation and commercial application value. The technical advantages and application value are as follows: 1. Anti-manipulation and anti-jailbreak capabilities: Through consistent training, the model remains stable when faced with fake instructions and guiding statements, preventing the generation of illegal commands.

[0128] 2. Internal and external dual-layer consistency constraints: taking into account the robustness of both the output layer and the internal activation layer, improving the consistency of decision paths and the robustness of the model.

[0129] 3. Alignment of facts and regulations: By combining with the RAG knowledge base, we ensure that the model output is fully compliant with laws and regulations.

[0130] 4. Unsupervised adaptive update: The system continuously collects cue variant samples to form a self-evolving consistent learning flywheel.

[0131] 5. Wide applicability: It can be used in scenarios such as insurance product Q&A, claims underwriting, customer consultation, and compliance testing, significantly reducing the risks of misjudgment, non-compliant responses, and public opinion.

[0132] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0133] It should be emphasized that, to further ensure the privacy and security of the above answer data, the answer data can also be stored in a node of a blockchain.

[0134] The blockchain referred to in this application is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0135] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0136] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0137] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0138] Further reference Figure 3 As a response to the above Figure 2 To implement the method shown, this application provides an embodiment of an artificial intelligence-based problem-solving device, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0139] like Figure 3 As shown, the AI-based problem-solving device 300 described in this embodiment includes: a receiving module 301, a first generation module 302, a combination module 303, a processing module 304, a verification module 305, a second generation module 306, and an output module 307. Wherein: The receiving module 301 is used to receive user input question data; The first generation module 302 is used to perform data generation processing on the problem data based on a preset enhancement generation strategy to obtain corresponding enhanced problem data; The combination module 303 is used to combine the problem data and the enhanced problem data to obtain corresponding combined data; The processing module 304 is used to perform inference processing on the combined data based on a preset target large language model to obtain an output result, and to perform consistency verification on the output result to obtain a corresponding consistency verification result. The verification module 305 is used to perform compliance detection and fact alignment verification on the output results to obtain the corresponding compliance detection results and fact verification results. The second generation module 306 is used to perform answer generation processing based on the consistency verification result, the compliance detection result and the fact verification result to obtain the corresponding answer data; The output module 307 is used to process the output of the answer data.

[0140] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based problem-solving method in the aforementioned implementation method, and will not be repeated here.

[0141] In some optional implementations of this embodiment, the first generation module 302 includes: The first processing submodule is used to perform text perturbation enhancement processing on the problem data to obtain the corresponding first problem data; The second processing submodule is used to perform context wrapping enhancement processing on the first problem data to obtain the corresponding second problem data; The third processing submodule is used to perform cross-style restatement processing on the second problem data to obtain the corresponding third problem data; The first determining submodule is used to use the third problem data as the enhanced problem data.

[0142] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based problem-solving method in the aforementioned implementation method, and will not be repeated here.

[0143] In some optional implementations of this embodiment, the processing module 304 includes: The inference submodule is used to call a pre-trained target large language model to perform parallel inference processing on the combined data and obtain the corresponding output results. The first verification submodule is used to perform external output consistency verification on the output result and obtain the corresponding first verification result. The second verification submodule is used to perform internal representation consistency verification on the output result to obtain the corresponding second verification result. The first integration submodule is used to integrate the first verification result and the second verification result to obtain the corresponding target verification result; The second determining submodule is used to use the target verification result as the consistency verification result.

[0144] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based problem-solving method in the aforementioned implementation method, and will not be repeated here.

[0145] In some optional implementations of this embodiment, the verification module 305 includes: Calling submodules is used to invoke the preset legal knowledge base; The query submodule is used to retrieve the corresponding regulatory information from the regulatory knowledge base based on the question data and the output results. The detection submodule is used to perform compliance detection on the output results and the regulatory information based on a preset regression classifier, and obtain the corresponding compliance detection results; An extraction submodule is used to extract key factual information from the output results; The third verification submodule is used to perform fact alignment verification on the key fact information and obtain the corresponding fact verification result.

[0146] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based problem-solving method in the aforementioned implementation method, and will not be repeated here.

[0147] In some optional implementations of this embodiment, the second generation module 306 includes: The second integration submodule is used to integrate the consistency verification result, the compliance detection result, and the fact verification result to obtain the corresponding integrated information. The analysis submodule is used to perform logical analysis based on the integrated information and obtain the corresponding logical analysis results. The filtering submodule is used to filter out the corresponding target answer template from the preset answer module library based on the logical analysis results; The first acquisition submodule is used to acquire content information corresponding to the target answer template; The fill submodule is used to fill the target answer template with information based on the content information to obtain the filled target answer template. The third determining submodule is used to use the filled target answer template as the answer data.

[0148] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based problem-solving method in the aforementioned implementation method, and will not be repeated here. In some optional implementations of this embodiment, the output module 307 includes: A submodule is constructed to build explanation data corresponding to the answer data; The optimization submodule is used to optimize the answer data based on a preset optimization strategy to obtain the corresponding target answer data; The second acquisition submodule is used to acquire the preset output format; The output submodule is used to process the target answer data and the explanation data based on the output format.

[0149] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based problem-solving method in the aforementioned implementation method, and will not be repeated here.

[0150] In some optional implementations of this embodiment, the optimized submodule includes: The first optimization unit is used to optimize the content richness of the answer data to obtain the corresponding first answer data; The second optimization unit is used to optimize the readability of the first answer data to obtain the corresponding second answer data. The third optimization unit is used to perform semantic style optimization on the second answer data to obtain the corresponding third answer data; The fourth optimization unit is used to perform scenario adaptation optimization on the third answer data to obtain the corresponding fourth answer data; A determining unit is used to use the fourth answer data as the target answer data.

[0151] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based problem-solving method in the aforementioned implementation method, and will not be repeated here. To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0152] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected via a system bus. It should be noted that only the computer device 4 with components 41-43 is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0153] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0154] The memory 41 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 41 may be an internal storage unit of the computer device 4, such as the hard disk or memory of the computer device 4. In other embodiments, the memory 41 may also be an external storage device of the computer device 4, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 4. Of course, the memory 41 may also include both the internal storage unit and its external storage device of the computer device 4. In this embodiment, the memory 41 is typically used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions for problem-solving methods based on artificial intelligence. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or will be output.

[0155] In some embodiments, the processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 42 is typically used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to execute computer-readable instructions stored in the memory 41 or to process data, for example, to execute computer-readable instructions of the artificial intelligence-based problem-solving method.

[0156] The network interface 43 may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the computer device 4 and other electronic devices.

[0157] This application also provides another embodiment, namely, providing a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor to cause the at least one processor to perform the steps of the artificial intelligence-based problem-solving method described above.

[0158] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0159] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A problem-solving method based on artificial intelligence, characterized in that, Includes the following steps: Receive user input for questions; The problem data is processed based on a preset enhancement generation strategy to obtain corresponding enhanced problem data; The problem data and the enhanced problem data are combined to obtain the corresponding combined data; The combined data is inferred based on the preset target large language model to obtain the output result, and the output result is subjected to consistency verification to obtain the corresponding consistency verification result. The output results are subjected to compliance testing and fact alignment verification to obtain the corresponding compliance testing results and fact verification results. Based on the consistency verification result, the compliance detection result, and the fact verification result, answer generation processing is performed to obtain the corresponding answer data; The answer data is then processed for output.

2. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The step of generating and processing the problem data based on a preset enhancement generation strategy to obtain corresponding enhanced problem data specifically includes: The problem data is subjected to text perturbation enhancement processing to obtain the corresponding first problem data; The first problem data is subjected to context wrapping enhancement processing to obtain the corresponding second problem data; The data for the second problem is subjected to cross-style restatement processing to obtain the corresponding data for the third problem; The third problem data is used as the enhanced problem data.

3. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The steps of performing inference processing on the combined data based on a preset target large language model to obtain an output result, and performing consistency verification on the output result to obtain a corresponding consistency verification result, specifically include: The pre-trained target large language model is invoked to perform parallel inference processing on the combined data to obtain the corresponding output results; Perform an external output consistency check on the output result to obtain the corresponding first check result; The output result is subjected to an internal representation consistency check to obtain the corresponding second check result. The first verification result and the second verification result are integrated to obtain the corresponding target verification result; The target verification result is used as the consistency verification result.

4. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The step of generating answer data based on the consistency verification result, the compliance detection result, and the fact verification result specifically includes: Call the preset legal knowledge base; Based on the problem data and the output results, the corresponding regulatory information is retrieved from the regulatory knowledge base; Based on a preset regression classifier, the output results and the regulatory information are subjected to compliance detection to obtain the corresponding compliance detection results; Extract key factual information from the output results; The key factual information is subjected to fact alignment verification to obtain the corresponding fact verification result.

5. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The step of generating answer data based on the consistency verification result, the compliance detection result, and the fact verification result specifically includes: The consistency verification result, the compliance detection result, and the fact verification result are integrated to obtain the corresponding integrated information. Logical analysis is performed based on the integrated information to obtain the corresponding logical analysis results; Based on the results of the logical analysis, the corresponding target answer template is selected from the preset answer module library; Obtain the content information corresponding to the target answer template; Based on the content information, the target answer template is filled with information to obtain the filled target answer template; The filled target answer template is used as the answer data.

6. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The step of outputting the answer data specifically includes: Construct explanatory data corresponding to the answer data; The answer data is optimized based on a preset optimization strategy to obtain the corresponding target answer data; Obtain the preset output format; The target answer data and the explanation data are output and processed based on the output format.

7. The problem-solving method based on artificial intelligence according to claim 6, characterized in that, The step of optimizing the answer data based on a preset optimization strategy to obtain the corresponding target answer data specifically includes: The answer data is optimized for content richness to obtain the corresponding first answer data; The first answer data is optimized for readability to obtain the corresponding second answer data; The second answer data is semantically style optimized to obtain the corresponding third answer data; The third answer data is optimized for scenario adaptation to obtain the corresponding fourth answer data; The fourth answer data is used as the target answer data.

8. A problem-solving device based on artificial intelligence, characterized in that, include: The receiving module is used to receive user-input question data; The first generation module is used to perform data generation processing on the problem data based on a preset enhancement generation strategy to obtain corresponding enhanced problem data; The combination module is used to combine the problem data and the enhanced problem data to obtain corresponding combined data; The processing module is used to perform inference processing on the combined data based on a preset target large language model to obtain the output result, and to perform consistency verification on the output result to obtain the corresponding consistency verification result. The verification module is used to perform compliance detection and fact alignment verification on the output results to obtain the corresponding compliance detection results and fact verification results. The second generation module is used to perform answer generation processing based on the consistency verification result, the compliance detection result and the fact verification result to obtain the corresponding answer data; The output module is used to process the output of the answer data.

9. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the problem-solving method based on artificial intelligence as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the problem-solving method based on artificial intelligence as described in any one of claims 1 to 7.

Images