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, which comprises the following steps of: analyzing input question data to obtain a triple; performing interference evaluation on the candidate options based on the triple to screen low-confidence options, and deleting the low-confidence options from the candidate options to obtain a first option; performing anti-fact verification on the first options to screen wrong options, and deleting the wrong options from the first options to obtain second options; if the number of the second options is multiple, performing multi-path reasoning on each second option to obtain a path reasoning result; performing confidence evaluation on each path reasoning result to obtain multiple pieces of confidence data; performing numerical value comparison on the confidence data to obtain a target option corresponding to the target confidence data with the highest numerical value; and generating answer data based on the target option and outputting the answer data. The method can be applied to question processing scenes in the field of financial science and technology, and the answer accuracy of question processing is 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] In the financial and insurance sector, the deep application of artificial intelligence (AI) technology has become a key force driving the industry's digital transformation and improving service efficiency and quality. Currently, AI has been widely integrated into multiple core business scenarios such as risk assessment, claims processing, product recommendation, and intelligent customer service, bringing significant efficiency improvements and business innovation to financial institutions and insurance companies.

[0003] Most existing intelligent systems in the financial and insurance sector rely on end-to-end large language models (LLMs) to achieve automatic question answering and decision support functions. These models, with their powerful language understanding and generation capabilities, can quickly process large amounts of text information, providing users with seemingly reasonable answers and decision suggestions, thus meeting basic business needs to a certain extent. However, in practical applications, end-to-end large language models reveal many limitations when handling specific complex problems, severely affecting the accuracy of answer data and thus hindering their further promotion and application in the financial and insurance field. Specifically, when faced with multiple-choice decisions or problems with distracting options, these models are highly susceptible to the drawbacks of "cognitive interference" and "misjudgment reasoning."

[0004] Taking insurance claims review as an example, insurance terms are typically highly specialized and complex, with different terms potentially having similar meanings but significant logical differences. When a model needs to accurately select the one that best suits the customer's situation from multiple similar terms, traditional large language models often struggle to accurately grasp the logical relationships between the terms, easily being misled by semantic similarity and thus selecting the logically incorrect but semantically closest option. Such misjudgment can not only lead to unfair claims results and customer disputes but also cause unnecessary economic losses and reputational risks for the insurance company.

[0005] In summary, existing intelligent financial and insurance systems that rely on end-to-end large language models suffer from low accuracy when dealing with complex problems, failing to meet the industry's demand for high-precision decision support. Therefore, developing a novel intelligent decision support system that can effectively overcome these problems and improve the accuracy of answer data has significant practical implications and broad application prospects. Summary of the Invention

[0006] The purpose of this application is to propose a problem-solving method, apparatus, computer device, and storage medium based on artificial intelligence, in order to solve the technical problem of low answer accuracy in existing problem-solving methods that rely on end-to-end large language models.

[0007] Firstly, an artificial intelligence-based problem-solving method is provided, including: Receive user input for questions; The problem data is parsed based on a preset semantic decomposition model to obtain corresponding triples; wherein, the triples include a target variable, a constraint set, and a candidate set, and the candidate set includes multiple candidate options; Based on the triplet, a preset interference evaluation network is used to evaluate the interference of the candidate option pairs to filter out low-confidence options, and the low-confidence options are removed from the candidate options to obtain the first option; Based on a pre-defined large language model, the first option is counterfactually verified to filter out incorrect options, and the incorrect options are then removed from the first option to obtain the second option. If there are multiple second options, then multi-path reasoning is performed on each second option based on a preset generation framework to obtain path reasoning results corresponding to each second option. The confidence scores of each path reasoning result are evaluated to obtain multiple corresponding confidence scores. The target confidence data with the highest value is determined by numerical comparison of all the confidence data, and the target option corresponding to the target confidence data is obtained. Generate corresponding answer data based on the target options, and output the answer data.

[0008] Secondly, an artificial intelligence-based problem-solving device is provided, comprising: The receiving module is used to receive user-input question data; The parsing module is used to parse the problem data based on a preset semantic decomposition model to obtain the corresponding triples; wherein, the triples include a target variable, a constraint set, and a candidate set, and the candidate set includes multiple candidate options; The first processing module is used to perform interference evaluation on the candidate option pair based on the triplet using a preset interference evaluation network to filter out low-confidence options, and delete the low-confidence options from the candidate options to obtain the first option; The second processing module is used to perform counterfactual verification on the first option based on a preset large language model to filter out incorrect options, and delete the incorrect options from the first option to obtain the second option. The reasoning module is used to perform multi-path reasoning processing on each of the second options based on a preset generation framework if there are multiple second options, so as to obtain the path reasoning results corresponding to each of the second options respectively. The evaluation module is used to evaluate the confidence of each path reasoning result to obtain multiple confidence data. The third processing module is used to perform numerical comparisons on all the confidence data to determine the target confidence data with the highest value, and to obtain the target option corresponding to the target confidence data; The output module is used to generate corresponding answer data based on the target options and output the answer data.

[0009] 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.

[0010] 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.

[0011] In the aforementioned 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 parsed based on a preset semantic decomposition model to obtain corresponding triples; wherein, the triples include a target variable, a constraint set, and a candidate set, and the candidate set includes multiple candidate options; then, based on the triples, a preset interference evaluation network is used to evaluate the interference of the candidate option pairs to filter out low-confidence options, and the low-confidence options are deleted from the candidate options to obtain a first option; subsequently, a counterfactual evaluation is performed on the first option based on a preset large language model. The process involves several steps: first, verifying and filtering out incorrect options; then, removing these incorrect options from the first option to obtain a second option; second, if there are multiple second options, performing multi-path reasoning on each second option based on a preset generation framework to obtain path reasoning results corresponding to each second option; third, evaluating the confidence of each path reasoning result to obtain multiple confidence data points; fourth, comparing the values ​​of all confidence data to determine the target confidence data with the highest value, and obtaining the target option corresponding to the target confidence data; and finally, generating corresponding answer data based on the target option and outputting the answer data. Thus, this problem-solving approach, by employing a thought process of eliminating incorrect options and gradually converging to the correct conclusion, can significantly reduce the cognitive interference and misjudgment rate of large language models. By using a reasoning mechanism combining interference evaluation, counterfactual verification, and multi-path verification, large language models can identify and eliminate potential interference and incorrect options when facing multiple-choice questions, ultimately generating accurate decision results based on confidence data, effectively improving the accuracy of problem-solving answers. Attached Figure Description

[0012] 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.

[0013] 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

[0014] 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.

[0015] 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.

[0016] 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.

[0017] 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.

[0018] 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.

[0019] 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.

[0020] 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.

[0021] 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.

[0022] 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.

[0023] 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.

[0024] 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 wireless connection methods. The implementing entity of this application is specifically a question processing system, which can be simply referred to as the system. Users can manually input the required question data through a specific interactive interface (such as a web form, mobile application interface, etc.). The system can provide diverse input methods according to different application scenarios and user needs. This application can be applied to question processing scenarios in the fintech field. In insurance claims scenarios, users can input questions described in natural language through text boxes, such as "If a customer's illness occurred 60 days after the policy took effect, is a claim payable?"; preset options can also be provided for users to choose from to simplify the input process, such as allowing users to select key information such as the time of the customer's illness occurrence and policy status, and the system will combine these selections into a complete question.

[0025] Step S202: The problem data is parsed based on a preset semantic decomposition model to obtain the corresponding triples; wherein, the triples include the target variable, the constraint set, and the candidate set, and the candidate set includes multiple candidate options.

[0026] In this embodiment, the candidate set contains multiple possible options (i.e., candidate options), and these candidate options need to undergo multiple subsequent evaluations to determine their rationality.

[0027] The specific implementation process of parsing the problem data based on the preset semantic decomposition model to obtain the corresponding triples will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0028] Step S203: Based on the triplet, use a preset interference evaluation network to evaluate the interference of the candidate option pair to filter out low-confidence options, and delete the low-confidence options from the candidate options to obtain the first option.

[0029] In this embodiment, the specific implementation process of using a preset interference evaluation network to evaluate the interference of the candidate option pairs based on the triples to filter out low-confidence options will be described in more detail in subsequent specific embodiments of this application, and will not be elaborated on here. Furthermore, after filtering out low-confidence options, i.e., interference options, from the candidate options, the low-confidence options are removed from the candidate options to obtain the first option. This allows the subsequent decision-making process to focus more on the truly valuable first option, thereby improving the accuracy and reliability of the decision-making process.

[0030] Step S204: Perform counterfactual verification on the first option based on a preset large language model to filter out incorrect options, and delete the incorrect options from the first option to obtain the second option.

[0031] In this embodiment, the specific implementation process of performing counterfactual verification on the first option based on a preset large language model to filter out incorrect options will be described in further detail in subsequent embodiments of this application, and will not be elaborated on here. Furthermore, after filtering out incorrect options from the first option, the incorrect options are removed from the first option to obtain the second option, making the subsequent decision-making process more focused on the truly valuable second option, thereby improving the accuracy and reliability of the decision processing. In addition, if there is only one second option, it is directly used as the final answer data and output to the user, thus completing the decision processing of the user-input question data.

[0032] Step S205: If there are multiple second options, then perform multi-path reasoning processing on each second option based on a preset generation framework to obtain path reasoning results corresponding to each second option.

[0033] In this embodiment, the above-mentioned generation framework can specifically adopt the improved Chain-of-Thought (CoT) generation framework.

[0034] The initialization of the generation framework includes: 1. Framework parameter configuration. Improvement point setting: Based on the specific business scenario and problem complexity, determine the improvement direction of the improved Chain-of-Thought (CoT) framework. For example, in an insurance claims scenario, it may be necessary to strengthen the framework's ability to parse complex insurance clauses and its ability to comprehensively utilize customer information. Inference step size and depth setting: Set the maximum step size for each inference step, i.e., how many small steps each inference can be broken down into; at the same time, determine the maximum depth of inference to avoid the inference process becoming too lengthy or falling into an infinite loop. 2. Knowledge base and rule base integration. Knowledge base preparation: Organize and store business-related knowledge, such as basic concepts of the insurance industry, common disease types, claims processes, etc., in the knowledge base. Ensure that the information in the knowledge base is accurate, comprehensive, and easy to retrieve. Rule base construction: Collect business rules, such as compensation conditions, deductible period provisions, exclusions, etc. in insurance clauses, to form a rule base. Set clear conditions and conclusions for each rule so that they can be matched and applied during the inference process. 3. Input data preprocessing. Data cleaning: Cleaning the remaining options and related contextual information after counterfactual elimination, removing irrelevant characters, correcting errors, and standardizing data format. Feature extraction: Extracting key features from the options and contextual information, such as the clause numbers involved in the options and basic customer information (age, gender, disease type, etc.), and using these features as the initial input for multi-path reasoning.

[0035] In addition, the implementation process of generating path reasoning results through multi-path parallel reasoning includes: 1. Path Planning. Path Generation Based on Options: For each remaining option (second option), plan multiple different reasoning paths based on its characteristics and business logic. For example, for a clause option in an insurance claim, a direct reasoning path based on the literal meaning of the clause can be planned, as well as a comprehensive reasoning path combining the customer's actual medical history and claims history. Consider Different Information Sources: Ensure that each reasoning path draws its basis from different information sources. For example, some paths mainly rely on the insurance clause, while others focus on the customer's historical claims records or medical reports. 2. Execution of the Reasoning Process. Step-by-Step Reasoning: Proceed step-by-step inference according to the planned path. In each reasoning step, generate the next reasoning direction and basis based on the input of the current step and the knowledge and rules already acquired. Dynamic Adjustment: During the reasoning process, dynamically adjust the reasoning path based on new information and reasoning results. For example, if information contradicting the initial assumption is found in a certain path, adjust subsequent reasoning steps in a timely manner or switch to other possible paths. 3. Path Convergence Judgment. Setting Convergence Conditions: Determine when the reasoning path can converge, that is, reach the final reasoning conclusion. Common convergence conditions include: clear and consistent inference results, reaching the maximum inference depth, and the inability to obtain any new valid information. Path merging and filtering: When multiple inference paths meet the convergence conditions, their results are analyzed and compared. If the inference results of multiple paths are consistent, the result is taken as the final inference result of that option (path inference result); if the results differ, they are filtered and merged based on factors such as path credibility and logical consistency.

[0036] In addition, each reasoning path retains logical nodes (thinking chain nodes), recorded in JSON structure. The specific implementation process includes: 1. Node definition and recording. Node content determination: In each reasoning step, the content of the logical nodes to be recorded is determined, including the terms on which the reasoning is based, customer information, intermediate reasoning results, etc. For example, when determining whether a customer meets the compensation conditions of a certain term, the specific content of the term, the customer's disease type, and medical history are recorded. JSON structure recording: Each logical node is recorded according to a predefined JSON structure to ensure the integrity and readability of the information. For example, the JSON structure of a logical node may contain fields such as "step name," "based information," and "reasoning result."

[0037] Step S206: Calculate the confidence level of each path reasoning result to obtain multiple confidence level data.

[0038] In this embodiment, the specific implementation process of evaluating the confidence of each path reasoning result to obtain the corresponding multiple confidence data will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0039] Step S207: Compare the values ​​of all the confidence data to determine the target confidence data with the highest value, and obtain the target option corresponding to the target confidence data.

[0040] In this embodiment, the target confidence level data with the highest value can be determined by integrating and comparing all the calculated confidence level data. The target option is the one from the second option that corresponds to the target confidence level data and is most likely to be correct.

[0041] Step S208: Generate corresponding answer data based on the target options, and output the answer data.

[0042] In this embodiment, the aforementioned target option can be used as the final answer, i.e., the answer data. This answer data is the most likely correct target option determined based on confidence data. Subsequently, the output processing of the answer data can be completed by sending the answer data and its explanation together to the user. The explanation of the answer data is a brief description of the reasoning process and confidence assessment, such as explaining why the answer data was selected and its confidence level.

[0043] Based on the above automated processing flow, this application first receives user-input question data; and parses the question data based on a preset semantic decomposition model to obtain corresponding triples; wherein, the triples include a target variable, a constraint set, and a candidate set, and the candidate set includes multiple candidate options; then, based on the triples, a preset interference evaluation network is used to evaluate the interference of the candidate option pairs to filter out low-confidence options, and the low-confidence options are deleted from the candidate options to obtain a first option; then, based on a preset large language model, the first option is counterfactually verified to filter out incorrect options, and the incorrect options are deleted from the first option to obtain a second option; if there are multiple second options, multi-path reasoning processing is performed on each second option based on a preset generation framework to obtain path reasoning results corresponding to each second option; subsequently, confidence evaluation is performed on each path reasoning result to obtain multiple corresponding confidence data; further, all the confidence data are numerically compared to determine the target confidence data with the highest value, and the target option corresponding to the target confidence data is obtained; finally, the corresponding answer data is generated based on the target option, and the answer data is output. Thus, the problem-solving approach of this application, by adopting a thinking mode of eliminating incorrect options and gradually converging to the correct conclusion, can significantly reduce the cognitive interference and misjudgment rate of the large language model. By employing a reasoning mechanism that combines interference assessment, counterfactual verification, and multi-path verification, the large language model can identify and eliminate potential interference and incorrect options when faced with multiple-choice questions, and finally generate accurate decision results based on confidence data, effectively improving the accuracy of the answer to the problem.

[0044] In some alternative implementations, step S202 includes the following steps: The problem data is preprocessed to obtain the corresponding processed data.

[0045] In this embodiment, the preprocessing includes: preliminary cleaning and standardization of the input problem data. Irrelevant characters, spaces, special symbols, etc., are removed from the problem data, and common expressions in the problem data are standardized, for example, "60 days after policy effective" is uniformly represented as "sixty days after policy effective," thereby obtaining corresponding processed data for better processing by subsequent modules.

[0046] Based on the semantic decomposition model, semantic analysis is performed on the processed data to obtain the corresponding semantic analysis results.

[0047] In this embodiment, the semantic decomposition model can specifically be a Transformer-based semantic decomposition model. This model has been trained on a large amount of semantic data and is capable of understanding the semantic structure of natural language problems. A pre-trained semantic decomposition model can be loaded into the system to ensure that the semantic decomposition model can operate normally.

[0048] The pre-processed input problem (processing data) can be input into the semantic decomposition model, which will then perform semantic analysis on the processed data, identify key information in the problem, and output the corresponding semantic analysis results.

[0049] The problem data is decomposed based on the semantic analysis results to obtain the corresponding decomposition results.

[0050] In this embodiment, based on the semantic analysis results, the input question data is decomposed into a triple of {target variable, constraint set, candidate set}. For example, for the question "If a customer's illness occurs 60 days after the policy takes effect, is a claim payable?", the target variable might be "whether a claim is payable", the constraint set might be "the customer's illness occurs 60 days after the policy takes effect", and the candidate set might be various situations related to claims in the insurance terms (although the candidate set may not be very obvious in this simple question, there will be clear candidate options in more complex questions).

[0051] The decomposition result is taken as the triplet.

[0052] In this embodiment, the generated triplet contains a candidate set containing multiple possible options (i.e., candidate options), which need to undergo multiple subsequent evaluations to determine their rationality.

[0053] Based on the above processing flow, this application preprocesses the problem data to obtain corresponding processed data; then, it performs semantic analysis on the processed data based on a semantic decomposition model to obtain corresponding semantic analysis results; subsequently, it decomposes the problem data based on the semantic analysis results to obtain corresponding decomposition results; and the decomposition results are then presented as triples. In this way, by using a semantic decomposition model to perform structured parsing of the input problem data, decomposing the problem data into triples, this application can clearly define the core objective, constraints, and possible options of the problem. This allows subsequent processes such as cognitive interference identification and counterfactual elimination to address the problem more effectively, thereby improving the efficiency and accuracy of decision-making.

[0054] In some optional implementations of this embodiment, step S203 includes the following steps: Invoke the preset interference assessment network.

[0055] In this embodiment, the aforementioned interference evaluation network is a pre-constructed interference evaluation network based on a bidirectional contrastive Transformer. This interference evaluation network can analyze the relationship between the options and the problem objective (i.e., the objective vector) from both positive and negative directions, thus more comprehensively evaluating the potential interference of the options.

[0056] For the first specified option, the interference evaluation network is used to calculate the semantic similarity and logical conflict between the first specified option and the target variable; wherein, the first specified option is any one of all the candidate options.

[0057] In this embodiment, the semantic vector representations of the first specified option and the question target (i.e., the target vector) can be calculated using an interference evaluation network. Then, the semantic similarity between these two semantic vector representations is obtained by calculating indices such as cosine similarity and Euclidean distance. The semantic similarity value generally ranges from 0 to 1; a higher value indicates a higher semantic similarity, meaning the option and the question target are semantically closer. If the semantic similarity between the option and the question target is high, it indicates that the option is semantically relevant to the question and is likely a reasonable option; conversely, a lower similarity suggests a lower relevance.

[0058] Furthermore, an interference evaluation network is used to determine whether the first specified option violates the logical relationships involved in the problem objective, and the corresponding logical conflict level is output. The logical conflict level typically ranges from 0 to 1; a higher value indicates a higher logical conflict level, meaning a greater logical contradiction between the option and the problem objective; a lower value indicates a lower logical conflict level, meaning a more logically consistent option and problem objective. Options with high logical conflict levels often interfere with correct decision-making and require close attention and screening.

[0059] The semantic similarity and logical conflict degree are calculated based on a preset scoring formula to obtain the corresponding score data.

[0060] In this embodiment, the scoring formula is specifically as follows: Where sim is semantic similarity, used to measure the similarity between options. With target variable Semantic similarity; conf represents logical conflict, used to measure the degree of agreement between options. With target variable Is there a logical contradiction? α is the adaptive weight of semantic similarity, used to adjust the importance of semantic similarity in the calculation of interference score. Different application scenarios and data characteristics have different requirements for semantic similarity. By adjusting the value of α, the influence of semantic similarity and logical conflict on interference score can be balanced. β is used to adjust the importance of logical conflict in the calculation of interference score. The value of β also needs to be adjusted according to the specific application scenario and data characteristics.

[0061] In addition, the selection of the adaptive weights for the semantic similarity can be determined according to actual needs. The corresponding adjustment criteria include: in some scenarios with high requirements for accurate semantic matching (such as text classification, information retrieval, etc.), the value of α can be appropriately increased to make semantic similarity have a greater impact on the interference score; while in some scenarios with more complex logical relationships (such as legal reasoning, medical diagnosis, etc.), it may be necessary to decrease the value of α, reduce the weight of semantic similarity, and pay more attention to the evaluation of logical conflict.

[0062] The selection of the adaptive weight for the above logical conflict degree can be determined according to actual needs. The corresponding adjustment criteria include: when logical relationship plays a key role in the decision-making process (such as in some fields that require strict adherence to logical rules), the value of β should be increased to make the logical conflict degree have a greater impact on the interference degree score; while in scenarios where semantic similarity has a greater impact on decision-making, the value of β can be appropriately reduced.

[0063] Furthermore, the corresponding score data can be obtained by substituting the semantic similarity, logical conflict degree, and adaptive weight into the above score calculation formula.

[0064] Determine whether the score data is less than a preset score threshold.

[0065] In this embodiment, a score threshold is preset according to actual needs, and options with interference scores lower than the score threshold are excluded, thereby reducing the number of options to be processed in the subsequent process and improving decision processing efficiency.

[0066] If so, the first specified option is determined as a low-confidence option.

[0067] In this embodiment, if the first specified selected score data is detected to be less than a preset score threshold, then the first specified option is determined as a low confidence option.

[0068] This application utilizes a pre-defined interference evaluation network. Then, for a first specified option, the interference evaluation network calculates the semantic similarity and logical conflict degree between the specified option and the target variable. The first specified option is any one of all candidate options. A pre-defined scoring formula is then used to calculate the semantic similarity and logical conflict degree to obtain corresponding score data. The score data is then checked against a pre-defined scoring threshold. If the score is less than a pre-defined threshold, the first specified option is identified as a low-confidence option. In this way, since some options may seem relevant to the problem in actual decision-making but actually contain semantic misleading or logical conflicts, this application, through calculation and filtering based on the interference evaluation network, can identify and eliminate these low-confidence interference options in advance. This helps the subsequent decision-making process focus more on truly valuable options, thereby improving the accuracy and reliability of decision processing.

[0069] In some alternative implementations, step S204 includes the following steps: Generate a hypothesis statement corresponding to the second specified option; wherein the second specified option is any one of all the first options; the hypothesis statement contains the reasoning result.

[0070] In this embodiment, by using a large model, the process of "hypothetical reasoning" can be imitated to generate a hypothetical statement containing the reasoning result for each remaining first option. For example, for option "Clause B", the hypothetical statement "If option Clause B is selected, the reasoning result is that it is claimable" is generated. The selection of the large model is not specifically limited; it can be selected from general large models based on actual business needs.

[0071] Based on known facts, the inference results are verified using the large language model.

[0072] In this embodiment, the aforementioned fact verification refers to verifying whether the reasoning result conflicts with known facts or context through a large language model. The large language model is a pre-built model with powerful language understanding and reasoning capabilities, capable of determining the reasonableness of the reasoning result based on known insurance terms, customer information, and other contextual content. For example, if it is known that the customer's illness is within the deductible period, and clause B stipulates that claims are not made within the deductible period, then the reasoning result "claimable" conflicts with the known fact. Here, fact verification is a qualitative judgment process, that is, determining whether the reasoning result contains a contradiction.

[0073] If the reasoning result is contradictory, then the strength of the conflict between the reasoning result and the known facts is calculated.

[0074] In this embodiment, based on the factual verification confirming the existence of a conflict, the conflict intensity calculation further quantifies the degree of this conflict. It not only identifies that the reasoning result contradicts known facts, but also clarifies the severity of the conflict, thus providing a more precise basis for subsequent option elimination.

[0075] The specific implementation process for calculating the conflict intensity between the reasoning result and the known facts will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0076] Determine whether the conflict intensity is greater than a preset conflict threshold.

[0077] In this embodiment, the selection of the aforementioned conflict threshold is not specifically limited and can be set according to actual business needs. If the conflict intensity of an option is greater than the conflict threshold, it indicates that the option contradicts the known facts more severely and should be excluded.

[0078] If so, the second specified option will be identified as an incorrect option.

[0079] In this embodiment, if the conflict intensity of the second designated option is detected to be greater than a preset conflict threshold, the second designated option is determined to be an incorrect option. Additionally, an "exclusion reason" text can be generated for the obtained incorrect options, such as "Option B conflicts with the deductible period rule, therefore excluded." These exclusion reasons clearly explain why a certain option is excluded, enhancing the interpretability of the decision and facilitating manual review and understanding of the decision-making process.

[0080] Based on the above processing flow, this application generates a hypothetical statement corresponding to a second specified option; wherein the second specified option is any one of all the first options; the hypothetical statement contains the reasoning result; then, based on known facts, a large language model is used to verify the reasoning result; if the reasoning result is contradictory, the conflict intensity between the reasoning result and the known facts is calculated; subsequently, it is determined whether the conflict intensity is greater than a preset conflict threshold; if so, the second specified option is determined as an incorrect option. Thus, counterfactual verification based on a large language model, on the basis of cognitive interference identification, further performs counterfactual verification on each remaining first option, actively eliminating incorrect options and ensuring the accuracy of the incorrect selection. By generating hypothetical statements and performing factual verification, the rationality of each option can be analyzed in depth from a logical perspective, thereby effectively avoiding incorrect selection of options due to superficial similarities.

[0081] In some optional implementations, calculating the conflict strength between the reasoning result and the known facts includes the following steps: Extract the key rules from the known facts.

[0082] In this embodiment, if both the known facts and the reasoning result involve some explicit rules or conditions, the conflict intensity can be calculated by examining the degree to which the reasoning result violates the rules. For example, in insurance claims, different types of diseases, treatment methods, etc., may correspond to different claim rules. The corresponding key rules can be obtained by extracting rules from the aforementioned known facts.

[0083] The data shows the degree to which the reasoning results violate the key rules.

[0084] In this embodiment, information related to the key rules in the reasoning results can be extracted, and then the degree to which the reasoning results violate the key rules can be statistically analyzed based on this information. For example, if it is known that a certain disease can only be claimed under a specific treatment method, and the treatment method in the reasoning results does not conform to the regulations, each instance of non-compliance increases the conflict intensity value.

[0085] Obtain the weight data corresponding to the key rule.

[0086] In this embodiment, the aforementioned weight data can be set according to the importance and influence of the key rules.

[0087] The degree data and the weight data are calculated and processed based on a preset conflict intensity calculation strategy to obtain the corresponding calculation results.

[0088] In this embodiment, the conflict intensity calculation strategy includes: weighted summation of the rule violation severity data to obtain the final conflict intensity. Suppose there are a total of rules, and the weight of the j-th rule is . The degree to which the reasoning result violates rule j is: The formula for calculating the conflict intensity Ci can be expressed as: .

[0089] Specifically, the above-mentioned degree data and weight data can be calculated and processed based on the calculation formula corresponding to the above-mentioned conflict intensity calculation strategy, and the obtained calculation result can be used as the conflict intensity between the above-mentioned reasoning result and the known facts.

[0090] The calculation result is used as the strength of the conflict between the reasoning result and the known facts.

[0091] Based on the above processing flow, this application extracts key rules from known facts; then statistically analyzes the degree of violation of key rules by the reasoning results; and obtains the weight data corresponding to the key rules. Subsequently, it calculates and processes the degree data and weight data based on a preset conflict intensity calculation strategy to obtain the corresponding calculation results. These calculation results are then used as the conflict intensity between the reasoning results and known facts. Thus, by extracting key rules from known facts and statistically analyzing the degree of violation of key rules by the reasoning results, and then calculating and processing the degree data and the weight data corresponding to the key rules based on the use of a conflict intensity calculation strategy, and using the obtained calculation results as the conflict intensity between the aforementioned reasoning results and known facts, this application can efficiently and accurately quantify the degree of contradiction between the reasoning results and known facts based on factual verification. This ensures the accuracy of the generated conflict intensity data and facilitates more objective and accurate selection and elimination of options based on the subsequent use of conflict intensity.

[0092] In some optional implementations of this embodiment, step S206 includes the following steps: Obtain the inference result of the specified path corresponding to the third specified option; wherein, the third specified option is any one of all the second options.

[0093] In this embodiment, the third specified option is any one of all the second options. When performing multi-path reasoning on the third specified option, multiple different reasoning paths are generated for it. Each path reasons from a different perspective or based on different information to comprehensively analyze the feasibility and rationality of the specified third option, and then outputs the path reasoning result corresponding to each reasoning path. Each reasoning path corresponds to one path reasoning result. The specified path reasoning result refers to all path reasoning results corresponding to the multiple reasoning paths of the third specified option.

[0094] Obtain the preset weighted fusion strategy.

[0095] In this embodiment, the weighted fusion strategy includes multiplying the score (path inference result) of each inference path by its corresponding path confidence and summing the results to obtain the final confidence score (confidence). Specifically, for each inference path, its confidence is determined by both path consistency and model confidence. Path consistency refers to the degree of consistency between the inference result of this path and the inference results of other paths, while model confidence refers to the model's confidence in the inference result of this path. By comprehensively considering these two factors, the confidence of each path can be evaluated more accurately.

[0096] The specific implementation process for determining path credibility includes: 1. Path Consistency Assessment. Data Collection and Organization: First, collect the reasoning results of all reasoning paths in the multi-path reasoning process. Organize these results into a result set, with each result corresponding to a reasoning path. For example, in an insurance claims review scenario, for a single claim case, there may be multiple reasoning paths based on different combinations of evidence and interpretations of clauses. Each path yields a preliminary conclusion on whether to pay out; these conclusions are collected. Consistency Measurement Method Selection: Select an appropriate consistency measurement method, such as the simple counting method, the voting method, or a more complex similarity measurement method. Simple Counting Method: Count how many paths have the same reasoning result as the current path. The more identical paths, the higher the consistency between the current path and other paths. Voting Method: Treat the reasoning result of each path as a "vote" and calculate the proportion of the current path's result among all results. The higher the proportion, the better the consistency. Similarity Measurement: If the reasoning result is numerical or can be converted to a numerical value (such as a probability value), methods such as cosine similarity and Euclidean distance can be used to calculate the similarity between the current path's result and other path results. The higher the similarity, the better the consistency. Calculate path consistency score: Calculate a consistency score for each inference path based on the selected consistency metric. For example, using a voting method, if there are 10 paths, and 7 of them have the same result as the current path, then the consistency score for the current path is 0.7.

[0097] 2. Model Confidence Assessment. Model Output Analysis: Analyze the confidence information output by the model (generative framework) during the reasoning process of each inference path. The form of confidence may differ for different models. Some models directly output a probability value representing the level of confidence in the inference result; others may use other indicators, such as the stability of the model's internal parameters. For example, in a machine learning-based inference model, when predicting whether a claim will be paid, the model outputs a probability value, which can be seen as the model's confidence in the reasoning result. Confidence Standardization: Since the range and scale of confidence output may differ between models, the confidence needs to be standardized to a uniform range (e.g., between 0 and 1). For example, if the model's output confidence range is -1 to 1, it can be mapped to 0 to 1 through a linear transformation, i.e., standardized confidence = (original confidence + 1) / 2. Determine the Model Confidence Score: Use the standardized confidence score as the model confidence score for each inference path.

[0098] 3. Determine Path Credibility Comprehensively. Weight Allocation: Determine the weights of path consistency and model confidence in determining path credibility. The weights can be adjusted based on specific scenarios and needs. For example, in some cases, model confidence may be given more weight, while in others, path consistency may be more important. Weights can be determined through expert experience, historical data analysis, or experimental verification. For instance, after analyzing multiple insurance claims cases, it was found that model confidence contributes significantly to prediction accuracy; therefore, the weight of model confidence is set to 0.6, and the weight of path consistency is set to 0.4. Calculate Path Credibility: Based on the determined weights, calculate the path credibility of each inference path. The formula is: w = a × consistency score + b × model confidence score, where a and b are the weights of path consistency and model confidence, respectively, and a + b = 1.

[0099] The weighted fusion strategy is used to perform weighted calculations on the inference results of the specified path to obtain the corresponding calculated data.

[0100] In this embodiment, based on the strategy content of the weighted fusion strategy described above, the path confidence level corresponding to the inference result of the specified path can be determined first, and then the inference result of the specified path and the path confidence level can be weighted and summed, and the calculated data obtained can be used as the specified confidence level data corresponding to the third specified option described above.

[0101] The calculated data is used as the specified confidence level data for the third specified option.

[0102] Based on the above processing flow, this application obtains the inference result of the specified path corresponding to the third specified option; wherein the third specified option is any one of all the second options; then, it obtains a preset weighted fusion strategy; subsequently, it performs weighted calculation processing on the specified path inference result based on the weighted fusion strategy to obtain the corresponding calculated data; and then uses the calculated data as the specified confidence level data of the third specified option. Thus, since different inference paths may yield different results in actual decision-making, this application, by using a weighted fusion strategy, can comprehensively consider the advantages and disadvantages of each path to calculate the specified confidence level data of the third specified option, ensuring the accuracy of the obtained specified confidence level data.

[0103] In some optional implementations of this embodiment, after step S208, the electronic device may further perform the following steps: Call the preset rendering engine.

[0104] In this embodiment, the rendering engine described above may specifically be a knowledge graph rendering engine.

[0105] The rendering engine generates an inference trajectory graph corresponding to the multi-path inference process.

[0106] In this embodiment, by using a rendering engine, the logical nodes recorded in the multi-path reasoning process described above can be treated as nodes in a graph, and the exclusion or support relationships between nodes can be treated as edges, thus constructing a visual reasoning trajectory graph. Nodes can be labeled with key information upon which the reasoning is based, and edges can be labeled with the reasons for exclusion or support.

[0107] Get the preset display method.

[0108] In this embodiment, the above-mentioned display method can specifically adopt a visual display method, and can adopt any of the following according to actual needs: 1) Visualization of the reasoning trajectory diagram itself. Node-edge graph (directed graph). Explanation: This is the most intuitive display method. Each logical node is drawn as a graphic of different shapes such as circles and squares as a node. The exclusion or support relationship between nodes is connected by lines (edges) with arrows. For example, green arrows represent support relationships, and red arrows represent exclusion relationships. The direction of the arrows clearly indicates the direction of the relationship. Key information on the reasoning basis can be marked inside the node, such as "According to clause X, this situation meets the compensation conditions"; the reasons for exclusion or support are marked on the edge, such as "This option is excluded due to the lack of key evidence Y". Advantages: Clearly and intuitively presents the reasoning path and the relationship between nodes, making it easy to quickly understand the entire reasoning process. Flowchart. Explanation: The reasoning process is presented in a certain order in the form of a flowchart. Starting from the starting node, it unfolds step by step according to the reasoning logic. Different branches represent different reasoning paths. Each step (node) is represented by a rectangle, and the reasoning basis is explained in detail inside the rectangle; the arrows connecting the steps indicate the reasoning direction, and the reasons for exclusion or support are marked on the arrows. For example, in insurance claims, starting from "submitting a claim application," there are multiple steps such as "document review" and "investigation and verification," each with corresponding judgments and bases. Advantages: It conforms to people's conventional thinking process, is easy to follow the reasoning steps, and is suitable for demonstrating a relatively linear reasoning process. 2) Comprehensive visualization methods combining other elements. Interactive webpage display. Explanation: The reasoning trajectory diagram is embedded in a webpage, utilizing the webpage's interactive functions, allowing claims adjusters to obtain more detailed information by clicking nodes or edges. For example, clicking a node will pop up a window displaying the complete clauses, cases, etc., on which the reasoning for that node is based; clicking an edge will display detailed explanations of the exclusion or supporting reasons and related evidence links. Simultaneously, zoom and drag functions can be set to facilitate viewing different parts of the entire reasoning trajectory diagram. Advantages: Provides a rich interactive experience, allowing claims adjusters to gain a deeper understanding of the reasoning details according to their needs, improving review efficiency. Dynamic demonstration video. Explanation: The reasoning process is made into a dynamic demonstration video, gradually showing the appearance of nodes, the connection of edges, and the explanation of the reasoning basis and reasons in the order of reasoning. Key nodes and relationships can be highlighted using animation effects. For example, when demonstrating supporting relationships, nodes and edges can be emphasized with highlights or blinking; when demonstrating reasons for exclusion, excluded options can be crossed out with strikethroughs. Simultaneously, voice narration provides a detailed explanation of the reasoning logic for each step. Advantages: It presents the reasoning process in a vivid and engaging way, making it particularly suitable for demonstrating to those unfamiliar with complex reasoning logic. It can quickly capture attention and convey key information.

[0109] The reasoning trajectory diagram is displayed based on the aforementioned display method.

[0110] In this embodiment, the aforementioned reasoning trajectory diagram can be displayed according to the selected display method. Specifically, in insurance claims review, the generated reasoning trajectory diagram can be directly embedded into the review system for claims adjusters to reference. This allows claims adjusters to clearly understand the entire decision-making process by viewing the reasoning trajectory diagram, including the reasoning path for each option, the reasons for exclusion, and the basis for the final answer.

[0111] Based on the above processing flow, this application calls a preset rendering engine; then generates an inference trajectory diagram corresponding to the multi-path inference process based on the rendering engine; subsequently obtains a preset display method; and then displays the inference trajectory diagram based on the display method. This application generates an inference trajectory diagram corresponding to the multi-path inference process based on the use of the rendering engine, and then displays the inference trajectory diagram based on the obtained display method, thereby visualizing the decision-making process and results, facilitating manual review, and enhancing the interpretability and auditability of the decision. Specifically, the visualized inference trajectory diagram enables reviewers to understand the decision-making process more intuitively and quickly, identify potential problems, and make corrections. At the same time, interpretability also meets the requirements of some industries for decision transparency and compliance, improving the credibility and acceptability of the decision.

[0112] In some alternative implementations, the user information obtained is subject to user consent and complies with relevant laws and policies.

[0113] 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.

[0114] Furthermore, this application can significantly reduce cognitive interference from large language models when facing multiple-choice and fuzzy tasks. Experimental results show that in insurance clause comparison and claims review tasks, this method improves accuracy compared to traditional LLM prompts and reduces the misjudgment rate under high interference conditions. Specifically, this solution has the following advantages: (1) High reliability: Improves decision stability through counterfactual verification and multi-path fusion; (2) Strong interpretability: Outputs complete exclusion process and logical chain, which is convenient for manual auditing; (3) Universal compatibility: Can be embedded into existing financial knowledge bases or customer service systems to achieve plug-and-play; (4) Compliance-friendly: The reasoning process is traceable and interpretable, meeting AI financial regulatory standards.

[0115] This application not only enhances the robustness of AI in complex financial reasoning tasks, but also provides a concrete path to solving the "black box problem," laying a technical foundation for reliable AI applications in fields such as intelligent insurance, intelligent customer service, and intelligent risk control.

[0116] 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.

[0117] 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.

[0118] 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.

[0119] 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.

[0120] 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).

[0121] 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.

[0122] 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.

[0123] like Figure 3 As shown, the AI-based problem-solving device 300 described in this embodiment includes: a receiving module 301, a parsing module 302, a first processing module 303, a second processing module 304, a reasoning module 305, an evaluation module 306, a third processing module 307, and an output module 308. Wherein: The receiving module 301 is used to receive user input question data; The parsing module 302 is used to parse the problem data based on a preset semantic decomposition model to obtain the corresponding triples; wherein, the triples include a target variable, a constraint set, and a candidate set, and the candidate set includes multiple candidate options; The first processing module 303 is used to perform interference evaluation on the candidate option pair based on the triplet using a preset interference evaluation network to filter out low-confidence options, and delete the low-confidence options from the candidate options to obtain the first option; The second processing module 304 is used to perform counterfactual verification on the first option based on a preset large language model to filter out incorrect options, and delete the incorrect options from the first option to obtain the second option. The reasoning module 305 is used to perform multi-path reasoning processing on each of the second options based on a preset generation framework if there are multiple second options, so as to obtain the path reasoning results corresponding to each of the second options respectively. Evaluation module 306 is used to evaluate the confidence of each path reasoning result to obtain corresponding multiple confidence data; The third processing module 307 is used to perform numerical comparisons on all the confidence data to determine the target confidence data with the highest value, and to obtain the target option corresponding to the target confidence data; The output module 308 is used to generate corresponding answer data based on the target options and output the answer data.

[0124] In some optional implementations of this embodiment, the parsing module 302 includes: The preprocessing submodule is used to preprocess the problem data to obtain the corresponding processed data; The analysis submodule is used to perform semantic analysis on the processed data based on the semantic decomposition model to obtain the corresponding semantic analysis results. The decomposition submodule is used to decompose the problem data based on the semantic analysis results to obtain the corresponding decomposition results; The first determining submodule is used to take the decomposition result as the triplet.

[0125] In some optional implementations of this embodiment, the first processing module 303 includes: Call the submodule to invoke the preset interference evaluation network; The first calculation submodule is used to calculate the semantic similarity and logical conflict degree between the first specified option and the target variable using the interference evaluation network for the first specified option; wherein, the first specified option is any one of all the candidate options; The second calculation submodule is used to calculate the semantic similarity and logical conflict degree based on a preset scoring calculation formula to obtain the corresponding score data. The first judgment submodule is used to determine whether the score data is less than a preset score threshold; The second determining submodule is used to determine the first specified option as a low-confidence option if the condition is met.

[0126] In some optional implementations of this embodiment, the second processing module 304 includes: A generation submodule is used to generate a hypothesis statement corresponding to a second specified option; wherein the second specified option is any one of all the first options; the hypothesis statement contains the reasoning result; The verification submodule is used to perform fact verification on the reasoning result using the large language model based on known facts; The third calculation submodule is used to calculate the conflict intensity between the reasoning result and the known facts if there is a contradiction in the reasoning result; The second judgment submodule is used to determine whether the conflict intensity is greater than a preset conflict threshold. The third determining submodule is used to determine the second specified option as an incorrect option if the condition is met.

[0127] In some optional implementations of this embodiment, the third calculation submodule includes: The extraction submodule is used to extract key rules from the known facts; The statistics submodule is used to collect data on the degree to which the reasoning results violate the key rules; The first acquisition submodule is used to acquire weight data corresponding to the key rule; The fourth calculation submodule is used to perform calculations on the degree data and the weight data based on a preset conflict intensity calculation strategy to obtain the corresponding calculation results; The fourth determination submodule is used to determine the calculation result as the conflict strength between the reasoning result and the known facts.

[0128] In some optional implementations of this embodiment, the evaluation module 306 includes: The second acquisition submodule is used to acquire the specified path reasoning result corresponding to the third specified option; wherein, the third specified option is any one of all the second options; The third acquisition submodule is used to acquire the preset weighted fusion strategy; The fifth calculation submodule is used to perform weighted calculation processing on the inference results of the specified path based on the weighted fusion strategy to obtain the corresponding calculation data; The fifth determining submodule is used to use the calculated data as the specified confidence level data for the third specified option.

[0129] In some optional implementations of this embodiment, the artificial intelligence-based problem-solving device further includes: The calling module is used to invoke the preset rendering engine; The generation module is used to generate an inference trajectory map corresponding to the multi-path inference processing based on the rendering engine; The acquisition module is used to acquire preset display methods; The display module is used to display the reasoning trajectory diagram based on the display method.

[0130] 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.

[0131] 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.

[0132] 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.

[0133] 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.

[0134] 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.

[0135] 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.

[0136] 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.

[0137] 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.

[0138] 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 question data; The problem data is parsed based on a preset semantic decomposition model to obtain corresponding triples; wherein, the triples include a target variable, a constraint set, and a candidate set, and the candidate set includes multiple candidate options; Based on the triplet, a preset interference evaluation network is used to evaluate the interference of the candidate option pairs to filter out low-confidence options, and the low-confidence options are removed from the candidate options to obtain the first option; Based on a pre-defined large language model, the first option is counterfactually verified to filter out incorrect options, and the incorrect options are then removed from the first option to obtain the second option. If there are multiple second options, then multi-path reasoning is performed on each second option based on a preset generation framework to obtain path reasoning results corresponding to each second option. The confidence scores of each path reasoning result are evaluated to obtain multiple corresponding confidence scores. The target confidence data with the highest value is determined by numerical comparison of all the confidence data, and the target option corresponding to the target confidence data is obtained. Generate corresponding answer data based on the target options, and output the answer data.

2. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The step of parsing the problem data based on a preset semantic decomposition model to obtain the corresponding triples specifically includes: The problem data is preprocessed to obtain the corresponding processed data; Based on the semantic decomposition model, semantic analysis is performed on the processed data to obtain the corresponding semantic analysis results; Based on the semantic analysis results, the problem data is decomposed to obtain the corresponding decomposition results; The decomposition result is taken as the triplet.

3. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The step of using a preset interference evaluation network to evaluate the interference of the candidate option pairs based on the triples in order to filter out low-confidence options specifically includes: Invoke the preset interference assessment network; For the first specified option, the interference evaluation network is used to calculate the semantic similarity and logical conflict degree between the first specified option and the target variable; wherein, the first specified option is any one of all the candidate options; The semantic similarity and logical conflict degree are calculated based on a preset scoring formula to obtain the corresponding score data. Determine whether the score data is less than a preset score threshold; If so, the first specified option is determined as a low-confidence option.

4. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The step of performing counterfactual verification on the first option based on a preset large language model to filter out incorrect options specifically includes: Generate a hypothesis statement corresponding to the second specified option; wherein the second specified option is any one of all the first options; the hypothesis statement contains the reasoning result; Based on known facts, the inference result is verified using the large language model. If the reasoning result is contradictory, then the strength of the conflict between the reasoning result and the known facts is calculated; Determine whether the conflict intensity is greater than a preset conflict threshold; If so, the second specified option will be identified as an incorrect option.

5. The problem-solving method based on artificial intelligence according to claim 4, characterized in that, The step of calculating the conflict strength between the reasoning result and the known facts specifically includes: Extract the key rules from the known facts; Statistical data on the degree to which the reasoning results violate the key rules; Obtain the weight data corresponding to the key rules; The degree data and the weight data are calculated and processed based on a preset conflict intensity calculation strategy to obtain the corresponding calculation results; The calculation result is used as the strength of the conflict between the reasoning result and the known facts.

6. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, The step of evaluating the confidence of each path reasoning result to obtain corresponding multiple confidence data specifically includes: Obtain the inference result of the specified path corresponding to the third specified option; wherein, the third specified option is any one of all the second options; Obtain the preset weighted fusion strategy; The weighted fusion strategy is used to perform weighted calculation on the inference results of the specified path to obtain the corresponding calculated data. The calculated data is used as the specified confidence level data for the third specified option.

7. The problem-solving method based on artificial intelligence according to claim 1, characterized in that, After the steps of generating corresponding answer data based on the target options and outputting the answer data, the method further includes: Call the preset rendering engine; Based on the rendering engine, a reasoning trajectory graph corresponding to the multi-path reasoning process is generated; Get the preset display method; The reasoning trajectory diagram is displayed based on the aforementioned display method.

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 parsing module is used to parse the problem data based on a preset semantic decomposition model to obtain the corresponding triples; wherein, the triples include a target variable, a constraint set, and a candidate set, and the candidate set includes multiple candidate options; The first processing module is used to perform interference evaluation on the candidate option pair based on the triplet using a preset interference evaluation network to filter out low-confidence options, and delete the low-confidence options from the candidate options to obtain the first option; The second processing module is used to perform counterfactual verification on the first option based on a preset large language model to filter out incorrect options, and delete the incorrect options from the first option to obtain the second option. The reasoning module is used to perform multi-path reasoning processing on each of the second options based on a preset generation framework if there are multiple second options, so as to obtain the path reasoning results corresponding to each of the second options respectively. The evaluation module is used to evaluate the confidence of each path reasoning result to obtain multiple confidence data. The third processing module is used to perform numerical comparisons on all the confidence data to determine the target confidence data with the highest value, and to obtain the target option corresponding to the target confidence data; The output module is used to generate corresponding answer data based on the target options and output 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