Method for discovering and generating problems through AI

Abstract

The invention belongs to the technical field of artificial intelligence and data management, and particularly relates to a method for discovering and generating problems through AI, which comprises the following steps: S1, receiving a natural language text of a service system as a work content, and distributing a global unique request identifier; s2, performing context analysis, standard comparative analysis and historical comparative analysis by adopting an AI problem discovery model; s3, fusing and sorting the content of the problem clue list based on the global unique request identifier, and generating a structured problem clue list; s4, querying a solution library according to clue attributes in the problem clue list, generating corresponding solution suggestions, and outputting the solution suggestions; and S5, according to metadata in the question clue list, a rule engine of the business system pushes the natural language question list and the solving suggestions to a corresponding terminal or system. The problem that problem grabbing is low in real-time performance and accuracy is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and data management, and particularly relates to a method for discovering and generating problems through AI. BACKGROUND

[0002] In the field of information technology, quickly and accurately discovering potential problems from massive unstructured natural language data (such as technical reports, user feedback, operation and maintenance logs, meeting minutes, etc.) is the key to effective decision-making and risk control. Traditional methods mainly rely on manual review, which is not only inefficient and costly, but also limited by individual experience and energy, and is prone to omissions and misjudgments, making it difficult to meet the analysis needs of large-scale data.

[0003] With the development of natural language processing technology, some automated text analysis tools have emerged. For example, some tools can perform keyword matching or sentiment analysis to identify negative comments, and some systems can perform simple compliance checks based on rule engines. However, these technical solutions have obvious limitations: first, they often analyze from a single dimension, lacking comprehensive consideration of deep-seated logical contradictions in the text, compatibility with external standards, and relevance to historical experience, resulting in incomplete and superficial problem discovery. Secondly, they can only output a classification label or risk score, but cannot generate clear and readable natural language problems to describe the discovered problems, thus failing to directly and effectively guide human experts to take action. SUMMARY

[0004] The application provides a method for discovering and generating problems through AI, solving the problems of low real-time and accuracy of problem extraction.

[0005] The basic scheme provided by the application is a method for discovering and generating problems through AI, comprising: S1: receiving natural language text of a business system as work content, and assigning a globally unique request identifier to each text; S2: analyzing the work content using an AI problem discovery model, the analysis content including: performing context analysis on the work content to identify internal logical and semantic consistency, semantic clarity, and information integrity, and generating internal clues; performing standard comparison analysis on the work content by comparing the text content with a pre-set domain knowledge base to generate compliance clues; performing historical comparison analysis on the work content by pattern matching the text content with a historical event database to generate historical clues; S3: Based on the globally unique request identifier, the internal clues, compliance clues and historical clues corresponding to the same text are fused and sorted to generate a structured question clue list; S4: According to the clue attribute in the question clue list, the solution library is queried to generate a corresponding solution suggestion, and the question clue list and the solution suggestion are converted into a natural language question list output; S5: According to the metadata in the question clue list and the rule engine of the business system, the natural language question list and the solution suggestion are pushed to the corresponding terminal or system.

[0006] Preferably, the type of the collected work content includes user feedback, operation monitoring, production collaboration, and document reporting.

[0007] Preferably, in S2, the context analysis includes: S2-1-1) using a pre-trained language model to encode the preprocessed text to generate a context-aware semantic vector representation of the text; S2-1-2) Based on the semantic vector representation, use a natural language reasoning model to analyze the logical contradiction of the sentences in the text to obtain a logical contradiction probability, and analyze the logical contradiction probability. When the logical contradiction probability is higher than the preset threshold, it is determined that there is a logical contradiction in the text, and a logical contradiction clue, an evidence fragment containing a contradictory sentence pair, and a confidence degree with the logical contradiction probability as the value are generated based on the determination; S2-1-3) Based on the semantic vector representation, use a coreference resolution model to analyze the text to obtain a coreference chain pointing to the same entity, and analyze the coreference chain. When a conflict is found in the description of the entity attributes or states in the same coreference chain, it is determined that there is a fact inconsistency in the text. Based on the determination, a fact inconsistency clue is generated, a coreference chain containing a conflict description is used as an evidence fragment, and a confidence degree is calculated based on the conflict degree; S2-1-4) Based on the semantic vector representation, use a pattern matching algorithm based on a pre-defined fuzzy word dictionary and rule set to analyze the text to obtain a matching result of the fuzzy expression, and analyze the matching result. When there is an expression in the text that matches the fuzzy word dictionary or conforms to the fuzzy rule, it is determined that the text is semantically ambiguous. Based on the determination, a semantic ambiguity clue is generated, an evidence fragment containing the ambiguous expression text is generated, and a confidence degree is assigned based on the rule matching strength; S2-1-5) Analyzing the text content using information extraction model based on semantic vector representation, obtaining the set of key information items mentioned in the text, and comparing the set with the pre-defined information item template library, when the key information items required by the template are missing in the text, it is determined that the text is incomplete, based on which the incomplete information clues, evidence fragments containing missing information items, and confidence level based on the importance of missing items are generated; S2-1-6) Structuring the analysis results of S2-1-2 to S2-1-5 to generate an internal clue list.

[0008] Further preferably, in S2, the standard comparison analysis includes S2-2-1) Using named entity recognition model or information extraction model to analyze the input text, automatically identifying and extracting key entities, indicators and their corresponding numerical values or state descriptions from the work content; S2-2-2) Taking the entities and indicators extracted in step S2-2-1 as query conditions, performing similarity retrieval in the domain knowledge base; S2-2-3) Comparing the status described in the text with the retrieved relevant standards for compliance, specifically including: a. Numerical comparison: when the indicators in the text and the standards in the domain knowledge base are specific numerical values, arithmetic or logical comparison is performed; b. State semantic comparison: when the text is a state description, calculate the semantic similarity between the state vector described in the text and the state vector described in the standard, if the semantic similarity is lower than the pre-set threshold, it is determined that the state is not consistent or deviates from the best practice; S2-2-4) Generating structured compliance clues based on the comparison results of step S2-2-3.

[0009] Further preferably, in S2, the historical comparison analysis includes S2-3-1) Using a pre-trained sentence embedding model to convert the text into a high-dimensional, numerical semantic vector; S2-3-2) Based on the semantic vector of the current text, retrieve in the pre-generated historical event vector database; S2-3-3) Calculate the cosine similarity between the semantic vector of the current text and each retrieved historical event vector, when there is a historical event whose similarity calculation result is higher than the pre-set matching threshold, it is determined that the current work content matches the historical event pattern; S2-3-4) Generating structured historical clues based on the matching results of step S2-3-3).

[0010] Further preferably, S3 includes S3-1) Based on the global unique request identifier, all clues output from the context analyzer, the standard comparator and the history comparator, which belong to the same original text, are collected to form an initial clue pool; S3-2) Determine the clue attribute in the clue pool; Wherein, the clue attribute includes fusion problem type, fusion confidence, fusion similarity, The fusion problem type selects the highest serious level clue type as the fusion problem type of the clue pool; The fusion confidence selects the maximum value of the confidence of the three clue types of the text as the fusion confidence; The fusion similarity is the similarity of the historical clue if there is a historical clue, otherwise it is 0; S3-3) Calculate the comprehensive score of each clue according to the following formula;

[0011] In the formula, The comprehensive score is, , , The weight coefficient is, The serious level corresponding to the fusion problem type is, The fusion confidence is, The fusion similarity is; S3-4) According to the comprehensive score, the sorted clue list is encapsulated as a structured problem clue list.

[0012] Further preferably, S4 includes S4-1) According to the clue attribute in the problem clue list, query the preset solution library, and match one or more solution suggestions for each text; S4-2) Combine the problem clue list with the solution suggestions obtained in step S4-1, and add a preset generation instruction to form a structured prompt word; S4-3) Input the prompt word constructed in step S4-2 to the large language model, arrange the structured problem clue and solution suggestion, and optimize the natural language problem draft generated by the large language model to generate a natural language problem list.

[0013] Further preferably, the preset generation instruction is set based on the business scenario target, the cognitive model of the target user role, the output reliability constraint and the readability quality standard.

[0014] The principle and advantages of the present application are: 1. Through AI model, multi-dimensional comprehensive analysis of context analysis, standard comparison analysis and history comparison analysis of text is carried out, the limitation of traditional single-dimensional analysis is broken through, the comprehensive mining and accurate identification of potential problems are realized, the accuracy of problem extraction is significantly improved, and the risk of problem omission or misjudgment is greatly reduced; 2. The fusion and sorting of multi-dimensional clues are realized through the global unique identifier, and combined with the business rule engine, it is pushed to the corresponding processing personnel or system, the efficiency of automatic processing of massive unstructured data is greatly improved, on the basis of ensuring the real-time processing of massive data, the accuracy of problem extraction is further improved through accurate aggregation of clues, the problem can be timely reached to the disposal subject, the problem flow cost is reduced, and the timeliness and pertinence of problem solving are ensured; 3. After the AI model generates a structured problem clue, the corresponding suggestion is matched and converted into a natural language problem list by linking the solution library, replacing the single output form of traditional tools that only output classification labels or risk scores, making the problem information more intuitive and accurate, ensuring the accuracy of information transmission after problem extraction, and directly providing clear disposal guidance for human experts to improve the real-time response of problems; 4. By constructing a dynamically updated domain knowledge base and historical event database, and linking the AI problem discovery model to realize real-time synchronization and analysis adaptation of data, it is ensured that the analysis basis is always consistent with the latest domain standards and historical experience, avoiding misjudgment due to lagging basis, further improving the accuracy of problem extraction, and at the same time, adapting to new scene data in real time to ensure the real-time analysis response. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The flowchart of the present application. DETAILED DESCRIPTION

[0016] The following will be further described in detail through specific embodiments: The specific implementation process is as follows: referring to Figure 1 A method for discovering problems and generating problems by AI, comprising: S1: receiving natural language text of a business system as work content, and assigning a global unique request identifier to each text; S1 includes S1-1) Work content collection; The types of collected work content include User feedback: customer service work order, user forum / community post, product evaluation, social media comment; Operation monitoring: system operation log, application program log, monitoring alarm information, business index report; Production cooperation: project progress report, code submission annotation, design review summary, meeting record; Document report type: stage summary, audit report, market analysis report, test report.

[0017] S1-2) Request identification and preprocessing; Assign a globally unique request identifier to each incoming text for tracking and correlating data throughout the processing chain; Clean and standardize the text to prepare for subsequent analysis; specifically, cleaning includes removing irrelevant characters, HTML tags, special symbols; standardization includes encoding Unicode, basic sentence segmentation, and in this embodiment, the text is converted to UTF-8.

[0018] S2: Analyze the work content using an AI problem discovery model; S2-1) Context analysis of work content, identify internal logical and semantic consistency, semantic clarity and information integrity of the text, generate internal clues; internal clues at least include clue type, evidence fragment and confidence; Specifically, the context analysis step includes S2-1-1) Use a pre-trained language model to encode the preprocessed text to generate a context-aware semantic vector representation of the text; in this embodiment, a BERT model based on Transformer is used for encoding; the vector output by the model is used as the underlying feature shared by all subsequent analysis sub-steps; S2-1-2) Based on the semantic vector representation, use a natural language reasoning model to analyze the logical contradiction of the sentences in the text to obtain a logical contradiction probability, and analyze the logical contradiction probability; when the logical contradiction probability is higher than the preset threshold, it is determined that there is a logical contradiction in the text, and based on the determination, a logical contradiction clue, an evidence fragment containing a contradiction sentence pair, and a confidence value based on the logical contradiction probability are generated; S2-1-3) Based on the semantic vector representation, use a coreference resolution model to analyze the text to obtain a coreference chain pointing to the same entity, and analyze the coreference chain; when a conflict is found in the description of the entity attributes or state within the same coreference chain, it is determined that there is a fact inconsistency in the text, and based on the determination, a fact inconsistency clue is generated, a coreference chain containing a conflict description is used as an evidence fragment, and a confidence value is calculated based on the conflict degree; S2-1-4) Based on the semantic vector representation, use a pattern matching algorithm based on a pre-defined fuzzy word dictionary and rule set to analyze the text to obtain a matching result of the fuzzy expression, and analyze the matching result; when there is an expression in the text that matches the fuzzy word dictionary or conforms to the fuzzy rule, it is determined that the text has semantic ambiguity, and based on the determination, a semantic ambiguity clue is generated, an evidence fragment containing the ambiguous expression text, and a confidence value is assigned based on the rule matching strength; S2-1-5) Analyzing the text content using an information extraction model based on semantic vector representation, obtaining a set of key information items mentioned in the text, and comparing the set with a pre-defined information item template library. When the key information item required by the template is missing in the text, it is determined that the text is incomplete, and based on this determination, an incomplete information clue, an evidence fragment containing a list of missing information items, and a confidence level based on the importance of the missing items are generated; wherein the information item template library is constructed based on compliance and mandatory specifications, domain knowledge and expert experience, historical high-quality document analysis, existing systems and data assets, and business goals and scenario requirements; S2-1-6) Structuring and packaging the analysis results of S2-1-2 to S2-1-5 to generate an internal clue list.

[0019] S2-2) Standard comparison analysis of work content, comparing text content with pre-set domain knowledge base to generate compliance clues; the compliance clues include clue type, standard value, reference basis, and confidence level; wherein the domain knowledge base is constructed based on international / national / industry standards and specifications, official regulations and policy documents, enterprise internal specifications and architecture constraints, verified professional knowledge and experience, and competitor and industry benchmark analysis; Specifically including the following steps: S2-2-1) Using a named entity recognition model or an information extraction model to analyze the input text, automatically identifying and extracting key entities, indicators, and their corresponding numerical values or state descriptions from the work content, for example, from the text "controller temperature has reached 45°C", extract the entity: "controller", indicator: "temperature", value: "45°C"; S2-2-2) Use the entities and indicators extracted in step S2-2-1 as query conditions to perform similarity search in the domain knowledge base; The search process is achieved by calculating the cosine similarity between the query vector and the vector of the standard document fragment in the domain knowledge base, and returning the Top-K most similar standards, rules or best practice document fragments as reference basis; S2-2-3) Compliance comparison of the conditions described in the text with the retrieved relevant standards, specifically including: a. Numerical comparison: when the indicators in the text and the standards in the domain knowledge base are both specific numerical values, perform arithmetic or logical comparison, for example, if the text value is 45°C and the retrieved standard value is "working temperature range: 0-40°C", it is determined that the standard is violated; b. State semantics comparison: when the text is a state description (such as "abnormal", "too high"), calculate the semantic similarity between the state vector of the text description and the state vector of the standard description, if the semantic similarity is lower than the preset threshold, it is determined that the state is inconsistent or deviates from the best practice; S2-2-4) Based on the comparison determination result of step S2-2-3, generate a structured compliance clue, If it is determined that the standard is violated or the practice is deviated, a clue of type standard violation or best practice deviation is generated.

[0020] S2-3) Historical comparison analysis is performed on the work content, and the text content is matched with the historical event database to generate historical clues; the historical clues include clue type, similarity, matched historical event ID, and confidence; The specific steps include: S2-3-1) Current event vectorization: A pre-trained sentence embedding model is used to convert the text into a high-dimensional, numerical semantic vector; the vector can capture the overall semantic features of the current text; S2-3-2) Based on the semantic vector of the current text, search in the pre-generated historical event vector database, wherein the search process is realized by K-nearest neighbor algorithm, find the K historical event vectors and their corresponding structured historical event records which are the nearest neighbors of the semantic vector of the current text in the vector space; S2-3-3) Calculate the cosine similarity between the semantic vector of the current text and each retrieved historical event vector, the formula is:

[0021] In the formula, S is the similarity, is the semantic vector of the current text, is the historical event vector; When there is a historical event whose similarity calculation result is higher than the preset matching threshold, it is determined that the current work content matches the historical event pattern; S2-3-4) Based on the matching determination result of step S2-3-3), generate a structured historical clue.

[0022] S3: Based on the globally unique request identifier, fuse and sort the internal clues, compliance clues and historical clues corresponding to the same text to generate a structured problem clue list; Specifically, this step is performed through the following sub-steps: S3-1) Based on the globally unique request identifier, all clues output from the context analyzer, standard comparator and history comparator belonging to the same original text are collected to form an initial clue pool; S3-2) Determine the clue attribute in the clue pool; Wherein, the clue attribute includes fusion problem type, fusion confidence, fusion similarity, The fusion problem type selects the highest severity level clue type as the fusion problem type of the clue pool; The fusion confidence selects the maximum value of the confidence of the three clue types of the text as the fusion confidence; The fusion similarity is the similarity of the historical clue if there is a historical clue; otherwise, it is 0; S3-3) Calculate the comprehensive score of each clue according to the following formula;

[0023] In the formula, is the comprehensive score, , , is the weight coefficient, is the severity level corresponding to the fusion problem type, is the fusion confidence, is the fusion similarity; S3-4) According to the comprehensive score, sort the clue list, and encapsulate the sorted clue list as a structured problem clue list, wherein the content of the problem clue list includes a globally unique request identifier, a clue attribute, a comprehensive score, an associated original clue ID list, and a fusion clue set creation time.

[0024] S4: According to the clue attribute in the problem clue list, query the solution library to generate corresponding solution suggestions, and convert the problem clue list and solution suggestions into a natural language problem list output; wherein the solution library is constructed according to historical problem root cause solutions, standard operation procedures (SOP) and operation manuals, official technical documents, internal and external knowledge accumulation, industry best practices and frameworks, and field expert experience; Specifically includes the following steps: S4-1) According to the clue attribute in the problem clue list, query the pre-set solution library to match one or more standardized and structured solution suggestions for each text; the solution library stores standard operation procedures, root cause repair solutions or best practices for different problem types; S4-2) Combine the problem clue list with the solution suggestion obtained in step S4-1, and add preset generation instructions to form a structured prompt word; the prompt word explicitly requires the large language model to generate fluent and professional natural language problems based on the provided problem clues and corresponding standard solutions; The preset generation instructions are set based on the business scenario target, the cognitive model of the target user role, the output reliability constraint, and the readability quality standard, aiming to guide the large language model to generate accurate, usable, and controllable natural language problems; S4-3) Input the prompt word constructed in step S4-2 into the large language model, integrate, polish, and reconstruct the structured problem clues and solution suggestions, output a preliminary natural language problem draft; and optimize the readability of the generated natural language problem draft (such as simplifying terms and adjusting sentence patterns), and finally encapsulate it into a specified data format to generate a natural language problem list.

[0025] S5: According to the metadata in the problem clue list and the rule engine of the business system, push the natural language problem list and solution suggestions to the corresponding terminal or system. The terminal can be a mobile APP or a webpage that interfaces different processing personnel, and the system can be a business process processing platform, such as a work order management system, a customer relationship management system, etc. When pushing, it is pushed according to the clue type, key entity / component, severity level, and belonging project / department; specifically, is the problem type "logical contradiction", "performance violation", or "security risk"? For example, performance problems are pushed to the operation and maintenance team, and logic problems are pushed to the product manager; for key entities / components, determine which core business object the problem involves? For example, if it involves "order service", it is pushed to the e-commerce team, and if it involves "payment gateway", it is pushed to the financial technology team; for the severity level, the severity level calculated according to the clue confidence and problem type (such as P0, P1, P2), P0-level problems may be pushed to both the front-line engineer and the team manager; for the belonging project / department, the attribution information identified from the text context or data source.

[0026] The above is only an embodiment of the present application, and the common knowledge of specific structures and characteristics in the scheme is not described in detail, and the ordinary skilled person in the art knows all the ordinary technical knowledge in the technical field of the present application before the application date or the priority date, can know all the prior art in the field, and has the ability to apply conventional experimental means before that date, and the ordinary skilled person in the art can improve and implement the present scheme under the guidance of the present application, and some typical known structures or known methods should not be an obstacle for the ordinary skilled person in the art to implement the present application. It should be pointed out that for those skilled in the art, without departing from the structure of the present application, a number of modifications and improvements can be made, which should also be considered as the protection scope of the present application, which will not affect the effect and practicality of the patent. The protection scope of the present application should be subject to the content of its claims, and the specific implementation mode and the like in the specification can be used to explain the content of the claims.

Claims

1. A method for discovering and generating problems by AI, characterized by, Comprise: S1: receive natural language text of a service system as a work content, and assign a globally unique request identifier to each text; S2: analyze the work content using an AI problem discovery model, the analysis content including: context analysis of the work content, identifying internal logic and semantic consistency, semantic clarity and information integrity within the text, generating internal clues; standard comparison analysis of the work content, comparing the text content with the pre-set domain knowledge base, generating compliance clues; historical comparison analysis of the work content, pattern matching the text content with the historical event database, generating historical clues; S3: based on the globally unique request identifier, fuse and sort the internal clues, compliance clues and historical clues corresponding to the same text to generate a structured problem clue list; S4: query the solution library according to the clue attributes in the problem clue list to generate corresponding solution suggestions, and convert the problem clue list and solution suggestions into a natural language problem list output; S5: according to the metadata in the problem clue list and the rule engine of the business system, push the natural language problem list and solution suggestions to the corresponding terminal or system. 2.The method of discovering and generating a problem by AI according to claim 1, wherein: The type of the collected work content includes user feedback, operation monitoring, production collaboration, and document reporting. 3.The method of discovering and generating a problem through AI according to claim 1, characterized in that: In S2, the context analysis includes: S2-1-1) using a pre-trained language model to encode the preprocessed text to generate a context-aware semantic vector representation of the text; S2-1-2) based on the semantic vector representation, using a natural language reasoning model to analyze the logical contradiction of the sentences in the text to obtain a logical contradiction probability, and analyzing the logical contradiction probability, when the logical contradiction probability is higher than a preset threshold, it is determined that there is a logical contradiction in the text, based on the determination, generating a logical contradiction clue, an evidence fragment containing a contradictory sentence pair, and a confidence degree taking the logical contradiction probability as a value; S2-1-3) based on the semantic vector representation, using a coreference resolution model to analyze the text to obtain a coreference chain pointing to the same entity, and analyzing the coreference chain, when a conflict is found in the description of the entity attributes or state within the same coreference chain, it is determined that there is a fact inconsistency in the text, based on the determination, generating a fact inconsistency clue, a coreference chain containing a conflict description as an evidence fragment, and a confidence degree calculated based on the conflict degree; S2-1-4) based on the semantic vector representation, using a pattern matching algorithm based on a pre-defined fuzzy word dictionary and rule set to analyze the text to obtain a matching result of the fuzzy expression, and analyzing the matching result, when there is an expression in the text that matches the fuzzy word dictionary or conforms to the fuzzy rule, it is determined that the text is semantically ambiguous, based on the determination, generating a semantic ambiguity clue, an evidence fragment containing the ambiguous expression text, and a confidence degree assigned based on the rule matching strength; S2-1-5) Analyzing the text content using information extraction model based on semantic vector representation, obtaining the set of key information items mentioned in the text, and comparing the set with the pre-defined information item template library. When the key information item required by the template is missing in the text, it is determined that the text is incomplete, and based on this determination, an incomplete information clue, an evidence fragment containing a list of missing information items, and a confidence level based on the importance of the missing item are generated; S2-1-6) Structuring the analysis results of S2-1-2 to S2-1-5 to generate an internal clue list. 4.The method of discovering and generating a problem by AI according to claim 3, characterized in that: In S2, the standard comparison analysis includes S2-2-1) Using a named entity recognition model or an information extraction model to analyze the input text, automatically identifying and extracting key entities, indicators, and their corresponding numerical values or state descriptions from the work content; S2-2-2) Using the entities and indicators extracted in step S2-2-1 as query conditions, performing similarity search in the domain knowledge base; S2-2-3) Comparing the status described in the text with the retrieved relevant standards for compliance, including: a. Numerical comparison: When both the indicators in the text and the standards in the domain knowledge base are specific numerical values, perform arithmetic or logical comparison; b. State semantic comparison: When the text is a state description, calculate the semantic similarity between the state vector described in the text and the state vector described in the standard. If the semantic similarity is lower than the pre-set threshold, it is determined that the state is not consistent or deviates from the best practice; S2-2-4) Based on the comparison results of step S2-2-3, generate structured compliance clues. 5.The method of discovering and generating a problem by AI according to claim 4, characterized in that: In S2, the historical comparison analysis includes S2-3-1) Using a pre-trained sentence embedding model to convert the text into a high-dimensional, numerical semantic vector; S2-3-2) Based on the semantic vector of the current text, search in the pre-generated historical event vector database; S2-3-3) Calculate the cosine similarity between the semantic vector of the current text and each retrieved historical event vector. When there is a historical event with a similarity calculation result higher than the pre-set matching threshold, it is determined that the current work content matches the historical event pattern; S2-3-4) Based on the matching determination result of step S2-3-3), generate structured historical clues. 6.The method of discovering and generating a problem by AI according to claim 5, characterized in that: S3 includes S3-1) Based on the globally unique request identifier, all clues from the context analyzer, standard comparator and historical comparator that belong to the same original text are collected to form an initial clue pool; S3-2) Determine the clue attributes in the clue pool; Wherein, the clue attributes include fusion problem type, fusion confidence, and fusion similarity, The fusion problem type is selected from the highest severity level of the clue type as the fusion problem type of the clue pool; The fusion confidence is the maximum value of the confidence of the three types of clues in the text; The fusion similarity is the similarity of the historical clue if there is one; otherwise, it is 0; S3-3) Calculate the comprehensive score of each clue according to the following formula; In the formula, is a comprehensive score, , , is a weight coefficient, is a severity level corresponding to the fusion question type, is a fusion confidence, is a fusion similarity; S3-4) Rank the list of clues according to the comprehensive score, and encapsulate the ranked list of clues as a structured list of question clues. 7.The method of discovering and generating a problem by AI according to claim 6, characterized in that: S4 includes S4-1) According to the clue attribute in the question clue list, query the preset solution library, and match one or more solution suggestions for each piece of text; S4-2) Combine the question clue list with the solution suggestions obtained in step S4-1, and add preset generation instructions to form a structured prompt word; S4-3) Input the prompt word constructed in step S4-2 into a large language model, organize the structured question clues and solution suggestions, and optimize the natural language question draft generated by the large language model to generate a natural language question list. 8.The method of discovering and generating a problem by AI according to claim 7, characterized in that: The preset generation instruction is set based on the business scenario target, the cognitive model of the target user role, the output reliability constraint, and the readability quality standard.

Images