A general domain thought chain corpus generation method and medium

Abstract

The application relates to a general field thought chain corpus generation method and medium, which comprises the following steps: acquiring general field corpora of different task types, converting the corpora into a unified standard format through structuring, and obtaining standardized corpora, wherein the task types include subjective types and objective types; generating initial thought chain texts and conclusions according to general matching reasoning links based on the standardized corpora of different task types; and verifying the initial thought chain texts and conclusions through a multi-dimensional automatic quality inspection mechanism to obtain final thought chain texts and conclusions. Compared with the prior art, the application has the advantages of generalization, process standardization, controllable quality and the like.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and knowledge engineering, and in particular to a method and medium for generating general-purpose domain thought chain corpora. Background Technology

[0002] With the increasing application of large language models in advanced cognitive tasks such as complex logical reasoning, mathematical calculation, and multi-step decision-making, the value of thought chain corpora is becoming increasingly prominent. These corpora, by explicitly demonstrating the complete thought process of problem decomposition, logical deduction, and conclusion induction, not only significantly improve the reasoning accuracy of models in general scenarios such as open-domain question answering, code generation, and scientific computing, but also greatly enhance the traceability and interpretability of model decisions, thus becoming a key training data resource driving leaps in model capabilities.

[0003] However, current mainstream CoT (CoT) corpus construction methods reveal significant systemic bottlenecks when facing the diverse and large-scale corpus generation needs of general domains: 1. Limitations of vertical domain customization: Existing solutions often customize generation logic and knowledge templates for specific vertical domains, making it difficult to flexibly migrate to other task types. Cross-domain reuse requires a lot of repetitive rule reconstruction and knowledge injection, which is inefficient. 2. Non-standardized generation process: From problem understanding and step decomposition to final expression, traditional CoT generation lacks a unified standardized path, the generation link has no fixed structure, the quality of the output corpus is uneven, and the interpretability of the reasoning logic is poor; different methods or manually written CoTs vary greatly in terms of logical rigor, step completeness and expression standardization, resulting in unstable quality of the output corpus, which seriously restricts the training effect of downstream models. 3. Inefficient reliance on manual quality inspection: The entire process still relies on expert verification as the final quality checkpoint. This approach is not only costly in terms of manpower and time-consuming, but also makes it difficult to form a large-scale, sustainable corpus production capacity, becoming the core bottleneck of corpus supply.

[0004] Currently, there is a lack of a general CoT corpus generation framework that can be reused across general domains, has a standardized generation path and automated quality inspection capabilities, resulting in low efficiency in obtaining high-quality CoT corpus required for training large models in general domains, which makes it difficult to meet the needs of large-scale training. Summary of the Invention

[0005] The purpose of this invention is to provide a method and medium for generating general domain thought chain corpora that can achieve universality, high quality, and reusability.

[0006] The objective of this invention can be achieved through the following technical solutions: A method for generating a general domain-specific thought chain corpus includes the following steps: We acquire general domain corpora of different task types, and convert them into a unified standard format through structured processing to obtain standardized corpora. The task types include subjective and objective categories. Based on standardized corpora of different task types, initial thought chain text and conclusions are generated according to a common matching reasoning link; Based on the initial thought chain text and conclusion, a multi-dimensional automated quality inspection mechanism is used to verify them, resulting in the final thought chain text and conclusion.

[0007] Furthermore, the structured corpus includes core fields, including: qid: Represents a unique identifier for the question; qtype: indicates the question type, which includes multiple choice, true / false, and short answer questions. Multiple choice, true / false, and short answer questions are all objective questions, while short answer questions are subjective questions. question: indicates the question text; choices: indicates options; answer: indicates the standard answer; explanation: indicates standard parsing; ref: Reference document or basis.

[0008] Furthermore, the steps for generating the initial thought chain text include: Based on the qtype, the task type and question type of the question are determined. Then, under the corresponding question type, the question text is combined to determine whether there is one or more value reasoning types among multi-step judgment, rule matching, and condition constraints. If they exist, the corresponding value reasoning needs to be executed during the answering process. If none of them exist, the answer is given directly. When none of them exist and the question can be answered directly from memory, the initial thought chain text of long text is not generated. Based on the corresponding value reasoning type, extract the core elements from the question text, including core conditions, limiting words, and implicit rules; Based on the core elements and the existing value reasoning types, the corresponding options or possible conclusions are verified one by one. After verification, an initial thought chain text is generated according to the set logic string. The corresponding options are those given in the objective category, and the possible conclusions are derived from the subjective category based on the question orientation of the question text, the standard analysis, and / or the reference documents.

[0009] Furthermore, the multi-step judgment refers to a question requiring two or more steps of deduction and / or calculation and / or comparison and / or elimination before an answer can be given; the rule matching refers to matching information in the question text with rules and / or definitions and / or clauses; and the condition constraints refer to the presence of constraint words or boundary conditions in the question text. The process of performing corresponding value reasoning during the response includes: When there are multiple steps involved, the intermediate steps are displayed and backtracking is ensured during the answering process. If there are no multiple steps involved, the answer can be given directly or by providing evidence and conclusions, generating a lightweight initial thought chain text. When a rule match exists, the answer should be provided by citing the rule points from the standard parsing and / or parameter document or by the basis and / or known knowledge points, and explaining the reason for the match. If no rule match exists, the rule points should not be forcibly introduced, and the answer should be provided according to common sense reasoning or calculation path. When constraints exist, during the answering process, the constraints and ranges are enumerated based on the constraint terms or boundary conditions, and then the corresponding options or possible conclusions are verified one by one. If there are no constraints, there is no need to enumerate them.

[0010] Furthermore, for objective categories, the steps for verifying each option one by one include: Semantic standardization is performed on all corresponding options to transcribe each corresponding option into a complete judgment statement, in which only the decidable facts or rule points are retained; The core elements in the question text are matched with each complete judgment statement one by one, and answers are given according to the existing value reasoning type to obtain the judgment result of the exclusion or retention of each complete judgment statement. Set prompt words, which indicate the generation of the initial thought chain text under the objective category based on the logical sequence of core elements in the question text → corresponding option judgment → elimination / retention → final conclusion.

[0011] Furthermore, for the subjective category, the step of verifying each possible conclusion includes: Based on the question text, identify the problem, analyze the requirements based on the problem, consult the reference documents or evidence, and answer according to the existing value reasoning type; Set prompt words, which indicate the generation of the initial thought chain text under the subjective category based on the logical sequence of problem confirmation → request analysis → document / basis query → response strategy.

[0012] Furthermore, for the everyday dialogue format in the subjective category, answers are given directly without generating initial thought chain text. If generation is required, only minimal analysis is retained, without proceeding to in-depth reasoning.

[0013] Furthermore, the steps for obtaining the final thought chain text and conclusion include: Define the dimensions for automatic quality inspection, including: completeness of issue representation, consistency of reasoning logic, and accuracy of factual content. The question text, the corresponding standard answer, the initial thought chain text, and the corresponding conclusion are input into the large language model, and validated under the automatic quality inspection dimension in conjunction with the prompt words. The validation result is output as either "correct" or "incorrect." During the large language model validation process... For objective questions, check if any of the following conditions are met: inconsistency between the standard answer and the corresponding conclusion, multiple answers for non-multiple-choice questions, or self-contradictory reasoning. If none of these conditions are met, output "correct". For subjective cases, verify whether the reasoning process is entirely derived from the content of the reference document or basis, and whether the standard answer is similar to the corresponding conclusion. If both are true, output "correct"; otherwise, output "incorrect". All initial thought chain texts and conclusions that are determined to be correct are taken as the final thought chain texts and conclusions.

[0014] Furthermore, it also includes confidence assessment and dynamic iteration steps to score the confidence of the generated thought chain text and dynamically adjust the prompt words, wherein the confidence scoring step includes: For the aforementioned thought chain text and conclusion, a confidence score is calculated, where the calculation expression is: , In the formula, , , For configurable weights, Scoring for logical consistency To score for consistency of conclusions, Score for the accuracy of the facts; Determine whether the confidence score is lower than the set minimum confidence threshold. If so, regenerate according to the general matching inference link; otherwise, no regeneration is required. The steps for dynamically adjusting the prompt words include: For initial thought chain texts and conclusions that fail validation, and thought chain texts and conclusions whose confidence scores do not reach the minimum confidence threshold, the prompt words are automatically adjusted according to their failure type: When the failure type is a logical reasoning problem, add constraint statements, including step-by-step analysis and no skipping steps, as prompts to constrain the generation process; When the failure type is a factual problem, add constraint phrases, including those based solely on reference documents or evidence, as prompts to constrain the generation process; When the failure type is expression redundancy, constraint words, including limiting the length of the thought chain text and compressing intermediate steps, are added as prompt words to constrain the generation process.

[0015] The present invention also provides a storage medium storing a program thereon, which, when executed, implements a general domain thinking chain corpus generation method as described above.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention achieves the production of a universal, high-quality, and reusable CoT corpus through a standardized automatic generation process by the triple integration of raw data standardization, universal matching reasoning link and automated verification.

[0017] (2) In the universal matching reasoning link proposed in this invention, each step from question type identification to logic string output has corresponding execution rules, avoiding unruly subjective analysis. On-demand generation is achieved through value reasoning type, avoiding the generation of long and redundant CoT for simple questions, and also preventing the CoT reasoning steps of complex questions from being missing, so that the level of detail of CoT is highly matched with the reasoning needs of the question.

[0018] (3) The universal matching reasoning link proposed in this invention shows the complete reasoning process from problem to conclusion. The reasoning chain is unbroken and truly reflects the deductive nature of the thinking chain, rather than simply giving a simple explanation of the conclusion. It can ensure that the thinking deviates during the CoT generation process and guarantee the generation quality.

[0019] (4) The universal matching reasoning link proposed in this invention can be designed for different task types. There is no need to design separate generation rules for different question types. One process can be adapted to all scenarios and has strong universality.

[0020] (5) The present invention also proposes a confidence assessment and dynamic iteration step. In this step, the generated thought chain text is scored with confidence and the thought chain text that fails the quality inspection or does not meet the confidence threshold is regenerated. In the regeneration process, the prompt words are adjusted according to the failure type. Thus, high-quality and verifiable thought chain content can be stably output under high-efficiency generation and is automatically evaluated without relying on manual verification. Attached Figure Description

[0021] Figure 1This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0023] This embodiment provides a general-domain thinking chain corpus generation method. This method aims to combine standardized data structures, streamlined CoT generation processes, and automated quality control mechanisms to construct a CoT corpus generation framework that can be reused in general domains. Through explicit prompt word templates, it guides the model to stably output high-quality, verifiable thinking chain content. Specifically, as... Figure 1 As shown, the method includes the following steps: S1, Raw Corpus Input We acquire general domain corpora from question banks, policy texts, question-and-answer data, and other sources. These general domain corpora include multiple task types, mainly subjective and objective. The subjective corpora include three question types: multiple choice, multiple selection, and true / false. The objective corpora mainly include one question type: question-and-answer.

[0024] S2, Structured Transformation The general domain corpora of different task types obtained in step S1 are standardized using a JSON / JSONL structure to obtain standardized corpora, whose core fields include: •qid: Unique identifier for the question •qtype: Question type (e.g., multiple choice, true / false, short answer) •question: Question text •choices: Options (required for objective questions) • Answer: Standard Answer •Explanation: Standard explanation (if applicable) •ref: Reference document or basis (required for Q&A questions) This standardized structure ensures that data from different sources and with different task types can be processed uniformly within the same generation framework.

[0025] S3, CoT Text and Conclusion Generation This step mainly involves matching and reasoning based on the task type according to a common matching reasoning link, in order to generate CoT text and conclusions (answers) under different task types.

[0026] Regardless of whether the question is objective or subjective, the general underlying process for generating this CoT text is as follows: 1. Task type identification: Determine whether the question is an objective or subjective question, and the specific question type, based on qtype; 2. Determining the value of reasoning: Under the corresponding question type, determine whether the question has multiple steps, rule matching, or condition constraints. If any one of these exists, the corresponding value reasoning needs to be performed during the answering process. If none of these exist, the answer can be given directly. in: (1) Multi-step judgment: The question requires two or more steps of derivation and / or calculation and / or comparison and / or elimination to reach a conclusion (common signals: the appearance of "first...then... / separately... / combined...", the need for intermediate variables or multiple conditions). If they exist: the intermediate steps must be explicitly listed and backtracking must be guaranteed; if they do not exist: a lightweight CoT with a sentence of "basis + conclusion" is allowed, or the conclusion can be given directly according to the task requirements.

[0027] (2) Rule matching: The essence of the question is to match the rules and / or definitions and / or clauses with the information in the question stem (common signals: the appearance of "compliant / non-compliant / belongs to / should / must not / according to... regulations"). If they exist: explicitly cite the corresponding rule point in the CoT (from the explanation / ref / known knowledge point) and explain the reason for the match or non-match; if they do not exist: do not forcibly introduce the rule statement, and output according to common sense reasoning or calculation path.

[0028] (3) Conditional constraints: The question stem contains constraint words or boundary conditions (such as "only / must / except / at least / not more than / under the condition of / which of the following is incorrect"). If they exist: in the process of answering, first enumerate the constraint conditions and ranges based on the constraint words or boundary conditions extracted from the question text, and then check the candidate conclusions one by one; if they do not exist: there is no need to expand the condition enumeration, and directly enter the element decomposition and candidate conclusion verification.

[0029] The enumeration logic is as follows: According to the definition in the document, the complete enumeration process follows this chain of steps: Source: Question text.

[0030] Extraction criteria: "Core elements" extracted from the question text, specifically including core conditions, limiting words, and implicit rules.

[0031] Direct enumeration is based on the "constraint words" or "boundary conditions" contained in the problem text (such as "only / must / except / at least / not more than" etc.).

[0032] (4) Output strategy: If the above three types of reasoning value are not obvious and the question can be answered directly from memory, it is not recommended to generate a long CoT; only retain the minimum necessary explanation or skip CoT generation (especially for subjective questions in daily conversation).

[0033] 3. Deconstruction of Reasoning Elements: Extract the core conditions, limiting words, and implicit rules from the question stem (problem); 4. Candidate conclusion enumeration: Verify each option or possible conclusion one by one. The options are those given in the objective category. For true / false questions, the options are considered as correct and incorrect. The possible conclusions are generated in the subjective category based on the question text, standard analysis and / or reference documents. 5. Logical chain output: Generate CoT text in the order of "condition → judgment → conclusion".

[0034] Specifically, the generation process of objective questions and subjective questions differs in the following ways: For objective questions: (1) Decomposition of reasoning elements Identify the objects of study (concepts / devices / rules / behaviors) - extract limiting conditions (time, scope, degree, negation words, etc.); (2) Standardization of option semantics Translate the options into complete judgment statements, eliminate differences in expression, and retain only the facts or rule points that can be determined; (3) Item-by-item matching and exclusion Compare the conditions in the question stem with the semantics of the options one by one - answer according to the existing value reasoning type, and clarify the basis for determining "match / disagree" - give clear reasons for excluding the disagreement; (4) Inference Link Connection This embodiment introduces prompt words to guide and drive the CoT text generation process. These prompt words require output in the logical sequence of "question conditions → option judgment → elimination / retention → final conclusion" - ensuring that each step can be traced back to the question stem or known rules.

[0035] CoT generation rule: Only retain reasoning steps directly related to the conclusion - prohibit background descriptions unrelated to the judgment.

[0036] Applicable to: knowledge-based questions, rule-matching questions, and logical reasoning questions.

[0037] This approach is not suitable for simple factual questions that can be answered directly from memory and have no reasoning value.

[0038] For subjective questions: (1) Questions with clear answers Identify the problem based on the question text, analyze the requirements based on the problem, and consult reference documents or evidence. Answer according to the existing value reasoning type. Then, generate the CoT text under the subjective category according to the logical sequence of question identification → requirement analysis → document / evidence query → response strategy based on the set prompt words.

[0039] Generation principles: - The reasoning process must be strictly derived from the provided reference documents or basis. - It is forbidden to introduce policies, regulations or conclusions that are not present in the documents. - Use fixed structured steps to express the thought process.

[0040] Recommended structure: 1. Problem confirmation 2. Request analysis 3. Policy / document search 4. Response strategy (2) Everyday conversation These types of questions typically do not involve complex reasoning and only require a direct answer: In principle, generating CoT is not recommended. If it is necessary to generate it, only a minimal analysis should be retained, and it should not be used in the deep inference process.

[0041] Preferably, this embodiment further divides the prompts introduced by the thinking chain into general prompts and prompts specifically for question-and-answer questions. Examples of general prompts are as follows: You need to engage in structured reasoning step by step, carefully analyzing the question content and maintaining a clear and logical thought process. Each reasoning step should have clear evidence, avoiding irrelevant or repetitive arguments. Please keep the length of your thought process within a reasonable range, ensuring conciseness and high information density.

[0042] The specific prompts for the question and answer questions include: Structured reasoning instructions: Step 1: Identify the problem and clarify the core and key requirements of the question.

[0043] Step 2: Request analysis to identify users' actual needs and goals.

[0044] Step 3: Policy research, strictly based on the content of the reference documents for analysis, and no expansion is allowed.

[0045] Step 4: Response strategy, formulate the final response based on the document content.

[0046] Constraints: - All reasoning must be derived from the referenced document. - If the question is outside the scope of the document, it must be explicitly rejected.

[0047] Preferably, this embodiment also imposes format constraints on the generated CoT text: When it is necessary to standardize the transcription of an existing thought process, the following should be followed: - Use Chinese numerals for step numbers (Step 1, Step 2, etc.) - Do not add or delete the original meaning - Do not use non-Chinese characters - Do not use line breaks or insert irrelevant content.

[0048] S4, Automated Quality Inspection This step involves multi-dimensional quality inspection, which is completed collaboratively through "rule verification + large model semantic verification".

[0049] 1. Definition of Validation Dimensions Automated quality inspection should include at least the following three dimensions: (1) Completeness of the question statement: Are there any missing conditions, ambiguities, or incomplete information in the question stem? (2) Logical consistency of the reasoning chain: Are there any contradictions, skips, or circular arguments in the CoT? (3) Accuracy of facts in the content: whether the conclusion is consistent with the standard answer or reference document.

[0050] 2. Verification prompt word process based on a large language model During the quality control phase, the following structured information is input into the large model: - Original question - Standard answer - CoT text and conclusion generated in step S3. Fixed prompts are used to instruct the model to output only "Correct / Incorrect" to avoid interpretive noise.

[0051] 3. Differences between objective questions and short-answer questions For objective questions: focus on comparing whether the final conclusion matches the answer; Question and Answer: Focus on verifying whether the reasoning is entirely derived from the content of the ref document.

[0052] (1) Quality inspection of objective questions • Core Rules: Does the final conclusion of CoT match the answer? For non-multiple-choice questions, are there multiple answers? Does self-contradictory reasoning exist? If any of these items appear, output an error; if none of them appear, output a correct answer. (2) Quality inspection of question and answer questions • Comparative validation using a large model: Verify whether the reasoning process is entirely derived from the content of the reference document or basis, and whether the standard answer and the corresponding conclusion are similar. If both are true, output "correct"; otherwise, output "incorrect".

[0053] Quality control output rules: - Only output "Correct" or "Incorrect" - No explanatory text is allowed. Only retain the CoT text and conclusions that are correct in the quality inspection. For those that fail the quality inspection, the corresponding steps in step S5 can be executed to regenerate them.

[0054] S5, Confidence Assessment and Dynamic Iteration 1. Confidence Score Components The confidence score is calculated by weighting multiple sub-scores, for example: Logical consistency score (L): Whether the reasoning is consistent Answer consistency score (A): Whether the conclusion matches the standard answer. Fact accuracy score (F): Whether there are factual errors The confidence score calculation formula is: , In the formula, , , Configurable weights; 2. Determination of Pass / Fail Threshold Set a minimum confidence threshold (e.g., 0.8); Corpus data below the threshold is entered into the regeneration process.

[0055] 3. Dynamic prompt word adjustment mechanism If the corpus fails the quality check or does not meet the confidence threshold, it will enter the regeneration process. During the regeneration process, the prompt words will be automatically adjusted according to the failure type. • Logical issues: Strengthen constraints such as "step-by-step analysis" and "no skipping steps"; • Factual issue: Strengthen the "based on reference documents only" restriction; • Expression redundancy: Limit CoT length or compress intermediate steps.

[0056] The adjusted prompt words re-trigger CoT generation until the confidence threshold is reached or manual intervention is required.

[0057] • Quantitative scoring of quality inspection results • Set configurable thresholds (such as pass rate, consistency score). • Samples that do not meet the standards will be automatically: Adjust prompt word constraints Regenerate CoT Re-entering the quality inspection process This mechanism effectively reduces the randomness and error rate caused by one-time generation.

[0058] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0059] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0060] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0064] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for generating a general domain-specific thought chain corpus, characterized in that, Includes the following steps: We acquire general domain corpora of different task types, and convert them into a unified standard format through structured processing to obtain standardized corpora. The task types include subjective and objective categories. Based on standardized corpora of different task types, initial thought chain text and conclusions are generated according to a common matching reasoning link; Based on the initial thought chain text and conclusion, a multi-dimensional automated quality inspection mechanism is used to verify them, resulting in the final thought chain text and conclusion.

2. The method for generating a general domain-specific thought chain corpus according to claim 1, characterized in that, The structured corpus includes core fields, including: qid: Represents a unique identifier for the question; qtype: indicates the question type, which includes multiple choice, true / false, and short answer questions. Multiple choice, true / false, and short answer questions are all objective questions, while short answer questions are subjective questions. question: indicates the question text; choices: indicates options; answer: indicates the standard answer; explanation: indicates standard parsing; ref: Reference document or basis.

3. The method for generating a general domain-specific thought chain corpus according to claim 2, characterized in that, The steps for generating the initial thought chain text include: Based on the qtype, the task type and question type of the question are determined. Then, under the corresponding question type, the question text is combined to determine whether there is one or more value reasoning types among multi-step judgment, rule matching, and condition constraints. If they exist, the corresponding value reasoning needs to be executed during the answering process. If none of them exist, the answer is given directly. When none of them exist and the question can be answered directly from memory, the initial thought chain text of long text is not generated. Based on the corresponding value reasoning type, extract the core elements from the question text, including core conditions, limiting words, and implicit rules; Based on the core elements and the existing value reasoning types, the corresponding options or possible conclusions are verified one by one. After verification, an initial thought chain text is generated according to the set logic string. The corresponding options are those given in the objective category, and the possible conclusions are derived from the subjective category based on the question orientation of the question text, the standard analysis, and / or the reference documents.

4. The method for generating a general domain-specific thought chain corpus according to claim 3, characterized in that, The term "multi-step judgment" refers to a question requiring two or more steps of deduction and / or calculation and / or comparison and / or elimination before an answer can be given. The term "rule matching" refers to matching information in the question text with rules and / or definitions and / or clauses. The term "conditional constraint" refers to the presence of constraint terms or boundary conditions in the question text. The process of performing corresponding value reasoning during the response includes: When there are multiple steps involved, the intermediate steps are displayed and backtracking is ensured during the answering process. If there are no multiple steps involved, the answer can be given directly or by providing evidence and conclusions, generating a lightweight initial thought chain text. When a rule match exists, the answer should be provided by citing the rule points from the standard parsing and / or parameter document or by the basis and / or known knowledge points, and explaining the reason for the match. If no rule match exists, the rule points should not be forcibly introduced, and the answer should be provided according to common sense reasoning or calculation path. When constraints exist, during the answering process, the constraints and ranges are enumerated based on the constraint terms or boundary conditions, and then the corresponding options or possible conclusions are verified one by one. If there are no constraints, there is no need to enumerate them.

5. The method for generating a general domain-specific thought chain corpus according to claim 3, characterized in that, For objective categories, the steps for verifying each option one by one include: Semantic standardization is performed on all corresponding options to transcribe each corresponding option into a complete judgment statement, in which only the decidable facts or rule points are retained; The core elements in the question text are matched with each complete judgment statement one by one, and answers are given according to the existing value reasoning type to obtain the judgment result of the exclusion or retention of each complete judgment statement. Set prompt words, which indicate the generation of the initial thought chain text for the objective category based on the logical sequence of core elements in the question text → corresponding option judgment → elimination / retention → final conclusion.

6. The method for generating a general domain-specific thought chain corpus according to claim 3, characterized in that, For the subjective category, the steps for verifying each possible conclusion include: Based on the question text, identify the problem, analyze the requirements based on the problem, consult the reference documents or evidence, and answer according to the existing value reasoning type; Set prompt words, which indicate the generation of the initial thought chain text under the subjective category based on the logical sequence of problem confirmation → request analysis → document / basis query → response strategy.

7. The method for generating a general domain-specific thought chain corpus according to claim 3, characterized in that, For the everyday dialogue format in the subjective category, answers can be given directly without generating initial thought chain text. If generation is required, only minimal analysis will be retained, and deep reasoning will not be entered.

8. The method for generating a general domain-specific thought chain corpus according to claim 2, characterized in that, The steps to obtain the final thought chain text and conclusion include: Define the dimensions for automatic quality inspection, including: completeness of issue representation, consistency of reasoning logic, and accuracy of factual content. The question text, the corresponding standard answer, the initial thought chain text, and the corresponding conclusion are input into the large language model, and validated under the automatic quality inspection dimension in conjunction with the prompt words. The validation result is output as either "correct" or "incorrect." During the large language model validation process... For objective questions, check if any of the following conditions are met: inconsistency between the standard answer and the corresponding conclusion, multiple answers for non-multiple-choice questions, or self-contradictory reasoning. If none of these conditions are met, output "correct". For subjective cases, verify whether the reasoning process is entirely derived from the content of the reference document or basis, and whether the standard answer is similar to the corresponding conclusion. If both are true, output "correct"; otherwise, output "incorrect". All initial thought chain texts and conclusions that are determined to be correct are taken as the final thought chain texts and conclusions.

9. A method for generating a general domain-specific thought chain corpus according to claim 5 or 6, characterized in that, It also includes confidence assessment and dynamic iteration steps to score the confidence of the generated thought chain text and dynamically adjust the prompt words, wherein the confidence scoring step includes: For the aforementioned thought chain text and conclusion, a confidence score is calculated, where the calculation expression is: ， In the formula, , , For configurable weights, Scoring for logical consistency To score for consistency of conclusions, Score for the accuracy of the facts; Determine whether the confidence score is lower than the set minimum confidence threshold. If so, regenerate according to the general matching inference link; otherwise, no regeneration is required. The steps for dynamically adjusting the prompt words include: For initial thought chain texts and conclusions that fail validation, and thought chain texts and conclusions whose confidence scores do not reach the minimum confidence threshold, the prompt words are automatically adjusted according to their failure type: When the failure type is a logical reasoning problem, add constraint statements, including step-by-step analysis and no skipping steps, as prompts to constrain the generation process; When the failure type is a factual problem, add constraint phrases, including those based solely on reference documents or evidence, as prompts to constrain the generation process; When the failure type is expression redundancy, constraint words, including limiting the length of the thought chain text and compressing intermediate steps, are added as prompts to constrain the generation process.

10. A storage medium having a program stored thereon, characterized in that, When the program is executed, it implements a general domain thinking chain corpus generation method as described in any one of claims 1-9.

Images