A method for generating synthetic question pairs based on DAG chain iteration reasoning

Abstract

The application discloses a method for generating a synthetic question pair based on DAG chain iteration reasoning, which comprises the following steps: generating an initial question according to input text or a knowledge base; evaluating the initial question to determine whether it meets the preset quality standard and obtaining an evaluation result; improving the initial question according to the evaluation result; generating an answer according to the initial question; decomposing the answer into a series of sub-propositions or assertions; verifying each sub-proposition or assertion; integrating all verification results to obtain a comprehensive evaluation of each assertion; if the initial answer is unreliable, modifying the answer according to the verification result and repeating the verification step until the answer is verified to be reliable; iterating until the question and the answer both meet the preset quality standard or reach the preset iteration number, forming a directed acyclic graph structure. The method effectively identifies and corrects errors in the LLM generation process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more particularly to a method for generating synthetic problem pairs based on chain-based iterative reasoning using DAG. Background Technology

[0002] Existing question-and-answer pair generation methods have many limitations, making it difficult to generate high-quality, reliable question-and-answer pairs. On the one hand, the quality of the generated questions and answers varies, lacking logical coherence and diversity; on the other hand, the accuracy and reliability of the answers are difficult to guarantee, easily leading to erroneous or contradictory results; more importantly, these methods cannot simulate human logical reasoning and multiple verification processes, making it difficult to generate high-quality question-and-answer pairs with deep thinking.

[0003] Currently available technical solutions include:

[0004] 1. Manually written question-and-answer pairing method

[0005] This method relies on human editors to write questions and answers, which ensures the quality of question-and-answer pairs, but it is costly and inefficient. Furthermore, due to limitations in the knowledge and perspectives of the editors, the diversity and innovation of the generated question-and-answer pairs are often limited. In addition, human writing struggles to simulate complex logical reasoning processes, making it impossible to generate high-quality question-and-answer pairs with deep thinking.

[0006] 2. Rule-based question-answering pair generation method

[0007] This method automatically generates questions and answers based on a set of predefined grammatical and semantic rules. While the question-and-answer pairs generated by this method are well-structured and standardized, they lack diversity, struggle to cover all scenarios, and cannot simulate complex logical reasoning processes.

[0008] 3. Retrieval-based question-answer pair generation method

[0009] This method retrieves relevant questions and answers from existing large-scale corpora, then assembles and rewrites them to generate new question-answer pairs. While this method can utilize existing data, it is limited by the scope of the corpus, resulting in limited diversity and originality in the generated question-answer pairs. Furthermore, the retrieval and assembly process struggles to guarantee the accuracy and consistency of the answers.

[0010] 4. LLM-based question-answering pair generation method

[0011] This method uses a large model (LLM) to directly generate questions and answers, producing diverse outputs. However, due to the lack of a clear reasoning process and verification mechanism, the quality of the generated question-and-answer pairs cannot be guaranteed.

[0012] Problems with existing technology:

[0013] 1. Rule-based retrieval methods lack diversity and innovation.

[0014] Traditional rule-based and retrieval-based question-answering generation methods typically rely on predefined rule templates and knowledge base queries, resulting in a lack of diversity and innovation in the generated question-answer pairs. This approach struggles to cover complex scenarios and knowledge points, fails to simulate deep human thinking and reasoning processes, and produces question-answer pairs of limited quality, thus failing to meet the ever-growing demands.

[0015] 2. Large model generation has low accuracy and reliability.

[0016] While large-model-based question-answering generation methods can produce diverse outputs, the accuracy and reliability of the generated questions and answers are low due to the lack of clear reasoning paths and verification steps.

[0017] 3. Manual coding is costly and has significant limitations.

[0018] While manually generating question-and-answer pairs can ensure a certain level of quality, it is costly and inefficient, making large-scale production impossible to meet ever-growing demand. Furthermore, limited by the knowledge and perspectives of the editors, manually generated question-and-answer pairs exhibit limited diversity and innovation, failing to cover a broad range of domain knowledge.

[0019] 4. Lack of genuine complex reasoning ability

[0020] The inability to truly simulate human logical thinking and complex reasoning processes, and the inability to automatically generate high-quality question-and-answer pairs with deep thinking and progressive layers, severely restricts the application and development of question-and-answer systems in various fields. Summary of the Invention

[0021] In view of the above problems, the present invention is proposed to provide a method for generating synthetic problem pairs based on chain iterative reasoning using DAG to overcome or at least partially solve the above problems.

[0022] According to one aspect of the present invention, a method for generating composition problem pairs based on chain-based iterative reasoning using DAG is provided, the method comprising:

[0023] Generate initial questions based on the input text or knowledge base;

[0024] The initial problem is evaluated to determine whether it meets the preset quality standards, and the evaluation result is obtained.

[0025] The initial problem was improved based on the evaluation results;

[0026] Generate an answer based on the initial question;

[0027] The answer is broken down into a series of sub-propositions or assertions;

[0028] Verify each subproposition or assertion;

[0029] By integrating all the verification results, a comprehensive evaluation of each assertion can be obtained;

[0030] If the initial answer is unreliable, modify the answer based on the verification results and repeat the verification steps until the answer is verified as reliable.

[0031] The process is iterated until both the question and the answer meet the preset quality standards or the preset number of iterations, forming a directed acyclic graph structure and recording the verification process.

[0032] Optionally, generating the initial question based on the input text or knowledge base specifically includes:

[0033] Input text corpus or knowledge base D

[0034] Using a Large Language Model (LLM), an initial question set Q = {q1, q2, ..., q} is generated based on the input. n}

[0035] Formula expression:

[0036] Q = LLM(D, P) question )

[0037] Among them, P question It is a prompt template used to guide large language model generation problems.

[0038] Optionally, evaluating the initial problem to determine whether it meets preset quality standards and obtaining evaluation results specifically includes:

[0039] The initial problem set Q = {q1, q2, ..., q} n};

[0040] Define the set of evaluation criteria C = {c1, c2, ..., c...} m}, each c i This represents an evaluation standard;

[0041] For each question q i Calculate the score s on each criterion. ij =Eval(q) i c j );

[0042] Calculate the overall score for each question Where w h It is standard C j The weights are used to obtain the evaluation result E for each problem. i=Assess(S i ,θ), where θ is a preset threshold;

[0043] The set of evaluation criteria includes: clarity, difficulty, relevance, importance, answerability, originality, and value.

[0044] Optionally, improving the initial problem based on the evaluation results specifically includes:

[0045] Question q i and its evaluation results E_i, scores of each dimension {s i1 s i2 , ..., s im};

[0046] Based on the evaluation results, determine the direction for improvement (D). i =Improve_Direction(q i ,{s i1 ,s i2 ,…,s im});

[0047] Optimize the problem based on the direction of improvement. i =LLM(q i ,P improve D i ).

[0048] Optionally, generating an answer based on the initial question specifically includes:

[0049] High-quality question q i ;

[0050] Generate the initial answer 'a' using a large language model. i =LLM(q i ,P answer ).

[0051] Optionally, generating an answer based on the initial question further includes:

[0052] Initial answer a i This includes: decomposing the answer into a series of subpropositions or assertions P = {p1, p2, ..., p...} k}, where each p j It represents an assertion that can be verified independently;

[0053] For each assertion p j Extract its key entities and relationships E j =Extract_Entities(p j ),R j =Extract_Relations(pj )

[0054] Formula expression:

[0055] P = Decompose(a i ,P decompose )

[0056] Among them, P decompose It is a prompting module used to guide the large language model in decomposing the answer.

[0057] Optionally, the verification of each subproposition or assertion specifically includes:

[0058] Each sub-proposition or assertion is validated using methods including external knowledge base queries, websites, logical reasoning, and cross-validation.

[0059] Each assertion in the answer is verified using multiple verification methods.

[0060] Optionally, the process of integrating all verification results to obtain a comprehensive evaluation of each assertion specifically includes:

[0061] Input: assertion p j Scores on each verification method {s kb ,s web ,s logic ,s cross};

[0062] Calculate the weighted composite score S j =w kb ·s kb +w web ·s web +w logic ·s logic +w cross ·s cross , where w kb ,w web ,w logic ,w cross These are the weights of each verification method;

[0063] The reliability R of the assertion is judged based on the overall score. j =Reliability(S j θ r ), where θ is a preset reliability threshold;

[0064] Formula expression:

[0065] Optionally, the step of modifying the answer based on the verification result and repeating the verification steps if the initial answer is unreliable specifically includes:

[0066] Enter answer a i Reliability assessment for each assertion {R1, R2, ..., R...} k}, verification information set {I kb ,I web ,I logic ,I cross}

[0067] Identify the set of unreliable assertions P unreliable ={p j |R j =0}

[0068] Generate revised assertions based on the verification information.

[0069] p′ j =LLM(p j , {I kb I web ,I logic ,I cross},P improve_assertion )

[0070] Update answer a' i =Replace(a i ,P unreliable {p'1, p'2, ...})

[0071] Formula: a' i =LLM(a i ,P unreliable ,{p'1,p'2,…},P improve_answer )

[0072] Among them, P improve_assertion and P improve_answer It is a prompt template used to guide large language models to improve assertions and answers.

[0073] Optionally, the iterative loop continues until both the question and the answer meet a preset quality standard or a preset number of iterations, forming a directed acyclic graph structure, and the verification process is recorded. Specifically, this includes:

[0074] Repeat the steps until the termination condition is met;

[0075] Termination conditions include: both the question and the answer meet the preset quality standards, or the preset maximum number of iterations Max_Iter is reached;

[0076] During the iteration process, based on the verification results and improvement strategies of the previous round, the quality of questions and answers is continuously optimized, while the entire verification and reasoning process is recorded to form a directed acyclic graph structure.

[0077] This invention provides a method for generating synthetic problem pairs based on chain-like iterative reasoning using a Directed Acyclic Graph (DAG). The method includes: generating an initial question based on input text or a knowledge base; evaluating the initial question to determine if it meets a preset quality standard and obtaining an evaluation result; improving the initial question based on the evaluation result; generating an answer based on the initial question; decomposing the answer into a series of sub-propositions or assertions; verifying each sub-proposition or assertion; integrating all verification results to obtain a comprehensive evaluation of each assertion; if the initial answer is unreliable, modifying the answer based on the verification result and repeating the verification steps until the answer is verified as reliable; iterating until both the question and answer meet the preset quality standard or reach a preset number of iterations, forming a DAG structure, and recording the verification process. This method effectively identifies and corrects errors in the LLM generation process and reduces the occurrence of illusions.

[0078] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0079] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0080] Figure 1 The flowchart illustrates a method for generating synthetic problem pairs based on chain-based iterative reasoning using DAG, as provided in an embodiment of the present invention. Detailed Implementation

[0081] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0082] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0083] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0084] like Figure 1 As shown, a chain-based iterative reasoning method for generating synthetic question pairs based on DAG is disclosed. The method includes: generating an initial question based on input text or a knowledge base; evaluating the initial question to determine if it meets a preset quality standard and obtaining an evaluation result; improving the initial question based on the evaluation result; generating an answer based on the initial question; decomposing the answer into a series of sub-propositions or assertions; verifying each sub-proposition or assertion; integrating all verification results to obtain a comprehensive evaluation of each assertion; if the initial answer is unreliable, modifying the answer based on the verification result and repeating the verification steps until the answer is verified as reliable; iterating until both the question and the answer meet the preset quality standard or reach a preset number of iterations, forming a directed acyclic graph structure, and recording the verification process.

[0085] (1) Initial question generation: Generate an initial question based on the input text or knowledge base.

[0086] Algorithm description:

[0087] Input: Text corpus or knowledge base D

[0088] Using a Large Language Model (LLM), an initial question set Q = {q1, q2, ..., q} is generated based on the input. n}

[0089] Formula expression:

[0090] Q = LLM(D,P) question )

[0091] Among them, P question It is a prompt template used to guide large language model generation problems.

[0092] (2) Problem assessment: Assess the problem and determine whether it meets the preset quality standards;

[0093] Algorithm description:

[0094] Input: Initial problem set Q = {q1, q2, ..., q} n};

[0095] Define the set of evaluation criteria C = {c1, c2, ..., c...} m}, each c i This represents an evaluation standard;

[0096] For each question q i Calculate its score s on each criterion. ij =Eval(q) i c j);

[0097] Calculate the overall score for each question Where w j It is standard C j The weights;

[0098] Obtain the evaluation result E for each problem. i =Assess(S i ,θ), where θ is a preset threshold.

[0099] The evaluation criteria set C includes the following dimensions: Clarity: Is the question clearly and unambiguously stated, avoiding ambiguity? Difficulty: Does the difficulty of the question meet expectations? Relevance: Is the question closely related to the text content? Importance: Does the question involve important knowledge points in the text? Answerability: Does the question have a clear answer? Novelty: Is the question innovative, avoiding simple repetition of the text content? Value: Does the question have educational or practical value?

[0100] (3) Problem Improvement: Improve the problem based on the evaluation results;

[0101] Algorithm description:

[0102] Input: Question q i and its evaluation results E i Scores for each dimension {s i1 s i2 ,…,s im}

[0103] Based on the evaluation results, determine the direction for improvement (D). i =Improve_Direction(q i ,{s i1 ,s i2 ,…,s im})

[0104] Optimize the problem based on the improvement direction.

[0105] Formula expression:

[0106] q′ i =LLM(q i ,P improve D i )

[0107] P improve ="""Please improve the following issues to make them more {improvement direction}: ""

[0108] Original question: {q i}

[0109] Rating results:

[0110] {Rating Dimension 1}: {Score 1}

[0111] {Rating Dimension 2}: {Score 2} ...

[0113] Areas for improvement: {D i}

[0114] The improved issues should be:

[0115] Maintain the core intent of the original question;

[0116] Improve performance on {low-scoring dimensions};

[0117] ...""".

[0118] (4) Answer generation: Generate an answer based on the question;

[0119] Input: High-quality question q i ;

[0120] Generate the initial answer 'a' using a large language model. i =LLM(q i ,P answer ).

[0121] The formula represents: a i =LLM(q i ,P answer )

[0122] (5) Answer decomposition: Decompose the answer into a series of sub-propositions or assertions;

[0123] Answer decomposition is a preliminary step in chain verification, breaking down the initial answer into a series of sub-propositions or assertions to facilitate refined verification.

[0124] Input: Initial answer a i

[0125] Decompose the answer into a series of subpropositions or assertions P = {p1, p2, ..., p...} k}, where each p j Represents an assertion that can be independently verified.

[0126] For each assertion p j Extract its key entities and relationships E j =Extract_Entities(p j ),R j =Extract_Relations(p j )

[0127] Formula expression:

[0128] P = Decompose(a i P decompose )

[0129] Among them, P decompose It is a prompting module used to guide large language models in decomposing answers.

[0130] (6) Chain verification: Verify each sub-proposition or assertion. Verification methods include external knowledge base queries, websites, logical reasoning, cross-validation, etc.

[0131] Chain verification is the core of this invention. It uses multiple verification methods to verify each assertion in the answer, ensuring the accuracy and reliability of the answer.

[0132] Algorithm description:

[0133] For each assertion p j The following methods were used for verification:

[0134] 6.1 External Knowledge Base Query Verification

[0135] Input: assertion p j Key entity set E j Relationship R j

[0136] Construct a knowledge base query statement q kb =Construct_Query(p j E j ,R j )

[0137] Retrieve relevant information from the knowledge base kb =Query_Knowledge_Base(q kb )

[0138] Calculate the consistency score between the assertion and the knowledge base information. kb =Consistency(p j ,I kb )

[0139] Formula expression:

[0140] s kb =Consistency(p j ,I kb Similarity can be calculated using methods such as cosine similarity.

[0141] 6.2 Website Content Verification

[0142] Input: assertion p j

[0143] Construct search query q web =Construct_Search_Query(p j )

[0144] Obtain relevant information from the website. web =Search_Web(q web )

[0145] Calculate the consistency score between the assertion and the website information. web =Consistency(p j ,I web )

[0146] 6.3 Logical Reasoning Verification

[0147] Input: assertion p j The set of verified assertions P verified

[0148] Constructing a logical reasoning task t logic =Construct_Logic_Task(p j ,P verified )

[0149] Conclusion c is obtained through logical reasoning. logic =Logic_Reasoning(t logic )

[0150] Calculate the consistency score s between the assertion and the logical reasoning conclusion. logic =Consistency(p j c logic ).

[0151] 6.4 Cross-validation

[0152] Input: assertion p j Other related assertion set P related

[0153] Analyze the consistency between assertions c cross =Cross_Check(p j P related )

[0154] Calculate the cross-validation score s cross =Evaluate(c cross )

[0155] Formula expression:

[0156] Cross-validation score s cross Represented as:

[0157]

[0158] (7) Results integration: Integrate all verification results to determine the reliability of the initial answer;

[0159] The results integration phase combines the results of various verification methods to arrive at a comprehensive evaluation of each assertion.

[0160] Input: assertion p j Scores on each verification method {s kb s web s logic s cross}

[0161] Calculate the weighted composite score S j =w kb ·s kb +w web ·s web +w logic ·s logic +w cross ·s cross , where w kb w web ,w logic ,w cross The weights of each verification method

[0162] The reliability R of the assertion is judged based on the overall score. j =Reliability(S j θ r ), where θ r It is a preset reliability threshold.

[0163] Formula expression:

[0164] (8) Answer improvement: If the initial answer is unreliable, modify the answer based on the verification results and repeat steps (5)-(7) until the answer is verified to be reliable;

[0165] When an answer contains unreliable assertions, the answer needs to be improved.

[0166] Input: Answer a i The reliability assessment of each assertion {R1, R2, ..., R} k}, verification information set {I kb ,I web ,I logic ,I cross}

[0167] Identify the set of unreliable assertions P unreliable ={pj |R j =0}

[0168] Generate a revised assertion a' based on the verification information. i =Replace(a i ,P unreliable ,{p'1,p'2,…})

[0169] Update answer a' i =Replace(a i P unreliable ({p'1, p'2, ...})

[0170] Formula expression:

[0171] a′ i =LLM(a i ,P unreliable ,{p′1,p′2,…},P improve_answer )

[0172] Among them, P improve_assertion and P improve_answer It is a prompt template used to guide large language models to improve assertions and answers.

[0173] (9) Iteration loop: Repeat steps (2) to (8) until the questions and answers meet the preset quality standards or the preset number of iterations, forming a directed acyclic graph structure and recording the verification process;

[0174] Repeat steps 2-8 until the termination condition is met.

[0175] Termination condition: Both the question and the answer meet the preset quality standard, or the preset maximum number of iterations Max_Iter is reached;

[0176] During the iteration process, the system continuously optimizes the quality of questions and answers based on the verification results and improvement strategies of the previous round, while recording the entire verification and reasoning process to form a directed acyclic graph structure.

[0177] Generate diverse and innovative high-quality question-and-answer pairs, covering a wide range of knowledge domains and complex scenarios, without being limited by rule templates and knowledge bases.

[0178] This invention provides a model based on large-scale model generation capabilities, which, through training on a large-scale corpus, endows the model with the ability to generate diverse and highly innovative question-answer pairs. Unlike traditional rule-based and query-based methods, this model does not rely on predefined templates and knowledge bases, and can autonomously generate rich and diverse question-answer content covering various domains, thus handling complex scenarios.

[0179] The generated question-and-answer pairs are guaranteed to be accurate and reliable, with good logical coherence, and truly simulate the human reasoning process.

[0180] This invention provides a chain-based verification-guided iterative reasoning mechanism based on DAG (Directed Acyclic Graph). The generated initial question-answer pairs are decomposed into a series of interrelated reasoning steps, and external knowledge sources are introduced in each step for verification, supplementation, and optimization. Finally, accurate, reliable, and logically coherent high-quality question-answer pairs are synthesized, truly simulating the complex human thought process of progressively deepening and mutually verifying each other.

[0181] By incorporating complex logical reasoning into the question-and-answer pair generation process, high-quality question-and-answer pairs with deep thinking and progressive layers are automatically generated, breaking through the bottlenecks of existing technologies.

[0182] This invention integrates complex logical reasoning capabilities into the question-and-answer pair generation process. By decomposing the reasoning path using a DAG structure and introducing external knowledge sources at each step, it achieves a progressive derivation of complex knowledge, ultimately automatically synthesizing high-quality question-and-answer pairs with deep thinking and rigorous logic. This breakthrough overcomes the bottlenecks of existing technologies and opens up new avenues for the development of question-and-answer systems.

[0183] Beneficial effects:

[0184] It can simulate the human logical thinking process and generate high-quality, highly reliable question-and-answer pairs.

[0185] By constructing a DAG to explicitly represent the reasoning process and verifying it after each step of reasoning, the high credibility of the answer can be ensured.

[0186] By employing an iterative reasoning approach, diverse and innovative question-and-answer pairs can be generated, covering more scenarios.

[0187] The entire process is automated, reducing human intervention and thus significantly lowering costs and increasing efficiency.

[0188] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating composition problem pairs based on chain-based iterative reasoning using DAG, characterized in that, The methods for generating and synthesizing problems include: Generate initial questions based on the input text or knowledge base; The initial problem is evaluated to determine whether it meets the preset quality standards, and the evaluation result is obtained. The initial problem was improved based on the evaluation results; Generate an answer based on the initial question; The answer is broken down into a series of sub-propositions or assertions; Verify each subproposition or assertion; By integrating all the verification results, a comprehensive evaluation of each assertion can be obtained; If the initial answer is unreliable, modify the answer based on the verification results and repeat the verification steps until the answer is verified as reliable. The process is iterated until both the question and the answer meet the preset quality standards or the preset number of iterations, forming a directed acyclic graph structure and recording the verification process.

2. The method for generating composition problem pairs based on chain-based iterative reasoning using DAG according to claim 1, characterized in that, The process of generating an initial question based on the input text or knowledge base specifically includes: Input: Text corpus or knowledge base D Using a Large Language Model (LLM), an initial question set Q = {q1, q2, ..., q} is generated based on the input. n } Formula expression: Q=LLM(D,P question ) Among them, P question It is a prompt template used to guide large language model generation problems.

3. The method for generating composition problem pairs based on chain-based iterative reasoning using DAG according to claim 1, characterized in that, The process of evaluating the initial problem to determine whether it meets the preset quality standards and obtaining the evaluation result specifically includes: The initial problem set Q = {q1, q2, ..., q} n }; Define the set of evaluation criteria C = {c1, c2, ..., c...} m }, each c i This represents an evaluation standard; For each question q i Calculate s for each standard ij =Eval(q) i c j ); Calculate the overall score for each question Where w j It is standard C j The weights are used to obtain the evaluation result E for each problem. i =Assess(S i ,θ), where θ is a preset threshold; The set of evaluation criteria includes: clarity, difficulty, relevance, importance, answerability, originality, and value.

4. The method for generating composition problem pairs based on chain-based iterative reasoning using DAG according to claim 1, characterized in that, The improvement of the initial problem based on the evaluation results specifically includes: Question q i and its evaluation results E i Scores for each dimension {s i1 s i2 , ..., s im }; Based on the evaluation results, determine the direction for improvement (D). i =Improve_Direction(q i , {s i1 s i2 , ..., s im }); Optimize the problem based on the direction of improvement. i =LLM(q i P improve D i ).

5. The method for generating composition problem pairs based on chain-based iterative reasoning using DAG according to claim 1, characterized in that, The process of generating an answer based on the initial question specifically includes: High-quality question q i ; Generate the initial answer 'a' using a large language model. i =LLM(q i P answer ).

6. The method for generating composition problem pairs based on chain-based iterative reasoning using DAG according to claim 1, characterized in that, The process of generating an answer based on the initial question also includes: Initial answer a i This includes: decomposing the answer into a series of subpropositions or assertions P = {p1, p2, ..., p...} k }, where each p j It represents an assertion that can be verified independently; For each assertion p j Extract its key entities and relationships E j =Extract_Entities(p j ),R j =Extract_Relations(p j ) Formula expression: P=Decompose(a i ,P decompose ) Among them, P decompose It is a prompting module used to guide the large language model in decomposing the answer.

7. The method for generating composition problem pairs based on chain-based iterative reasoning using DAG according to claim 1, characterized in that, The verification of each subproposition or assertion specifically includes: Each sub-proposition or assertion is validated using methods including external knowledge base queries, websites, logical reasoning, and cross-validation. Each assertion in the answer is verified using multiple verification methods.

8. The method for generating composition problem pairs based on chain-based iterative reasoning using DAG according to claim 1, characterized in that, The process of integrating all verification results to obtain a comprehensive evaluation of each assertion specifically includes: Input: assertion p j Scores on each verification method {s kb ,s web ,s logic ,s cross }; Calculate the weighted composite score S j =w kb ·s kb +w web ·s web +w logic ·s logic +w cross ·s cross , where w kb ,w web ,w logic ,w cross These are the weights of each verification method; The reliability R of the assertion is judged based on the overall score. j =Reliability(S j ,θ r ), where θ is a preset reliability threshold; Formula expression:

9. The method for generating composition problem pairs based on chain-based iterative reasoning using DAG according to claim 1, characterized in that, If the initial answer is unreliable, the answer is modified based on the verification result, and the verification steps are repeated until the answer is verified as reliable. Specifically, this includes: Enter answer a i The reliability assessment of each assertion {R1, R2, ..., R} k }, verification information set {I kb ,I web ,I logic ,I cross } Identify the set of unreliable assertions P unreliable ={p j |R j =0} Generate revised assertions based on the verification information. p′ j =LLM(p j ,{I kb ,I web ,I logic ,I cross },P improve_assertion ) Update answer a' i =Replace(a i ,P unreliable ,{p'1,p'2,…}) Formula: a' o =LLM(a i ,P unreliable ,{p'1,p'2,…},P improve_answer ) Where P improve_assertion and P improve_answer It is a prompt template used to guide large language models to improve assertions and answers.

10. The method for generating composition problem pairs based on chain-based iterative reasoning using DAG according to claim 1, characterized in that, The iterative cycle continues until both the question and the answer meet the preset quality standard or the preset number of iterations, forming a directed acyclic graph structure. The verification process is recorded, specifically including: Repeat the steps until the termination condition is met; Termination conditions include: both the question and the answer meet the preset quality standards, or the preset maximum number of iterations Max_Iter is reached; During the iteration process, based on the verification results and improvement strategies of the previous round, the quality of questions and answers is continuously optimized, while the entire verification and reasoning process is recorded to form a directed acyclic graph structure.

Images