Title generation method, apparatus, and storage medium

Abstract

This application relates to a question generation method, apparatus, and storage medium. The question generation method includes: acquiring raw corpus; preprocessing the raw corpus to obtain standardized corpus; combining the standardized corpus with a large language model using specific prompts to generate an initial reasoning chain containing reasoning steps; processing the initial reasoning chain in parallel through an evidence fusion enhancement engine to output enhanced result data; wherein the evidence fusion enhancement engine includes factual evidence verification, logical consistency self-checking, knowledge graph alignment, and counterfactual and multi-perspective expansion; inputting the enhanced result data into a rule arbitration model to determine whether it passes; if it passes, fusing the initial reasoning chain with the enhanced result data to form an enhanced reasoning chain, and generating multiple types of questions based on the enhanced reasoning chain. By constructing a multi-source evidence fusion enhancement engine to perform parallel verification and correction of the reasoning chain, it can improve the accuracy of the generated question content.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to question generation methods, apparatus and storage media. Background Technology

[0002] Currently, the use of large language models for automated content generation has been widely applied in fields such as educational assessment, knowledge Q&A, and intelligent tutoring. The generation of questions, analysis, and reasoning content through large models has been widely used.

[0003] Currently, related technologies typically rely on calling a single large model, requiring the model to directly generate target content based on input materials by designing specific prompts. However, due to the inherent "illusion" phenomenon of large language models, the directly generated content often contains factual errors, fabricated data, or false citations. This results in the problem of low accuracy in the generated content of current large model question generation technologies.

[0004] There is currently no effective solution to the problem of low accuracy in the generated content in related technologies. Summary of the Invention

[0005] This embodiment provides a question generation method, apparatus, and storage medium to address the problem of low accuracy in generated content in related technologies.

[0006] Firstly, this embodiment provides a question generation method, including:

[0007] Obtain the raw corpus and perform preprocessing operations to obtain standardized corpus;

[0008] By combining the standardized corpus and calling the large language model through specific prompts, an initial inference chain containing inference steps is generated; the specific prompts include role definitions, task instructions, constraints, input materials, and format requirements.

[0009] The initial inference chain is processed in parallel by an evidence fusion enhancement engine to output enhanced result data. The evidence fusion enhancement engine includes factual evidence verification, logical consistency self-check, knowledge graph alignment, and counterfactual and multi-perspective expansion. The enhanced result data includes support score, logical consistency score, knowledge depth score, extended value score, correction suggestions, problem report, structured metadata, and derived content.

[0010] The enhanced result data is input into the rule arbitration model to determine whether it passes; if it passes, the initial reasoning chain is merged with the enhanced result data to form an enhanced reasoning chain, and multiple types of questions are generated based on the enhanced reasoning chain; the rule arbitration model calculates the comprehensive quality score by weighting the scores of each dimension in the enhanced result data with preset weights, determines whether the comprehensive quality score is greater than a preset passing threshold and combines it with a preset veto item, and outputs a decision result of passing or failing.

[0011] In some embodiments, the raw corpus includes paper abstracts, full-text papers, patents, question-and-answer pairs, and full-text textbooks; the preprocessing operations include format recognition, content parsing, and multi-level content analysis.

[0012] In some embodiments, the step of processing the initial inference chain in parallel through an evidence fusion enhancement engine to output enhanced result data includes:

[0013] Extract the key entities and the assertions corresponding to the key entities from the initial inference chain, query the preset multi-source knowledge base in parallel, and output the support score of the assertions and correction suggestions.

[0014] The initial reasoning chain is converted into a structured representation, and circular arguments and contradictory nodes in the structured representation are detected. A logical consistency score and a problem report are then output.

[0015] The conceptual entities in the initial reasoning chain are linked to a preset subject knowledge graph. The reasoning relationships in the initial reasoning chain are compared with the standard knowledge paths in the subject knowledge graph. Based on the comparison results, structured metadata containing knowledge point tags and cognitive levels and a knowledge depth score are output.

[0016] Key variables in the initial inference chain are identified, and counterfactual inference chains are generated by changing the values ​​of the key variables. The initial inference chain is then reconstructed based on preset perspectives from other disciplines to generate multidisciplinary perspective inference chains. The counterfactual inference chains and the multidisciplinary perspective inference chains are output as derived content, and the derived content is output as extended value scores according to preset extended value assessment rules.

[0017] In some embodiments, the step of extracting key entities and corresponding assertions from the initial inference chain, querying a pre-defined multi-source knowledge base in parallel, and outputting the support score of the assertions and correction suggestions includes:

[0018] Extract the key entities and the assertions corresponding to the key entities from the initial inference chain;

[0019] For each assertion, query the pre-defined multi-source knowledge base in parallel;

[0020] An attention mechanism is used to perform credibility-weighted fusion of the query results from the multi-source knowledge base;

[0021] A support score is calculated based on the degree of consistency between the query results and the assertions, and correction suggestions are generated.

[0022] In some embodiments, the step of converting the initial inference chain into a structured representation, detecting circular arguments and contradictory nodes in the structured representation, and outputting a logical consistency score and a problem report includes:

[0023] The initial inference chain is parsed into a structured directed graph containing nodes and logical relationships between nodes;

[0024] The structured directed graph is traversed according to the preset logical conflict detection rules to identify the circular argument structure and contradictory node pairs in the structured directed graph.

[0025] Calculate the logical consistency score based on the identification results and generate a problem report that includes the location of abnormal nodes.

[0026] In some embodiments, linking the conceptual entities in the initial inference chain to a preset subject knowledge graph, comparing the inference relationships in the initial inference chain with standard knowledge paths in the subject knowledge graph, and outputting structured metadata and a knowledge depth score containing knowledge point tags and cognitive levels based on the comparison results include:

[0027] Link the conceptual entities in the initial reasoning chain to the corresponding conceptual nodes in the preset subject knowledge graph;

[0028] Based on the linked concept nodes, the reasoning relationships between the concept nodes in the initial reasoning chain are extracted to form a path to be matched;

[0029] The path to be matched is compared with the predefined standard knowledge paths in the subject knowledge graph. Based on the comparison results, the core knowledge point tags and cognitive levels involved in the initial reasoning chain are determined, and structured metadata and knowledge depth scores are output.

[0030] In some embodiments, the process involves identifying key variables in the initial inference chain, generating a counterfactual inference chain by changing the values ​​of the key variables, reconstructing the initial inference chain based on preset perspectives from other disciplines to generate a multidisciplinary perspective inference chain, outputting the counterfactual inference chain and the multidisciplinary perspective inference chain as derived content, and outputting an extended value score for the derived content according to preset extended value assessment rules, including:

[0031] Identify the key variables in the initial inference chain and the current values ​​of the key variables;

[0032] At least one controllable counterfactual scenario is generated by changing the values ​​of the key variables;

[0033] Based on the counterfactual scenario, a counterfactual reasoning chain is derived, and the initial reasoning chain is reconstructed according to preset other disciplinary perspectives to generate a multidisciplinary perspective reasoning chain.

[0034] The counterfactual reasoning chain and the multidisciplinary perspective reasoning chain will be output as derivative content.

[0035] The derived content is then evaluated according to a preset extended value assessment rule, which outputs an extended value score.

[0036] In some embodiments, the step of inputting the enhanced result data into the rule arbitration model to determine whether it passes further includes:

[0037] If the decision is not made, the reason for failure output by the rule arbitration model will be fed back to the specific prompt.

[0038] After optimizing the specific prompt, the large language model is called again to generate the initial inference chain, and the process is repeated until it is determined to pass or the preset iteration limit is reached.

[0039] Secondly, this embodiment provides a question generation device, including: an initial processing module, a fusion enhancement module, and a rule arbitration module; wherein:

[0040] The initial processing module is used to acquire the raw corpus, preprocess the raw corpus to obtain standardized corpus, and combine the standardized corpus with a large language model through specific prompts to generate an initial inference chain containing inference steps; the specific prompts include role definitions, task instructions, constraints, input materials and format requirements;

[0041] The fusion enhancement module is used to process the initial inference chain in parallel through the evidence fusion enhancement engine and output enhanced result data. The evidence fusion enhancement engine includes factual evidence verification, logical consistency self-check, knowledge graph alignment, and counterfactual and multi-perspective expansion. The enhanced result data includes support score, logical consistency score, knowledge depth score, extended value score, correction suggestions, problem report, structured metadata, and derived content.

[0042] The rule arbitration module is used to input the enhanced result data into the rule arbitration model to determine whether it passes; if it passes, the initial reasoning chain and the enhanced result data are merged to form an enhanced reasoning chain, and multiple types of questions are generated based on the enhanced reasoning chain; the rule arbitration model calculates the comprehensive quality score by weighting the scores of each dimension in the enhanced result data with preset weights, determines whether the comprehensive quality score is greater than a preset passing threshold and combines it with a preset veto item, and outputs a decision result of passing or failing.

[0043] Thirdly, this embodiment provides a storage medium storing a computer program that, when executed by a processor, implements the steps of the question generation method described in the first aspect above.

[0044] Compared with related technologies, the question generation method, apparatus, and storage medium provided in this embodiment are different. The question generation method first acquires raw corpus and preprocesses it to obtain standardized corpus. Second, it combines the standardized corpus with a large language model using specific prompts to generate an initial reasoning chain containing reasoning steps. The specific prompts include role definitions, task instructions, constraints, input materials, and format requirements. Further, the initial reasoning chain is processed in parallel using an evidence fusion enhancement engine to output enhanced result data. The evidence fusion enhancement engine includes factual evidence verification, logical consistency self-checking, knowledge graph alignment, and counterfactual and multi-perspective expansion. The enhanced result data includes support for… The system evaluates and refines the data based on several criteria: degree of accuracy, logical consistency, knowledge depth, extended value, suggested corrections, issue reports, structured metadata, and derived content. Finally, the enhanced result data is input into a rule-based arbitration model to determine whether it passes. If it passes, the initial reasoning chain is merged with the enhanced result data to form an enhanced reasoning chain, which then generates multiple types of questions. The rule-based arbitration model uses preset weights to weight the scores of each dimension in the enhanced result data to obtain a comprehensive quality score. It then determines whether the comprehensive quality score exceeds a preset passing threshold and, combined with preset veto options, outputs a pass or fail decision. By constructing a multi-source evidence fusion enhancement engine to perform parallel verification and correction of the reasoning chain, it can improve the accuracy of the generated question content.

[0045] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0046] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0047] Figure 1 This is a hardware structure block diagram of a terminal for a title generation method according to an embodiment of this application;

[0048] Figure 2 This is a flowchart of a title generation method according to an embodiment of this application;

[0049] Figure 3 This is a flowchart of a question generation method according to one embodiment of this application;

[0050] Figure 4 This is a structural block diagram of a title generation apparatus according to an embodiment of this application. Detailed Implementation

[0051] To better understand the purpose, technical solution, and advantages of this application, the application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0052] Unless otherwise defined, the technical or scientific terms used in this application shall have the general meaning understood by a person skilled in the art to which this application pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this application do not indicate quantitative limitation and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this application are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, or B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. The terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific order of objects.

[0053] The method embodiments provided in this example can be executed on a terminal, computer, or similar electronic device with a certain computing power. For example, it can run on a terminal. Figure 1 This is a hardware structure block diagram of a terminal for a title generation method according to an embodiment of this application. For example... Figure 1 As shown, a terminal may include one or more ( Figure 1Only one is shown in the diagram. A processor 102 and a memory 104 for storing data are also included. The processor 102 may be, but is not limited to, a microprocessor (MCU) or a programmable logic device (FPGA). The terminal may also include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that… Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the terminal described above. For example, the terminal may also include components that are larger than... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown are illustrated.

[0054] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the question generation method in this embodiment. The processor 102 executes various functional applications and data processing by running the computer programs stored in the memory 104, thereby implementing the above-described method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0055] The transmission device 106 is used to receive or send data via a network. This network includes a wireless network provided by the terminal's communication provider. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 can be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0056] One embodiment of this application provides a question generation method. Figure 2 This is a flowchart of a question generation method according to an embodiment of this application, such as... Figure 2 As shown, the process includes the following steps:

[0057] Step S210: Obtain the original corpus and perform preprocessing operations to obtain standardized corpus.

[0058] Corpus refers to the original text materials used to generate questions. Acquired original corpora include academic papers, patent documents, question-and-answer pairs, and textbook texts. After obtaining different original corpora through a unified interface (such as an application programming interface, file upload interface, or database connection interface), preprocessing operations are performed to obtain standardized corpora. Specifically, preprocessing operations can include format parsing, content cleaning, chapter division, and priority sorting. Format parsing is used to unify the encoding and structure of different format files in the original corpus; content cleaning is used to remove noise information from the original corpus; chapter division is used to divide the structured text of the original corpus into different functional areas; and priority sorting is used to select core content as input for subsequent processing.

[0059] Step S220: Combine standardized corpus and call the large language model through specific prompts to generate an initial inference chain containing inference steps; the specific prompts include role definition, task instructions, constraints, input materials and format requirements.

[0060] Specifically, the specific prompts employ a dynamic domain-adaptive construction approach, automatically filling in relevant fields based on the topic domain identified during the preprocessing stage to guide the large language model in generating an initial inference chain that conforms to scientific norms. The specific prompts comprise five core components: role definition, task instructions, constraints, input materials, and format requirements.

[0061] The system includes several key components: Role Definition, Task Instructions, and Constraints. Role Definition specifies the expert role the model will play, such as "You are an expert in [domain] and need to perform rigorous reasoning analysis based on the following materials," where the domain is filled in adaptively based on the requirements. Task Instructions specify the requirements for generating the initial reasoning chain, such as "Please generate a complete reasoning chain according to the structure of premises → derivation steps → intermediate conclusions → final conclusions." Constraints limit the reasoning basis and output scope, such as "1. Reasoning steps must be clear and verifiable. 2. Each derivation step must have a clear logical relationship. 3. Conclusions must be based on the given materials and cannot include external knowledge." Input Materials are used to fill in standardized corpora, serving as the object of model reasoning. Format Requirements specify the structured output template, requiring it to conform to the preset JavaScript Object Notation (JavaScript Object Notation). The output is in JSON format, such as "{"reasoning_chain":"complete inference chain text","key_concepts":["list of key concepts"],"assumptions":["inference-dependent assumptions"]}". By setting a standard JSON format, it is ensured that the initial inference chain contains core fields such as the complete inference chain text, the list of key concepts, and the inference-dependent assumptions, thus achieving a unified and standardized output format. The design of structured specific prompts can effectively guide large language models to generate interpretable and traceable initial inference chains.

[0062] For example, the large language model application programming interface (API) is invoked, and the following optimization parameters are configured: the temperature parameter is set between 0.2 and 0.4 to control the randomness of the generated content; a lower temperature value ensures the consistency and stability of the output. The generated length is dynamically adjusted to 2 to 3 times the length of the original corpus to adapt to inference needs of different scales. The repetition penalty coefficient is set to 1.2 to reduce the probability of repeated phrases or sentence patterns in the generated content. A kernel sampling (Top-p) strategy is adopted, with a p-value of 0.9, to balance the diversity and quality of the generated content while avoiding the output homogenization caused by an excessively narrow sampling range. Through the above parameter configuration, the diversity and stability of the output can be balanced while ensuring the quality of the initial inference chain.

[0063] Step S230: The initial inference chain is processed in parallel through the evidence fusion enhancement engine to output enhanced result data; wherein, the evidence fusion enhancement engine includes factual evidence verification, logical consistency self-check, knowledge graph alignment, and counterfactual and multi-perspective expansion; the enhanced result data includes support score, logical consistency score, knowledge depth score, extended value score, correction suggestions, problem report, structured metadata, and derived content.

[0064] The initial reasoning chain is processed in parallel using an evidence fusion enhancement engine. This engine verifies and expands the initial reasoning chain in parallel across multiple dimensions, including factual evidence verification, logical consistency self-checking, knowledge graph alignment, and counterfactual and multi-perspective expansion.

[0065] Among them, the factual evidence verification verifies the assertions in the initial reasoning chain by querying multi-source knowledge bases, ensuring that each assertion is based on evidence, and outputs the support score and correction suggestions for each assertion; the logical consistency self-check focuses on the inherent rigor of the reasoning process, and outputs the logical consistency score and problem report by detecting logical problems such as circular arguments, contradictory nodes and missing premises; the knowledge graph alignment links the conceptual entities in the initial reasoning chain to the subject knowledge graph, positioning them in the macro-level subject knowledge framework, and outputs structured metadata containing knowledge point tags and cognitive levels and knowledge depth score; the counterfactual and multi-perspective extension generates controllable counterfactual scenarios and multi-disciplinary perspective reasoning by identifying key variables, breaking through the subject limitations of the original corpus, and outputs derivative content with evaluation value, including counterfactual reasoning chains and multi-disciplinary perspective reasoning chains, and outputs extension value scores for the derivative content according to the preset extension value evaluation rules.

[0066] The enhanced outcome data includes support scores, logical consistency scores, knowledge depth scores, extended value scores, correction suggestions, issue reports, structured metadata, and derived content from the various dimensions mentioned above, which serve as the basis for subsequent rule arbitration decisions.

[0067] Step S240: Input the enhanced result data into the rule arbitration model to determine whether it passes; if it passes, merge the initial reasoning chain with the enhanced result data to form an enhanced reasoning chain, and generate multiple types of questions based on the enhanced reasoning chain; the rule arbitration model calculates the comprehensive quality score by weighting the scores of each dimension in the enhanced result data with preset weights, determines whether the comprehensive quality score is greater than the preset passing threshold, and outputs the decision result of passing or failing in combination with the preset veto item.

[0068] The enhanced outcome data is input into the rule-based arbitration model to determine whether it passes or fails. The rule-based arbitration model receives enhanced outcome data from four different dimensions and makes a pass or fail decision through a transparent quantitative scoring mechanism.

[0069] Specifically, the rule-based arbitration model employs an interpretable weighted scoring model, transforming multi-dimensional qualitative analysis into quantitative decision-making criteria. First, it obtains dimensional scores for four different dimensions: support score S_fact, logical consistency score S_logic, knowledge depth score S_knowledge, and expanded value score S_expansion. Then, it performs a weighted calculation according to preset weights to obtain the comprehensive quality score Q = w1 × S_fact + w2 × S_logic + w3 × S_knowledge + w4 × S_expansion, where w1 to w4 are configurable weights, with the factual evidence verification dimension having the highest weight. Simultaneously, the rule-based arbitration model includes a veto clause; for example, if S_fact falls below a preset threshold or a fatal logical fallacy is detected, the decision is directly rejected, and other dimensional scores are no longer considered.

[0070] If the overall quality score reaches the preset passing threshold and no veto item is triggered, it is considered passed. At this point, the initial inference chain and the enhanced result data are intelligently fused to form an enhanced inference chain. The fusion operation includes: embedding support scores and sources of evidence after the assertions of the initial inference chain, including the name of the knowledge base supporting the assertions of the initial inference chain, document identifiers, or data sources; linking relevant knowledge path information in the knowledge graph in the mechanism inference part; and adding considered logical alternative explanations in the conclusion part, ultimately outputting a verified, corrected, and enriched enhanced inference chain.

[0071] If the rule arbitration model determines that the rule has passed, multiple types of questions are generated based on the enhanced reasoning chain. Using the enhanced reasoning chain as the core input, the question construction process is driven by multi-path reasoning information, realizing question stem generation, option construction, logical consistency verification, and automatic annotation of question meta-information, forming a standardized question bank data that can be directly used for model evaluation or teaching assessment.

[0072] Specifically, question generation includes multiple-choice question generation, true / false question generation, and short-answer and open-ended question generation. Multiple-choice question generation involves invoking an enhanced reasoning chain, extracting key semantics from the core conclusion to reconstruct the question stem, and using the final conclusion of the initial reasoning chain as the correct option. A counterfactual reasoning chain generation mechanism is introduced to perturb or logically replace the original premises, generating alternative conclusions that appear reasonable but have key logical deviations. Distractors with high semantic similarity and clear error types are selected, ensuring a high correlation between distractors and correct options and improving question differentiation. True / false question generation constructs propositions based on logical consistency analysis and contradiction detection results. Verifiable propositions are extracted from the initial reasoning chain, and expressions consistent with or conflicting with known knowledge are identified through logical relationship analysis. When logical contradictions are detected, incorrect propositions are generated; when consistency is detected, correct propositions are generated, ensuring that the right / wrong attribute stems from clear logical relationships rather than simple text rewriting. The generation of short-answer and open-ended questions is based on a multidisciplinary reasoning chain. It constructs multi-angle analysis of problems from different reasoning paths or knowledge dimensions, retains the key analytical dimensions in the initial reasoning chain, and designs open-ended expressions that require comprehensive explanation, comparative analysis, or multi-factor argumentation. This reflects the respondent's comprehensive understanding and reasoning depth, while maintaining consistency between knowledge objectives and reasoning direction.

[0073] Furthermore, after question generation, attribute labeling can be automatically performed, with each question inheriting the knowledge point tags from the source knowledge unit. The cognitive level of the generated questions is automatically determined by combining indicators such as the number of reasoning steps, the level of conceptual abstraction, the similarity of distractors, and logical complexity. The difficulty of the questions is estimated using a preset difficulty assessment model for subsequent screening, tiered evaluation, and adaptive testing. Domain experts conduct quality checks on the generated questions regarding their domain scope, research value, and difficulty level. Questions that pass the quality check are stored in a preset question bank database for subsequent model evaluation or teaching assessment.

[0074] In related technologies, there is often a reliance on a single large model call. Through designed prompts, traditional large language models are required to directly generate target content based on input materials. However, due to the inherent "illusion" phenomenon of large language models, the content they directly generate often contains factual errors, fabricated data, or false citations. For example, they may fabricate non-existent experimental data, distort the research conclusions in the original text, incorrectly cite authors or publication years, confuse the definition boundaries of similar concepts, or even fabricate completely non-existent external knowledge sources. These factual errors lack effective detection and correction mechanisms in the single model call mode, resulting in the problem of low accuracy in the content generated by current large language models in question generation technology.

[0075] To address the issue of low accuracy in the generated content in related technologies, this embodiment, through steps S210 to S240, firstly, acquires the original corpus and preprocesses it to obtain standardized corpus; secondly, combines the standardized corpus with a large language model using specific prompts to generate an initial inference chain containing inference steps; the specific prompts include role definitions, task instructions, constraints, input materials, and format requirements; subsequently, the initial inference chain is processed in parallel by an evidence fusion enhancement engine to output enhanced result data; wherein, the evidence fusion enhancement engine includes factual evidence verification, logical consistency self-check, knowledge graph alignment, and anti-faulty logic verification. The process involves expanding the evidence and providing multiple perspectives. The enhanced result data includes support scores, logical consistency scores, knowledge depth scores, extended value scores, correction suggestions, issue reports, structured metadata, and derived content. Finally, the enhanced result data is input into a rule-based arbitration model to determine whether it passes. If it passes, the initial reasoning chain is merged with the enhanced result data to form an enhanced reasoning chain, which generates multiple types of questions. The rule-based arbitration model uses preset weights to weight the scores of each dimension in the enhanced result data to obtain a comprehensive quality score. It then determines whether the comprehensive quality score exceeds a preset passing threshold and, combined with a preset veto option, outputs a pass or fail decision. By constructing a multi-source evidence fusion enhancement engine to perform parallel verification and correction of the reasoning chain, it can improve the accuracy of the generated question content.

[0076] Optionally, in one embodiment, the original corpus includes paper abstracts, full texts of papers, patents, question-and-answer pairs, and full texts of textbooks; the preprocessing operations include format recognition, content parsing, and multi-level content analysis.

[0077] After various corpora are accessed through a unified interface, they are converted into a unified lightweight markup language (such as Markdown) representation through format recognition and content parsing. Format recognition identifies different format types of input files, including Portable Document Format (PDF), Microsoft Word Document (DOC / DOCX), Extensible Markup Language (XML), and Hypertext Markup Language (HTML), and performs encoding consistency and integrity checks. Content parsing converts the identified files into a unified structured representation, including extracting the main text, filtering references and non-text information, removing headers and footers, eliminating formatting noise, repairing cross-page breaks, and cleaning up citation marks.

[0078] For example, taking the full text of a paper as an example, after receiving the file, the encoding is first standardized and the integrity is checked; secondly, a combination of rule matching and model recognition is used to automatically divide the text into chapters such as title, abstract, introduction, methods, experimental results, discussion, and conclusion, filtering out references and non-text content; then, standardization processing is performed, including removing headers and footers, eliminating formatting noise, repairing cross-page breaks, and cleaning up citation marks; finally, the chapters are prioritized, with summary paragraphs such as abstract, background, and conclusion being selected as input corpus.

[0079] Subsequently, after format recognition and content parsing, multi-level content analysis was performed on the corpus. Specifically, for scientific entity recognition and classification, a named entity recognition model trained on a large-scale scientific terminology corpus was used to identify key scientific research elements in the text. Specifically, for gene and protein entities, the recognition results were "OsNAC6 transcription factor" (classified as "transcription factor"), "OsP5CS gene", and "OsAPX2 gene"; for biological terms, the recognition results were "drought stress", "reactive oxygen species homeostasis", and "osmotic regulation"; for experimental method entities, the recognition results were "chromatin immunoprecipitation sequencing" and "real-time quantitative polymerase chain reaction"; for phenotypic and data entities, the recognition results were "survival rate", "leaf water potential", and "proline content". For argumentation role labeling, the functional role of each sentence or paragraph in the scientific argument is identified. For example, "Plants overexpressing a certain gene have a higher survival rate under drought conditions" is labeled as "experimental finding," "Because a certain gene activates the expression of another gene" is labeled as "mechanism explanation," and "Therefore, a certain gene enhances drought resistance by coordinating multiple pathways" is labeled as "research conclusion." For logical relationship extraction, the scientific logical relationships between entities and assertions are analyzed and extracted. For example, a series of causal relationship chains are established, such as "overexpression of a certain gene leads to upregulation of the expression of another gene," "upregulation of the expression of a certain gene leads to increased proline accumulation," and "increased proline accumulation contributes to improved drought resistance."

[0080] Furthermore, in one embodiment, the initial inference chain is processed in parallel by an evidence fusion enhancement engine to output enhanced result data, including:

[0081] The system extracts key entities and their corresponding assertions from the initial inference chain, queries a pre-defined multi-source knowledge base in parallel, and outputs the support scores of the assertions and suggestions for correction. It then converts the initial inference chain into a structured representation, detects circular arguments and contradictory nodes in the structured representation, and outputs a logical consistency score and a problem report. Next, it links the conceptual entities in the initial inference chain to a pre-defined subject knowledge graph, compares the reasoning relationships in the initial inference chain with the standard knowledge paths in the subject knowledge graph, and outputs structured metadata containing knowledge point tags and cognitive levels, along with a knowledge depth score, based on the comparison results. Finally, it identifies key variables in the initial inference chain, generates a counterfactual inference chain by changing the values ​​of these key variables, and reconstructs the initial inference chain based on pre-defined other subject perspectives to generate a multi-disciplinary perspective inference chain. The counterfactual inference chain and the multi-disciplinary perspective inference chain are output as derived content, and the derived content is then output with an extended value score according to pre-defined extended value assessment rules.

[0082] Specifically, the evidence fusion enhancement engine verifies and expands the initial reasoning chain through four different dimensions, including factual evidence verification, logical consistency self-check, knowledge graph alignment, and counterfactual and multi-perspective expansion.

[0083] Factual evidence verification first extracts key entities and their corresponding assertions from the initial inference chain. Key entities refer to the core concepts or objects involved in the initial inference chain, and assertions refer to statements about the attributes or relationships of those key entities. For each assertion, a pre-defined multi-source knowledge base is queried in parallel. This knowledge base includes authoritative domain databases and literature resources. Based on the consistency between the query results and the assertions, a support score and correction suggestions are output. The support score reflects the factual accuracy of the assertions, and the correction suggestions provide guidance for modifying erroneous or biased assertions.

[0084] The logical consistency self-check first converts the initial reasoning chain into a structured representation, which explicitly expresses the logical relationships between reasoning steps. It then checks for logical problems such as circular arguments and contradictory points within the structured representation. Circular arguments refer to arguments where the premises and conclusions are interdependent, while contradictory points refer to conflicting assertions within the same initial reasoning chain. Based on the check results, it outputs a logical consistency score and a problem report. The logical consistency score reflects the rigor of the initial reasoning chain, while the problem report helps pinpoint specific logical flaws.

[0085] Knowledge graph alignment first links the conceptual entities in the initial reasoning chain to corresponding nodes in a pre-built, structured knowledge graph of a subject. The subject knowledge graph is a pre-constructed structured knowledge base containing subject concepts and their relationships. The reasoning relationships in the initial reasoning chain are compared with standard knowledge paths in the subject knowledge graph, which refer to predefined chains of correct reasoning relationships between concepts. Based on the comparison results, structured metadata and a knowledge depth score are output. The structured metadata includes tags and cognitive levels of core knowledge points involved in the initial reasoning chain, while the knowledge depth score reflects the closeness of the connection between the initial reasoning chain and the subject knowledge system.

[0086] The counterfactual and multi-perspective expansion approach first identifies key variables in the initial reasoning chain, which refer to the core conditions or parameters affecting the reasoning conclusion. By changing the values ​​of these key variables, a counterfactual reasoning chain is generated, simulating changes in the reasoning conclusion under assumed conditions. Simultaneously, the initial reasoning chain is reconstructed based on pre-set perspectives from other disciplines, generating a multi-disciplinary perspective reasoning chain. This chain examines the same issue from different disciplinary angles. The counterfactual and multi-disciplinary perspective reasoning chains are output as derived content, and an extended value score is generated according to pre-set extended value assessment rules. For example, the extended value assessment rules can include evaluating the rationality of the derived content based on the logical consistency and scientific credibility of the counterfactual scenario; evaluating the innovativeness of the derived content based on the novelty and heuristic value of the multi-disciplinary perspective; and calculating the extended value score by weighting the rationality score and the innovativeness score. The extended value score reflects the degree to which the derived content contributes to the quality of the generated question.

[0087] The evidence fusion enhancement engine processes data in parallel across four different dimensions, and the outputs of each dimension—support score, logical consistency score, knowledge depth score, extended value score, correction suggestions, issue reports, structured metadata, and derived content—serve as input for subsequent arbitration decisions.

[0088] In one embodiment, key entities and their corresponding assertions are extracted from the initial inference chain, a pre-defined multi-source knowledge base is queried in parallel, and the support scores of the assertions and correction suggestions are output, including:

[0089] Extract key entities and their corresponding assertions from the initial inference chain; query a pre-defined multi-source knowledge base in parallel for each assertion; use an attention mechanism to perform credibility-weighted fusion of the query results from the multi-source knowledge base; calculate the support score based on the consistency between the query results and the assertions, and generate correction suggestions.

[0090] Factual evidence verification plays the role of a "scientific fact checker" in the evidence fusion enhancement engine, ensuring that every empirical assertion in the initial inference chain is supported by evidence. Specifically, firstly, it extracts key entities and their corresponding assertions from the initial inference chain. Key entities refer to the core concepts or objects involved in the initial inference chain, such as gene names, protein names, experimental conditions, or phenotypic indicators; assertions refer to statements about the attributes or relationships of these key entities, such as "overexpression of a certain gene leads to a certain phenotype" or "a certain protein has a regulatory relationship with a certain pathway."

[0091] Secondly, for each assertion, a pre-defined multi-source knowledge base is queried in parallel. This multi-source knowledge base includes multiple authoritative domain databases and literature resources, such as gene ontology databases, protein databases, pathway databases, and related academic literature databases. Parallel querying allows for the simultaneous acquisition of evidence from multiple knowledge sources, improving verification efficiency. Specifically, the query includes two methods: internal consistency checking and external knowledge base verification. Internal consistency checking involves comparing the claimed experimental results in the initial inference chain with the original data (such as digitized charts or structured abstracts) extracted from the same source paper to verify whether the data has been correctly cited and interpreted. External knowledge base verification involves linking the background knowledge or principle assertions in the initial inference chain to authoritative domain knowledge bases for cross-validation, assessing their universality and accuracy. For example, the statement "NAC transcription factor family is generally involved in stress response" is linked to authoritative domain knowledge bases (such as general protein resource libraries, gene ontology knowledge bases, and plant transcription factor databases) for cross-validation to assess its universality and accuracy.

[0092] Then, an attention mechanism is used to perform credibility-weighted fusion of query results from multiple knowledge bases. Different knowledge bases may have varying levels of authority and reliability. The attention mechanism automatically assigns different weights based on the credibility of each knowledge base, fusing evidence from multiple sources. Knowledge bases with higher credibility receive higher weights, giving query results from those bases a more significant role in the fusion result. Furthermore, the strength of evidence supporting each key assertion can be graded and labeled. Specifically, evidence with direct experimental support is labeled as "Direct Experimental Evidence (Level A)," such as "Chromatin immunoprecipitation sequencing peak plots show that a transcription factor binds to the promoter region of a gene." Evidence based on indirect or correlation analysis is labeled as "Indirect or Correlation Evidence (Level B)," such as "Co-expression network analysis suggests that a gene may be involved in a metabolic pathway."

[0093] Finally, a support score is calculated based on the consistency between the weighted fusion result and the assertion, and correction suggestions are generated. The support score reflects the factual accuracy of the assertion; the higher the consistency and the stronger the evidence, the higher the support score. If there are deviations or contradictions between the fusion result and the assertion, correction suggestions are generated based on the specific content of the deviation, such as pointing out erroneous information in the assertion and providing correct alternative statements or supplementing missing sources of evidence.

[0094] Factual evidence verification can generate quantifiable support scores for each assertion in the initial chain of reasoning, ensuring that there is evidence to support the claim and providing a quantitative basis for accuracy in subsequent arbitration decisions.

[0095] In one embodiment, the initial reasoning chain is converted into a structured representation, and circular arguments and contradictory nodes in the structured representation are detected. A logical consistency score and a problem report are then output. This includes: parsing the initial reasoning chain into a structured directed graph containing nodes and logical relationships between nodes; traversing the structured directed graph according to preset logical conflict detection rules to identify circular argument structures and contradictory node pairs in the structured directed graph; calculating a logical consistency score based on the identification results; and generating a problem report containing the location of abnormal nodes.

[0096] In the evidence fusion enhancement engine, logical consistency self-checking acts as a "logician," focusing on the inherent rigor of the reasoning process. Specifically, it first parses the initial reasoning chain into a structured directed graph. This structured directed graph contains various node types, including "premise" nodes (representing the initial conditions or assumptions upon which the reasoning depends), "experimental operation" nodes (representing the experiments or analytical operations performed), "intermediate discovery" nodes (representing the stage conclusions in the reasoning process), "mechanism inference" nodes (representing the reasoning explanation of the underlying mechanism), and "final conclusion" nodes (representing the final output conclusion of the initial reasoning chain). The directed edges between nodes represent the logical relationships in the reasoning steps. Through this scientific argumentation schema construction, the reasoning process described in natural language can be transformed into a computable and analyzable graph structure.

[0097] Secondly, the structured directed graph is traversed according to pre-defined logical conflict detection rules to identify logical fallacies and loopholes. Specific types of reasoning problems detected include: causal confusion (mistaking correlation for causation), overgeneralization (extending conclusions from limited conditions to broader scenarios), circular reasoning (the premises and conclusions of the reasoning are interdependent), lack of evidence (the reasoning steps lack necessary supporting evidence), and contradictory nodes (conflicting statements within the same initial reasoning chain). Then, alternative explanation compatibility analysis can be performed to detect logical conflicts. Alternative explanation compatibility analysis assesses whether existing evidence excludes other reasonable competing hypotheses. For example, in the reasoning that "a certain transcription factor enhances drought resistance," it analyzes whether the experimental design is sufficient to distinguish whether the cause is the direct transcriptional regulation function of the transcription factor or a non-specific stress response triggered by its overexpression. If alternative explanations are not excluded, they are marked in the problem report. Finally, uncertainty can be explicitly expressed to achieve logical conflict detection. Uncertainty explicitness is achieved by identifying and labeling links in the initial inference chain that rely on statistical significance (such as p-value), limitations of experimental techniques, or sample size. These uncertainties are then made explicit in the enhancement chain, for example, by labeling "This conclusion is based on statistical significance of p < 0.05, but the sample size is small and the results need further verification."

[0098] Finally, a logical consistency score is calculated based on the identification results, and a problem report containing the location of abnormal nodes is generated. The logical consistency score reflects the rigor of the initial reasoning chain, and the problem report is used to locate the specific type and location of logical defects, along with relevant explanations.

[0099] The logical consistency self-check can automatically identify logical problems in the initial reasoning chain, providing a quantitative basis for logical rigor in subsequent arbitration decisions and guiding the correction and optimization of the initial reasoning chain.

[0100] In one embodiment, the conceptual entities in the initial reasoning chain are linked to a predefined subject knowledge graph. The reasoning relationships in the initial reasoning chain are compared with the standard knowledge paths in the subject knowledge graph. Based on the comparison results, structured metadata containing knowledge point tags and cognitive levels, as well as a knowledge depth score, are output, including:

[0101] The conceptual entities in the initial reasoning chain are linked to the corresponding conceptual nodes in the pre-defined subject knowledge graph. Based on the linked conceptual nodes, the reasoning relationships between the conceptual nodes in the initial reasoning chain are extracted to form a path to be matched. The path to be matched is compared with the predefined standard knowledge paths in the subject knowledge graph. Based on the comparison results, the core knowledge point tags and cognitive levels involved in the initial reasoning chain are determined, and structured metadata and knowledge depth scores are output.

[0102] In the evidence fusion enhancement engine, knowledge graph alignment acts as a "subject taxonomist," positioning the specific initial reasoning chain within a macro-level subject knowledge framework. This involves: First, linking the conceptual entities in the initial reasoning chain to corresponding conceptual nodes in a pre-defined subject knowledge graph. This linking operation includes concept standardization, where entities in the initial reasoning chain are linked to standardized subject terminology ontologies to resolve synonym and alias issues. For example, the gene name "OsAPX2" is linked to the standard term "peroxidase activity" in the gene ontology knowledge base, thus eliminating ambiguity caused by different expressions of the same concept. Second, based on the linked conceptual nodes, the reasoning relationships between the conceptual nodes in the initial reasoning chain are extracted to form the matching path. Reasoning relationships refer to the logical connections between conceptual nodes, such as causal relationships, regulatory relationships, or functional associations. By extracting these relationships, the reasoning process described in natural language is transformed into a structured path representation.

[0103] Then, the path to be matched is compared with the predefined standard knowledge paths in the subject knowledge graph. Standard knowledge paths refer to the predefined chains of reasoning relationships between correct concepts in the knowledge graph, reflecting generally accepted knowledge connections within the subject area. Simultaneously, the core mechanism described in the initial reasoning chain—that is, the causal or regulatory relationships between entities involved in the initial reasoning chain, such as "a transcription factor activates the expression of a gene"—is compared with existing knowledge connections in the predefined subject knowledge graph to determine whether the core mechanism already exists in the knowledge graph, exists partially, or is completely novel. Specifically, if the core mechanism matches a known pathway in the knowledge graph (i.e., a series of predefined functionally related entities and their interactions, such as metabolic pathways or signal transduction pathways), the mechanism is determined to be known knowledge; if the core mechanism partially matches a known pathway, the mechanism is determined to be an extension of existing knowledge; if the core mechanism has no associated record in the knowledge graph, the mechanism is determined to be a potential new discovery. Through the above analysis, the degree of association between the initial reasoning chain and the existing knowledge system is evaluated, providing a basis for calculating the knowledge depth score.

[0104] Next, based on the comparison results between the path to be matched and the standard knowledge path, as well as the comparison results between the core mechanism and existing knowledge associations, the core knowledge point tags and cognitive levels involved in the initial reasoning chain are determined, and structured metadata is output. The knowledge point tags are used to identify the subject concepts involved in the initial reasoning chain, such as "transcriptional regulation," "drought stress response," and "reactive oxygen species metabolism," and can be labeled using a pre-defined subject classification system. The cognitive level is used to label the thinking level to which the initial reasoning chain belongs, including levels such as memory, understanding, application, analysis, evaluation, or creation. For example, an initial reasoning chain involving the analysis of relationships between multiple concepts can be labeled as the "analysis" level. The structured metadata, by integrating the above combination of knowledge point tags and cognitive levels, is output in a structured format (such as JSON) to automatically inherit the knowledge domain and difficulty level of the question.

[0105] Simultaneously, the contribution of the initial inference chain is evaluated based on the alignment results. The contribution type of the initial inference chain is determined according to the degree of matching between the initial inference chain and the standard knowledge path and its related position in the existing knowledge system: if the initial inference chain verifies the role of a known pathway in a new technology system, it is evaluated as a "verification contribution"; if the initial inference chain extends a known concept to a new function or mechanism, it is evaluated as an "expansion contribution"; if the initial inference chain reveals a completely new regulatory module or knowledge association, it is evaluated as an "innovation contribution".

[0106] Finally, based on the degree of matching between the path to be matched and the standard knowledge path, as well as the evaluation results of the contribution of the initial inference chain, a knowledge depth score is output according to the preset knowledge depth evaluation rules. The knowledge depth score reflects the closeness of the connection between the initial inference chain and the subject knowledge system, as well as its academic contribution value. The higher the score, the greater the knowledge depth of the initial inference chain and the higher its integration with the subject knowledge system.

[0107] By linking concepts, comparing the path to be matched with the standard knowledge path, analyzing the connection between the core mechanism and existing knowledge, and evaluating the contribution of the initial reasoning chain, knowledge graph alignment can locate the initial reasoning chain within a macro-level disciplinary knowledge framework, providing structured metadata and quantified knowledge depth scores for subsequent question generation.

[0108] Furthermore, in one embodiment, key variables in the initial inference chain are identified, and a counterfactual inference chain is generated by changing the values ​​of the key variables. The initial inference chain is then reconstructed based on preset perspectives from other disciplines to generate a multidisciplinary perspective inference chain. The counterfactual inference chain and the multidisciplinary perspective inference chain are output as derived content, and the derived content is output with an extended value score according to preset extended value assessment rules, including:

[0109] Identify key variables and their current values ​​in the initial inference chain; generate at least one controllable counterfactual scenario by changing the values ​​of the key variables; derive a counterfactual inference chain based on the counterfactual scenario, and reconstruct the initial inference chain according to preset perspectives from other disciplines to generate a multidisciplinary perspective inference chain; output the counterfactual inference chain and the multidisciplinary perspective inference chain as derived content; output an extended value score for the derived content according to preset extended value assessment rules.

[0110] Counterfactual and multi-perspective extensions play the role of a "thought experiment engine" in the evidence fusion enhancement engine. First, key variables in the initial inference chain and their current values ​​are identified. Key variables refer to the core conditions, parameters, or hypotheses that influence the conclusion of the initial inference chain, such as experimental conditions (e.g., "drought stress"), gene function (e.g., "overexpression of a certain gene"), or environmental factors (e.g., "temperature"). The current value refers to the specific set value of the variable in the initial inference chain, such as "moderate drought stress intensity" or "a certain gene is in an overexpression state." For example, from a paper on "a certain transcription factor enhancing drought tolerance," key variables such as "water deprivation," "OsNAC6 transcriptional activity," and "normal downstream target gene function" and their current values ​​can be deconstructed.

[0111] Secondly, at least one controllable counterfactual scenario is generated by changing the values ​​of key variables. A counterfactual scenario is a hypothetical situation constructed by systematically perturbing key variables, such as changing "overexpression of a gene" to "knockout of a gene," or changing "drought stress" to "salt stress." The changes can be univariate perturbations (changing only the value of one key variable) or multivariate combined perturbations (changing the values ​​of multiple key variables simultaneously) to generate counterfactual scenarios of varying complexity.

[0112] Then, a counterfactual reasoning chain is derived based on the counterfactual scenario. Specifically, by changing the values ​​of key variables, the logical structure of the initial reasoning chain is re-examined to arrive at new conclusions or intermediate steps under the assumed scenario. For example, if the initial reasoning chain is "overexpression of a certain gene leads to upregulation of the expression of another gene, which in turn leads to increased proline accumulation, ultimately improving drought resistance," changing the key variable "overexpression of a certain gene" to "knockout of a certain gene" results in a counterfactual reasoning chain that is "knockout of a certain gene leads to downregulation of the expression of another gene, which in turn leads to decreased proline accumulation, ultimately reducing drought resistance." Counterfactual scenarios include two types: univariate perturbation and mechanism isolation. Univariate perturbation involves changing the value of a key variable, such as "If the experiment were conducted under high salt stress instead of drought stress, what changes might occur in the regulatory network and phenotypic results of the OsNAC6 transcription factor?" Mechanism isolation involves specifically disrupting one function while preserving another, such as "If gene editing specifically disrupts the activation of the OsNAC6 transcription factor on the OsP5CS osmotic regulatory gene, but preserves its activation on a certain reactive oxygen species scavenging gene, how will the plant's drought tolerance change?" Through these operations, the generated counterfactual reasoning chain can simulate the changes in the reasoning conclusion under the hypothetical scenario, providing high-quality distractor material for multiple-choice questions.

[0113] Simultaneously, the initial reasoning chain is reconstructed based on pre-set perspectives from other disciplines, generating a multidisciplinary perspective reasoning chain. These other disciplinary perspectives refer to fields of study different from the discipline to which the initial reasoning chain belongs, such as shifting from a molecular biology perspective to an evolutionary biology perspective, a synthetic biology perspective, or an ecological perspective. Based on the theoretical framework and conceptual system of the selected disciplinary perspective, the entities, relationships, and conclusions in the initial reasoning chain are reinterpreted and reconstructed, generating alternative reasoning paths from other disciplinary perspectives. For example, the same research can be re-examined from an evolutionary biology perspective or a synthetic biology perspective. From an evolutionary biology perspective, questions can be raised such as, "From an evolutionary perspective, what adaptive advantages might the multi-pathway response mechanism coordinated by the OsNAC6 transcription factor confer on plants?" From a synthetic biology perspective, questions can be raised such as, "Can the OsNAC6 transcription factor and its downstream modules be designed as a portable 'drought-resistant circuit' to improve other crops?" Through these operations, the multidisciplinary perspective reasoning chain can examine the same problem from different disciplinary angles, providing a blueprint for designing open-ended questions with multi-faceted analysis for short-answer questions.

[0114] Finally, counterfactual reasoning chains and multidisciplinary perspective reasoning chains are output as derivative content. Simultaneously, extended value scores are output according to pre-defined extended value assessment rules. The extended value score reflects the degree to which the derivative content contributes to the quality of the generated questions. The pre-defined extended value assessment rules are specifically based on the rationality and heuristic nature of the derivative content to evaluate it: rationality refers to the logical consistency and scientific credibility of the counterfactual scenario, i.e., whether there are contradictions within the reasoning chain generated after changing key variables, and whether the hypothetical situation is established within the existing scientific knowledge framework; heuristic nature refers to the novelty of the multidisciplinary perspective reasoning chain and its value in designing open-ended questions, i.e., whether the introduced disciplinary perspectives exceed the conventional framework of the original analysis, and whether they can lead to new insights or new questions beyond the original conclusions.

[0115] Through operations such as key variable deconstruction, counterfactual scenario generation and deduction, and multidisciplinary perspective reconstruction, counterfactual and multi-perspective expansion can break through the limitations of the original corpus and initial reasoning chain, generate derivative content with evaluation value, provide high-quality distractors for multiple-choice questions, and provide open-ended question design ideas for short-answer questions with multi-angle analysis.

[0116] In one embodiment, the process of inputting the enhanced result data into the rule arbitration model to determine whether it passes or fails further includes: if it fails, feeding back the failure reason output by the rule arbitration model to a specific prompt; optimizing the specific prompt and then calling the large language model again to generate the initial inference chain, and re-executing until it passes or reaches the preset iteration limit.

[0117] When the enhanced result data input into the rule arbitration model results in a failure, the following steps are taken: First, a root cause diagnosis is performed. Based on the rule arbitration model's score and the problem reports output from different dimensions, the main reason for the failure is diagnosed. For example, the diagnosis might be "the factual score is too low because the statement 'change in the expression level of a certain gene' is seriously inconsistent with the original data." Second, structured feedback instructions are generated. The diagnosis result is transformed into actionable suggestions for modifying the prompt, such as adding the constraint "When describing an experimental result, it must be clearly stated that a certain gene is upregulated in treatment group A, but the change is not significant in treatment group B." Then, iterative regeneration is triggered. The optimized prompt is fed back, and the large language model is called again to generate the initial inference chain. The newly generated initial inference chain is then processed in parallel by the evidence fusion enhancement engine, and the enhanced result data is input into the rule arbitration model for judgment until it is judged as passing or the preset iteration limit is reached. A maximum number of iterations (e.g., 3 times) is set to avoid infinite loops. If the problem is deemed successful within the preset number of iterations, the problem generation process begins; if the problem is still unsuccessful after reaching the maximum number of iterations, the process is terminated and a failure report is output for manual review.

[0118] The aforementioned iterative closed-loop mechanism can automatically optimize specific prompts and regenerate the initial inference chain until the quality requirements are met or the iteration limit is reached, which can effectively improve the quality stability and accuracy of the generated content.

[0119] Figure 3 This is a flowchart of a title generation method according to one embodiment of this application. Figure 3 As shown, the method for generating this question includes the following steps:

[0120] Step S301: Input the raw corpus. Obtain the raw corpus and preprocess it to obtain standardized corpus.

[0121] Step S302: The large language model generates an initial inference chain. Using standardized corpora and prompts, the large language model is invoked to generate an initial inference chain containing inference steps.

[0122] Step S303, Evidence Fusion Enhancement Processing. The initial inference chain is input into the multi-evidence fusion enhancement engine for parallel processing. The multi-evidence fusion enhancement engine includes four dimensions: factual evidence verification, logical consistency self-check, knowledge graph alignment, and counterfactual and multi-perspective expansion. It verifies and expands the initial inference chain from different dimensions and outputs enhanced result data.

[0123] Step S304: Determine whether the enhanced result data passes the comprehensive evaluation of the rule arbitration model. The enhanced result data is input into the rule arbitration model for comprehensive evaluation. The rule arbitration model calculates the comprehensive quality score according to preset weighted scoring rules and makes a pass or fail decision based on a veto item. If it is determined to pass, proceed to step S307; if it is determined to fail, proceed to step S305.

[0124] Step S305: Determine whether the current iteration count has reached the preset iteration count limit. If the limit has not been reached, proceed to step S306; if the limit has been reached, proceed to step S308.

[0125] Step S306, Feedback and Iterative Optimization. Based on the evaluation results of the rule arbitration model and the problem report, diagnose the cause of failure, and transform the cause of failure into structured feedback instructions to optimize the prompts. Then return to step S302 to re-execute, while incrementing the iteration count by one.

[0126] Step S307: Generation of multiple question types. The initial reasoning chain is fused with the enhanced result data to form an enhanced reasoning chain, and multiple question types are generated based on the enhanced reasoning chain, including multiple choice questions, true / false questions, and short answer questions.

[0127] Step S308, Failure Termination and Manual Handling. Generate a failure analysis report, terminate the current process, and submit the failure report for manual review and processing.

[0128] Step S309, Large Model and Expert Quality Inspection. The questions generated in step S307 undergo automated quality inspection and expert manual quality inspection. Questions that pass the quality inspection are stored in a pre-set question bank database.

[0129] Steps S301 to S309 above, by constructing a multi-source evidence fusion enhancement engine to perform parallel verification and correction of the reasoning chain, and by generating better prompts through iterative loops, can improve the accuracy of the generated question content.

[0130] Figure 4 This is a structural block diagram of a title generation apparatus 40 according to an embodiment of this application, as shown below. Figure 4 As shown, the question generation device 40 includes: an initial processing module 42, a fusion enhancement module 44, and a rule arbitration module 46; wherein: the initial processing module 42 is used to acquire raw corpus, preprocess the raw corpus to obtain standardized corpus; combine the standardized corpus with a large language model through specific prompts to generate an initial reasoning chain containing reasoning steps; the specific prompts include role definitions, task instructions, constraints, input materials, and format requirements; the fusion enhancement module 44 is used to process the initial reasoning chain in parallel through an evidence fusion enhancement engine and output enhanced result data; wherein the evidence fusion enhancement engine includes factual evidence verification, logical consistency self-check, and knowledge graph. Spectral alignment and counterfactual and multi-perspective extensions; enhanced result data includes support score, logical consistency score, knowledge depth score, extended value score, correction suggestions, issue reports, structured metadata, and derived content; rule arbitration module 46 is used to input the enhanced result data into the rule arbitration model to determine whether it passes; if it is determined to pass, the initial reasoning chain and the enhanced result data are merged to form an enhanced reasoning chain, and multiple types of questions are generated based on the enhanced reasoning chain; the rule arbitration model calculates the comprehensive quality score by weighting the scores of each dimension in the enhanced result data with preset weights, determines whether the comprehensive quality score is greater than the preset pass threshold, and combines the preset veto item to output the decision result of pass or fail.

[0131] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination.

[0132] Furthermore, in conjunction with the question generation methods provided in the above embodiments, this embodiment can also provide a storage medium for implementation. This storage medium stores a computer program; when executed by a processor, the computer program implements any of the question generation methods described in the above embodiments.

[0133] It should be understood that the specific embodiments described herein are merely illustrative of the application and not intended to limit it. All other embodiments derived by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0134] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0135] Obviously, the accompanying drawings are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar situations based on these drawings without any creative effort. Furthermore, it is understood that although the work done in this development process may be complex and lengthy, for those skilled in the art, certain design, manufacturing, or production modifications made based on the technical content disclosed in this application are merely conventional technical means and should not be considered as insufficient disclosure of this application.

[0136] The term "embodiment" in this application refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily imply the same embodiment, nor does it imply that it is mutually exclusive with or independent of other embodiments. It will be clearly or implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0137] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of patent protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A method for generating questions, characterized in that, The question generation method includes: Obtain the raw corpus and perform preprocessing operations to obtain standardized corpus; By combining the standardized corpus and calling the large language model through specific prompts, an initial inference chain containing inference steps is generated; the specific prompts include role definitions, task instructions, constraints, input materials, and format requirements. The initial inference chain is processed in parallel by an evidence fusion enhancement engine to output enhanced result data. The evidence fusion enhancement engine includes factual evidence verification, logical consistency self-check, knowledge graph alignment, and counterfactual and multi-perspective expansion. The enhanced result data includes support score, logical consistency score, knowledge depth score, extended value score, correction suggestions, problem report, structured metadata, and derived content. The enhanced result data is input into the rule arbitration model to determine whether it passes; if it passes, the initial reasoning chain is merged with the enhanced result data to form an enhanced reasoning chain, and multiple types of questions are generated based on the enhanced reasoning chain; the rule arbitration model calculates the comprehensive quality score by weighting the scores of each dimension in the enhanced result data with preset weights, determines whether the comprehensive quality score is greater than a preset passing threshold and combines it with a preset veto item, and outputs a decision result of passing or failing.

2. The question generation method according to claim 1, characterized in that, The original corpus includes paper abstracts, full texts of papers, patents, question-and-answer pairs, and full texts of textbooks; the preprocessing operations include format recognition, content parsing, and multi-level content analysis.

3. The question generation method according to claim 1, characterized in that, The process of parallel processing the initial inference chain through an evidence fusion enhancement engine to output enhanced result data includes: Extract the key entities and the assertions corresponding to the key entities from the initial inference chain, query the preset multi-source knowledge base in parallel, and output the support score of the assertions and correction suggestions. The initial reasoning chain is converted into a structured representation, and circular arguments and contradictory nodes in the structured representation are detected. A logical consistency score and a problem report are then output. The conceptual entities in the initial reasoning chain are linked to a preset subject knowledge graph. The reasoning relationships in the initial reasoning chain are compared with the standard knowledge paths in the subject knowledge graph. Based on the comparison results, structured metadata containing knowledge point tags and cognitive levels and a knowledge depth score are output. Key variables in the initial inference chain are identified, and counterfactual inference chains are generated by changing the values ​​of the key variables. The initial inference chain is then reconstructed based on preset perspectives from other disciplines to generate multidisciplinary perspective inference chains. The counterfactual inference chains and the multidisciplinary perspective inference chains are output as derived content, and the derived content is output as extended value scores according to preset extended value assessment rules.

4. The question generation method according to claim 3, characterized in that, The process involves extracting key entities and corresponding assertions from the initial inference chain, querying a pre-defined multi-source knowledge base in parallel, and outputting the support scores and correction suggestions for the assertions, including: Extract the key entities and the assertions corresponding to the key entities from the initial inference chain; For each assertion, query the pre-defined multi-source knowledge base in parallel; An attention mechanism is used to perform credibility-weighted fusion of the query results from the multi-source knowledge base; A support score is calculated based on the degree of consistency between the query results and the assertions, and correction suggestions are generated.

5. The question generation method according to claim 3, characterized in that, The process of converting the initial reasoning chain into a structured representation, detecting circular arguments and contradictory nodes in the structured representation, and outputting a logical consistency score and a problem report includes: The initial inference chain is parsed into a structured directed graph containing nodes and logical relationships between nodes; The structured directed graph is traversed according to the preset logical conflict detection rules to identify the circular argument structure and contradictory node pairs in the structured directed graph. Calculate the logical consistency score based on the identification results and generate a problem report that includes the location of abnormal nodes.

6. The question generation method according to claim 3, characterized in that, The process involves linking the conceptual entities in the initial inference chain to a preset subject knowledge graph, comparing the inference relationships in the initial inference chain with the standard knowledge paths in the subject knowledge graph, and outputting structured metadata and a knowledge depth score containing knowledge point tags and cognitive levels based on the comparison results. Link the conceptual entities in the initial reasoning chain to the corresponding conceptual nodes in the preset subject knowledge graph; Based on the linked concept nodes, the reasoning relationships between the concept nodes in the initial reasoning chain are extracted to form a path to be matched; The path to be matched is compared with the predefined standard knowledge paths in the subject knowledge graph. Based on the comparison results, the core knowledge point tags and cognitive levels involved in the initial reasoning chain are determined, and structured metadata and knowledge depth scores are output.

7. The question generation method according to claim 3, characterized in that, The process involves identifying key variables in the initial inference chain, generating a counterfactual inference chain by changing the values ​​of these key variables, and reconstructing the initial inference chain based on preset perspectives from other disciplines to generate a multidisciplinary perspective inference chain. The counterfactual inference chain and the multidisciplinary perspective inference chain are then output as derived content, and the derived content is used to output an extended value score according to preset extended value assessment rules, including: Identify the key variables in the initial inference chain and the current values ​​of the key variables; At least one controllable counterfactual scenario is generated by changing the values ​​of the key variables; Based on the counterfactual scenario, a counterfactual reasoning chain is derived, and the initial reasoning chain is reconstructed according to preset other disciplinary perspectives to generate a multidisciplinary perspective reasoning chain. The counterfactual reasoning chain and the multidisciplinary perspective reasoning chain will be output as derivative content. The derived content is then evaluated according to a preset extended value assessment rule, which outputs an extended value score.

8. The question generation method according to claim 1, characterized in that, The step of inputting the enhanced result data into the rule arbitration model to determine whether it passes also includes: If the decision is not made, the reason for failure output by the rule arbitration model will be fed back to the specific prompt. After optimizing the specific prompt, the large language model is called again to generate the initial inference chain, and the process is repeated until it is determined to pass or the preset iteration limit is reached.

9. A question generation device, characterized in that, include: The module consists of an initial processing module, a fusion enhancement module, and a rule arbitration module; among which: The initial processing module is used to acquire the raw corpus, preprocess the raw corpus to obtain standardized corpus, and combine the standardized corpus with a large language model through specific prompts to generate an initial inference chain containing inference steps; the specific prompts include role definitions, task instructions, constraints, input materials and format requirements; The fusion enhancement module is used to process the initial inference chain in parallel through the evidence fusion enhancement engine and output enhanced result data. The evidence fusion enhancement engine includes factual evidence verification, logical consistency self-check, knowledge graph alignment, and counterfactual and multi-perspective expansion. The enhanced result data includes support score, logical consistency score, knowledge depth score, extended value score, correction suggestions, problem report, structured metadata, and derived content. The rule arbitration module is used to input the enhanced result data into the rule arbitration model to determine whether it passes; if it passes, the initial reasoning chain and the enhanced result data are merged to form an enhanced reasoning chain, and multiple types of questions are generated based on the enhanced reasoning chain; the rule arbitration model calculates the comprehensive quality score by weighting the scores of each dimension in the enhanced result data with preset weights, determines whether the comprehensive quality score is greater than a preset passing threshold and combines it with a preset veto item, and outputs a decision result of passing or failing.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the question generation method according to any one of claims 1 to 8.

Images