A method and system for constructing thinking chain data from a dissertation

Abstract

The present application belongs to the field of artificial intelligence and natural language processing, and specifically relates to a method and system for constructing thinking chain data from a dissertation, aiming to solve the technical problems of high cost, low efficiency, unreliable quality and lack of traceable basis when constructing thinking chain data by artificial construction or directly generating thinking chain data by a large model. The present application includes: analyzing the title, table of contents and abstract of the dissertation to drive a large language model to condense key scientific problems; generating a core reasoning framework based on the problems and chapter directory, and extracting detailed argumentation content from the corresponding chapters step by step to integrate into a preliminary reasoning chain; using a long window large language model to compare and verify the reasoning chain and the full text of the dissertation and filter. The present application realizes the automatic and high-fidelity extraction of high-quality thinking chain data from the original dissertation, ensuring the logical rigor and traceability of the data.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and natural language processing, and specifically relates to a method and system for constructing thought chain data from academic theses. Background Technology

[0002] With the rapid development of large language models, their performance in complex reasoning tasks has become a key indicator for measuring their capabilities. Chain-of-Thought, as a data format that simulates the step-by-step reasoning process of humans, has demonstrated immense value in training and guiding models to engage in deep logical thinking and improving problem-solving abilities. Therefore, acquiring large-scale, high-quality chain-of-Thought data is crucial for driving the technological advancement of large language models.

[0003] Currently, the construction of thought chain data mainly relies on two approaches. The first is manual compilation by domain experts. This method ensures data accuracy and logical rigor, but it is costly, time-consuming, and highly dependent on expert resources, making large-scale production difficult. The second approach utilizes large language models for generation; given a question and answer, the model completes the reasoning process. While this method can quickly generate large amounts of data, its quality is entirely limited by the knowledge boundaries and reasoning capabilities of the generating model, easily leading to factual errors or logical fallacies—the so-called "illusion" phenomenon—and the generated content lacks verifiable evidence.

[0004] Academic literature, especially rigorously structured and well-argued doctoral or master's theses, is itself a record of a systematic research process centered around one or more core scientific questions, naturally containing a complete logical chain from question formulation and analysis to solution. However, current technology lacks an effective solution for automatically and faithfully extracting and transforming this implicit reasoning logic hidden within lengthy, unstructured texts into structured thought chain data. Therefore, how to efficiently and reliably utilize massive amounts of dissertation resources to address the insufficient production capacity of high-quality thought chain data is a pressing challenge in the current technological field. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, namely the high cost, low efficiency, unreliable quality, and lack of traceability associated with generating thought chain data through manual construction or direct generation of large models, this invention provides a method and system for constructing thought chain data from dissertations.

[0006] In a first aspect, this invention proposes a method for constructing thought chain data from dissertations, comprising:

[0007] The document of the thesis is parsed to extract the thesis title, table of contents, and abstract;

[0008] The paper title, chapter list, and abstract are input into the first language model to generate text containing key scientific questions, and the key scientific questions are identified from the text.

[0009] The key scientific questions and chapter list are input into the second language model to generate a core reasoning framework, which includes multiple reasoning steps and the corresponding chapters of the paper for each reasoning step.

[0010] For each reasoning step in the core reasoning framework, the text content of the corresponding thesis chapter is extracted from the thesis, and the text content and the key scientific question are input into the third language model to generate detailed content for the reasoning step. The detailed content of all reasoning steps is then integrated into a preliminary reasoning logic chain.

[0011] A long-window, large-language model is used to compare and verify the preliminary reasoning logic chain with the full text of the thesis. Based on the verification results, the final reasoning logic chain is output.

[0012] Furthermore, at least two of the first, second, and third large language models are the same large language model.

[0013] Furthermore, the core reasoning framework includes a mapping relationship between each reasoning step and at least one chapter of the paper; the method also includes:

[0014] When generating the detailed content of the reasoning steps, the text content input to the third language model is limited to the text content of the mapped paper chapters based on the mapping relationship.

[0015] Furthermore, the steps for generating detailed information for this reasoning step include:

[0016] For each reasoning step, an independent context is created and input into the third language model. The independent context includes only the text content of the paper chapter corresponding to the reasoning step and the key scientific question.

[0017] Furthermore, the core reasoning framework is generated and stored in the form of structured data objects.

[0018] Furthermore, the structured data object is a JSON format object.

[0019] Further steps in the screening process include:

[0020] The preliminary reasoning logic chain and the full text of the thesis are input into the long window large language model to evaluate multiple preset evaluation dimensions and obtain the confidence score of each evaluation dimension.

[0021] A comprehensive evaluation score is calculated based on the confidence scores of each of the aforementioned evaluation dimensions;

[0022] The initial reasoning logic chain that scores below the preset quality threshold is filtered out.

[0023] Furthermore, the multiple evaluation dimensions include at least: semantic conformity, logical coherence, and reasoning rigor;

[0024] The semantic conformity degree is used to characterize the degree of consistency between the content of the reasoning logic chain and the meaning expressed in the full text of the thesis;

[0025] The logical coherence is used to characterize the logical progression relationship between multiple reasoning steps in the reasoning logic chain;

[0026] The rigor of reasoning is used to characterize the degree to which the argument of the reasoning logic chain is supported by the corresponding evidence in the full text of the thesis.

[0027] Furthermore, the step of obtaining the confidence scores for each evaluation dimension includes:

[0028] For each evaluation dimension, the Logits values ​​corresponding to preset positive and negative judgments are extracted from the output of the long-window large language model;

[0029] The Logits values ​​are normalized using the Softmax function to obtain the probability value of a positive judgment, and this probability value is used as the confidence score for this evaluation dimension.

[0030] In a second aspect, the present invention proposes a system for constructing thought chain data from dissertations, and a method for constructing thought chain data from dissertations, the system comprising:

[0031] The data processing module is configured to parse the document of the thesis to extract the thesis title, table of contents, and abstract;

[0032] The paper title, chapter list, and abstract are input into the first language model to generate text containing key scientific questions, and the key scientific questions are identified from the text.

[0033] The scientific question identification module is configured to input the key scientific questions and chapter directory into the second language model to generate a core reasoning framework, which includes multiple reasoning steps and the corresponding paper chapters for each reasoning step.

[0034] The thought chain extraction module is configured to extract the text content of the corresponding thesis chapters from the thesis for each reasoning step in the core reasoning framework, input the text content and the key scientific question into the third language model, generate detailed content for the reasoning step, and integrate the detailed content of all reasoning steps into a preliminary reasoning logic chain.

[0035] The quality assessment module is configured to use a long-window, large-language model to compare and verify the preliminary reasoning logic chain with the full text of the thesis, and to filter based on the verification results, outputting the final reasoning logic chain.

[0036] The beneficial effects of this invention are:

[0037] This invention achieves end-to-end construction of structured, high-quality thought chain data from raw, unstructured PDF dissertations through a series of automated steps. Compared to manual writing methods relying on domain experts, this invention significantly reduces labor costs and time consumption, substantially improves the production efficiency of high-quality thought chain data, and is feasible for large-scale application.

[0038] Unlike techniques that rely on large models to create their own reasoning processes, this invention consistently uses the original text of the thesis as its factual basis. By explicitly anchoring each step in the core reasoning framework to a specific chapter of the thesis and extracting detailed arguments from it, this ensures that the generated thought chain remains faithful to the author's original research logic, effectively avoiding illusions, and ensuring that each reasoning step has a clear and traceable textual source.

[0039] This invention introduces a global verification and filtering process based on a long-window, large language model. By comprehensively evaluating the initially generated reasoning logic chain and the full text of the paper from multiple dimensions (such as semantic compliance, logical coherence, and reasoning rigor), and filtering based on quantitative scores, low-quality and inconsistent data can be effectively eliminated, thereby ensuring the high reliability and consistency of the final output dataset.

[0040] This invention not only extracts the surface information of the paper, but also deeply explores the core scientific questions that run through the entire text and the complete argumentation process for solving those questions through the semantic understanding capabilities of a large language model. Finally, it outputs the data in a structured format (such as JSON), facilitating direct processing and utilization by computers, thus laying a solid data foundation for subsequent applications such as AI model training and knowledge graph construction. Attached Figure Description

[0041] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0042] Figure 1 This is a flowchart illustrating a method for constructing thought chain data from a thesis according to the present invention.

[0043] Figure 2 This is a structural diagram of a system for constructing thought chain data from a thesis according to the present invention;

[0044] Figure 3 This is a schematic diagram of the structure of a computer system used to implement the methods, systems, and devices of this application. Detailed Implementation

[0045] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0046] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0047] The first embodiment of the present invention provides a method for constructing thought chain data from dissertations, comprising:

[0048] Step S10: Parse the document of the thesis to extract the thesis title, table of contents and abstract;

[0049] Step S20: Input the paper title, chapter list and abstract into the first language model to generate text containing key scientific questions, and identify the key scientific questions from the text;

[0050] Step S30: Input the key scientific questions and chapter list into the second language model to generate a core reasoning framework. The core reasoning framework includes multiple reasoning steps and the corresponding chapters of the paper for each reasoning step.

[0051] Step S40: For each reasoning step in the core reasoning framework, extract the text content of the corresponding thesis chapter from the thesis, input the text content and the key scientific question into the third language model to generate detailed content for the reasoning step, and integrate the detailed content of all reasoning steps into a preliminary reasoning logic chain.

[0052] Step S50: Using a long-window large language model, the preliminary reasoning logic chain is compared and verified with the full text of the thesis, and the final reasoning logic chain is output based on the verification results.

[0053] To more clearly illustrate the method of constructing thought chain data from dissertations according to the present invention, the following will be combined with... Figure 1 The steps in the embodiments of the present invention are described in detail below:

[0054] Step S10: Parse the document of the thesis to extract the thesis title, table of contents and abstract;

[0055] This invention discloses a method for constructing thought chain data from dissertations. The initial step involves parsing the dissertation document to extract the title, table of contents, and abstract. In practice, the method first acquires electronic documents of the target dissertations in batches via a pre-defined programming interface (e.g., APIs provided by major academic databases) or by accessing the local file system. These documents are typically in PDF format. After acquisition, specialized document parsing libraries such as PyMuPDF or pdfplumber are used to perform in-depth processing of the PDF files.

[0056] The process begins with comprehensive text extraction to capture all visible text content and key metadata within the document. This metadata provides detailed information for subsequent structured analysis and specifically includes basic formatting and positional information such as font, font size, bold status, color, and page coordinates for each text block. After the initial extraction of text and metadata, structured analysis is then performed.

[0057] Specifically, the identification of the paper's title is accomplished by comprehensively analyzing significant visual features of the text block on the first page of the document, such as the largest font size, centered or top-aligned position, and bold style. The extraction of the paper's abstract involves searching the entire text for predefined keywords such as "Abstract" to locate its starting position, and then continuously extracting text until encountering a next chapter separator such as "Keywords" or "Chapter 1 Introduction," thereby defining and obtaining the complete abstract content. The construction of the chapter headings employs a rule-based and pattern recognition method. This method traverses the entire document, identifying all text lines with commonalities in format, such as uniform font, fixed indentation, and specific numbering formats, as potential chapter headings. Subsequently, by logically parsing heading numbers such as "Chapter 1," "1.1," and "2.3.4," the hierarchical relationship and order between them are inferred, ultimately reconstructing the tree-like directory structure of the entire paper.

[0058] Finally, all accurately identified and extracted paper titles, abstracts, and hierarchical chapter lists are integrated and transformed into a standardized structured data object. This could be a JSON object representing the chapter list as a nested list like this: [{"chapter_id": "1", "title": "Introduction"}, {"chapter_id": "2.1", "title": "Related Theoretical Foundations"}, ...]. Here, `chapter_id` represents the chapter number, and `title` represents the chapter title. This structured data object serves as clean and well-organized input for subsequent processing steps such as identifying key scientific questions. Ultimately, a structured list of chapters is formed, in the following format:

[0059] [{"chapter_id": "1", "title": "Introduction"}, {"chapter_id": "2.1", "title":"Relevant Theoretical Foundations"}, ...].

[0060] Step S20: Input the paper title, chapter list and abstract into the first language model to generate text containing key scientific questions, and identify the key scientific questions from the text;

[0061] After parsing the dissertation and extracting its title, table of contents, and abstract, the next step of this method is to input the title, table of contents, and abstract into the first large language model to generate text containing the key scientific question, and then identify the key scientific question from this text. Specifically, this step involves combining the dissertation title, table of contents, and abstract obtained in the previous step and embedding them together into a pre-designed prompt template to construct a complete input text. This prompt template is designed with explicit text instructions, aiming to guide the large language model to act as a senior academic expert, requiring it to accurately summarize the single and most critical scientific question that the dissertation aims to solve based on the provided meta-information, and to clearly state it in one sentence. The prompt template is designed as follows:

[0062] "You are a senior academic expert. Please refer to the following meta-information from academic papers:"

[0063] Paper title: {T};

[0064] Table of Contents: {C};

[0065] Abstract: {A};

[0066] Please accurately summarize the most critical scientific question that this paper aims to solve. Please state this question clearly in one sentence.

[0067] The constructed complete input text is provided to the first major language model for processing. After understanding the semantics and instructions of the input text, the first major language model generates a concise statement of the key scientific question. Its output form is usually a complete sentence with a guiding prefix, such as generating text in the form of "This study aims to solve the problem of..." or "The core scientific question of this paper is...". Guided by prompt word templates, the model understands academic text and outputs concise statements in the form of "This study aims to solve the problem of..." or "The core scientific question of this paper is...".

[0068] To identify the standardized key scientific questions from the model-generated text, post-processing is required. Specifically, text processing techniques such as regular expression matching are applied to identify and remove leading prefixes such as "This study aims to solve" or "The core scientific question of this thesis is..." This allows for the precise separation and recording of the remaining core text content, which is then identified as the key scientific question of the thesis for subsequent use. The model output is post-processed to remove leading prefixes and recorded as the standardized key scientific question text.

[0069] Step S30: Input the key scientific questions and chapter list into the second language model to generate a core reasoning framework. The core reasoning framework includes multiple reasoning steps and the corresponding chapters of the paper for each reasoning step.

[0070] In this embodiment, the core reasoning framework includes a mapping relationship between each reasoning step and at least one chapter of a paper; the method further includes:

[0071] When generating the detailed content of the reasoning steps, the text content input to the third language model is limited to the text content of the mapped paper chapters based on the mapping relationship.

[0072] The core reasoning framework is generated and stored in the form of structured data objects, which are in JSON format.

[0073] After identifying the aforementioned key scientific questions, the next step in this method is to input the key scientific questions and chapter list into the second large language model to generate a core reasoning framework. Specifically, the key scientific questions are first combined with the extracted chapter list and integrated into a prompt word template designed for the second large language model, containing specific logical reasoning instructions, to construct the complete input text. This prompt word template aims to guide the second large language model to simulate the thinking of the paper's author, deduce the core logical steps required to solve the key scientific questions, and force the model to map each deduced logical step to at least one most relevant chapter of the paper based on the provided chapter list. The prompt words are designed to guide the model in logical reasoning, as follows:

[0074] Given the table of contents structure of an academic paper: {C};

[0075] And its core scientific question: {K};

[0076] Imagine yourself as the author of the paper and deduce the key logical reasoning steps required to solve this core problem. Please list the core steps.

[0077] Then, based on the table of contents, map each of the above reasoning steps to one or more of the most relevant specific chapters.

[0078] Meanwhile, the prompt word template also includes strict formatted output instructions, requiring the model to output the result in the form of a structured data object. In this embodiment, the structured data object is specifically a JSON format object. "Please output in strict JSON format:"

[0079] {

[0080] "reasoning_chain_outline": [

[0081] {"step_id": 1, "reasoning_step": "Description of the first step", "relevant_chapters":["Chapter ID1", "Chapter ID2"]}, ... ]

[0084] }”

[0085] Here, reasoning_chain_outline is the reasoning chain outline, and step_id is the step identifier.

[0086] After the second language model receives and processes the input text, it generates and outputs a JSON object containing the core reasoning framework. This object defines a list of multiple reasoning steps, each accompanied by a mapping relationship containing at least one chapter identifier from the paper. The model outputs a structured JSON object containing the backbone of the reasoning chain (reasoning_step) and the mapping relationship between each step and a specific chapter in the paper (relevant_chapters). This JSON object represents the initial form of the associative data chain. This JSON object, containing the mapping relationships, not only provides a structured extraction of the overall argumentative structure of the paper but, more importantly, serves as a crucial basis for subsequent steps in generating detailed reasoning steps.

[0087] Specifically, during the subsequent detailed content extraction for each reasoning step, the text content input to the third language model will be strictly limited to the text content of the specific paper chapter mapped to that reasoning step, rather than the full text of the paper or other irrelevant chapters, based strictly on the mapping relationship generated here. In this way, this step generates and stores a core reasoning framework that precisely defines the core reasoning steps and their correspondence with paper chapters, providing necessary and accurate guidance for subsequent refined content filling by chapter.

[0088] Step S40: For each reasoning step in the core reasoning framework, extract the text content of the corresponding thesis chapter from the thesis, input the text content and the key scientific question into the third language model to generate detailed content for the reasoning step, and integrate the detailed content of all reasoning steps into a preliminary reasoning logic chain.

[0089] In this embodiment, the step of generating detailed content for this reasoning step includes:

[0090] For each reasoning step, an independent context is created and input into the third language model. The independent context includes only the text content of the paper chapter corresponding to the reasoning step and the key scientific question.

[0091] Among them, at least two of the first large language model, the second large language model and the third large language model are the same large language model.

[0092] After generating a core reasoning framework that includes each reasoning step and its mapping relationship with the chapters of the paper, this method then performs a refined content filling operation for each reasoning step in the core reasoning framework.

[0093] 1. Chapter Content Preparation: Based on the relevant_chapters list obtained from S30, the complete text content of the corresponding chapters is extracted from the original paper PDF using a parser.

[0094] 2. Detailed Extraction by Chapter: For each key chapter, its complete text, Content_Chapter, is input into the large language model along with the key scientific question K. Hint words are as follows:

[0095] "Core scientific question: {K};"

[0096] The full text of the current chapter: {Content_Chapter};

[0097] Based on the aforementioned scientific question and the content of this chapter, please elaborate on how this chapter serves to solve that scientific question. Please extract the **core argumentation process**, **key experimental data / theoretical derivation conclusions**, or **connecting logical nodes** within this chapter.

[0098] The output should be specific, coherent, and directly support the solution of the core scientific questions.

[0099] 3. Content Integration: The model outputs a detailed reasoning summary for each chapter. The system then "fills" the detailed content of each chapter into the corresponding reasoning steps according to the order of `reasoning_chain_outline` defined in S30. For example, the detailed content of chapter "3.2" is associated with `reasoning_step` with step_id 2. Ultimately, a complete reasoning logic chain L is formed, consisting of "step skeleton → chapter mapping → detailed content".

[0100] In practice, based on the mapping relationship generated in the previous step, the complete text content of the specific thesis chapter corresponding to the reasoning step is located and accurately extracted from the stored text data of the original thesis.

[0101] To ensure the model can make accurate inferences based on local and highly relevant information, a separate context will be created for the current inference step. This separate context is strictly limited in content; it only includes the text content of the extracted paper section that directly corresponds to the current inference step, as well as the previously identified key scientific questions, and does not include other parts of the paper.

[0102] This constructed, complete input text, containing independent context, is provided to the third language model for processing. This input is accompanied by specific prompts to guide the third language model in elaborating on how the current chapter serves to solve the key scientific problem, and extracting the core argumentation process, key experimental data, theoretical derivations, or logical connecting points within the chapter, thereby generating a detailed text for the current reasoning step. After traversing all reasoning steps in the core reasoning framework and generating detailed content for each, all generated detailed text paragraphs are integrated and pieced together according to the inherent logical order within the core reasoning framework—that is, the chronological order of the reasoning steps—to construct a complete and coherent logical chain of reasoning as a preliminary result for use in subsequent quality verification steps.

[0103] Furthermore, in this embodiment, the third large language model used to perform this step, and at least two of the first and second large language models used in the preceding steps, can be the same specific large language model instance. For example, the same deployed model service can be called uniformly throughout the entire method flow to complete all tasks related to the large language model.

[0104] Those skilled in the art can choose any existing large language model according to actual needs. This invention does not make any specific limitations, and any large language model is within the protection scope of this invention.

[0105] Step S50: Using a long-window large language model, the preliminary reasoning logic chain is compared and verified with the full text of the thesis, and the final reasoning logic chain is output based on the verification results.

[0106] In this embodiment, the screening step includes:

[0107] Step S51: Input the preliminary reasoning logic chain and the full text of the thesis into the long window large language model to evaluate multiple preset evaluation dimensions and obtain the confidence score of each evaluation dimension.

[0108] Step S52: Calculate the comprehensive evaluation score based on the confidence scores of each evaluation dimension;

[0109] Step S53: Filter out the preliminary reasoning logic chains whose comprehensive evaluation scores are lower than the preset quality threshold.

[0110] In another embodiment, the screening step can also be implemented using a pre-trained discriminative classification model. The initial inference logic chain and the full text of the paper are provided to the model as input, and the model directly outputs a 'keep' or 'reject' binary classification result to filter out low-quality data.

[0111] Among them, the multiple evaluation dimensions include at least: semantic conformity, logical coherence, and reasoning rigor;

[0112] The semantic conformity degree is used to characterize the degree of consistency between the content of the reasoning logic chain and the meaning expressed in the full text of the thesis;

[0113] The logical coherence is used to characterize the logical progression relationship between multiple reasoning steps in the reasoning logic chain;

[0114] The rigor of reasoning is used to characterize the degree to which the argument of the reasoning logic chain is supported by the corresponding evidence in the full text of the thesis.

[0115] The steps for obtaining the confidence scores for each evaluation dimension in this embodiment include:

[0116] Step S511: For each evaluation dimension, extract the Logits values ​​corresponding to the preset positive and negative judgments from the output of the long-window large language model; wherein, the Logits values ​​are the unnormalized output scores in the machine learning model.

[0117] Step S512: Normalize the Logits value using the Softmax function to obtain the probability value of a positive judgment, and use this probability value as the confidence score for this evaluation dimension.

[0118] After generating the initial reasoning logic chain, this method proceeds to the crucial quality verification and screening step, which is implemented as follows:

[0119] First, a large language model with long context processing capabilities is used. The previously generated preliminary reasoning logic chain and its corresponding complete full-text dissertation are used as common inputs to comprehensively compare and verify the preliminary reasoning logic chain. A multi-dimensional quality assessment system is constructed, with dimensions including at least semantic conformity, logical coherence, and reasoning rigor. This verification process is not a single-dimensional judgment but is achieved through a refined screening process. The first step of this process is to evaluate each of the pre-defined assessment dimensions and obtain the confidence score for each dimension. These assessment dimensions are designed to include at least three core aspects:

[0120] The first is semantic conformity, which is specifically used to characterize and measure the degree of consistency between the content of the reasoning logic chain and the overall meaning to be expressed in the full text of the thesis. Semantic conformity: whether the content of the thought chain is highly consistent with the meaning expressed in the full text of the thesis, and whether there are factual errors or omissions of core arguments.

[0121] The second is logical coherence, which is used to characterize and evaluate whether there is a clear and reasonable logical progression relationship between multiple reasoning steps within the reasoning logic chain.

[0122] Thirdly, there is the rigor of reasoning. This dimension is used to characterize and confirm the extent to which the arguments and conclusions in the reasoning logic chain can be supported by specific evidence from the corresponding chapters or paragraphs in the full text of the thesis.

[0123] Furthermore, the semantic conformity score is obtained as follows: the long-window large language model performs deep semantic understanding and alignment between the preliminary reasoning logic chain and the full text of the thesis. By decomposing them into semantic units and constructing fine-grained mapping relationships, the system analyzes the degree of support and correspondence of each claim, fact, and conclusion in the reasoning logic chain in the context of the full text, and then comprehensively evaluates the overall content consistency and generates a quantitative score.

[0124] The logical coherence score is obtained as follows: the long-window large language model performs structural deconstruction and sequence analysis on the multiple reasoning steps contained in the preliminary reasoning logic chain. By identifying logical connectors, premise-conclusion dependencies, and argument transition relationships between steps, it verifies whether the sequence of steps constitutes a formal reasoning process that is interlocked, progressively layered, and without internal contradictions. Based on this, a quantitative score representing the logical rationality is generated.

[0125] The score for the rigor of the reasoning is obtained as follows: the long-window large language model performs source tracing and evidence retrieval on each key argument node in the preliminary reasoning logic chain within the entire text of the thesis. By accurately associating and analyzing the argument nodes with experimental data, theoretical derivations, cited arguments and related descriptions in the full text, the system judges whether the argument is supported by sufficient, direct and diverse original text evidence, and generates a quantitative score representing the solidity of the argument based on this.

[0126] In the specific implementation process, the technical details of generating quantitative scores for the three evaluation dimensions are as follows. For the evaluation of semantic conformity, the initial reasoning logic chain and the full text of the thesis first need to be preprocessed and segmented into texts, dividing them into segments or sentences with complete semantics, i.e., semantic units. Subsequently, a pre-trained semantic encoding model (such as Sentence-BERT or a similar text embedding model) is called to vectorize each semantic unit, resulting in a high-dimensional semantic vector representation. The core of the evaluation process lies in performing semantic unit-level alignment: the system calculates the cosine similarity between each semantic unit vector in the reasoning logic chain and all semantic unit vectors in the full text of the thesis, finding the most matching corresponding part for each unit in the logic chain within the entire text, and recording the highest similarity score. Furthermore, the system will focus on comparing specific types of semantic units such as core claims, key facts, and concluding statements. By using a weighted average or threshold judgment method, it will synthesize the alignment scores of all units and combine them with semantic judgment criteria defined by domain knowledge (e.g., whether the core claim is supported by multiple parts of the text, whether the expression of key facts is unambiguous with the original text, etc.) to finally synthesize a confidence score between 0 and 1. This score represents the coverage and fidelity of the overall content of the reasoning logic chain to the main theme of the paper.

[0127] For assessing logical coherence, the initial logical chain of reasoning is first parsed into a series of discrete reasoning steps. A long-window, large-language model is prompted or invoked with specific logical analysis functions to perform structured parsing of these steps. This process includes identifying subject-verb-object structures, conditional adverbs, conclusive words (such as "therefore," "proves," "leads to"), and explicit logical connectors between steps (such as "first," "next," "however," "in conclusion"). Crucially, the model needs to analyze implicit dependencies between steps, such as determining whether the conclusion of a previous step constitutes a premise or input condition for the subsequent step. The system can integrate or access a pre-defined logical rule base and coherence templates, containing common valid reasoning patterns (such as deduction, induction, and causal chains) and typical logical fallacies. By matching and verifying the parsed step relationships with these rules and templates, the system can determine whether the entire sequence of steps contains logical contradictions, argument gaps, or disordered order. Finally, based on the rule compliance and the model's own comprehensive judgment of logical fluency, a quantitative score is generated, reflecting the consistency and rationality of the argumentation process.

[0128] The assessment of the rigor of reasoning begins with the identification and extraction of each independent argument node in the initial logical chain of reasoning. Each argument node typically represents a claim or assertion. Subsequently, a long-window, large-language model uses each argument node as a query to conduct in-depth retrieval and evidence location within the context of the entire dissertation. This is not a simple keyword matching process, but rather a search for textual evidence that supports, explains, or corroborates the assertion within the entire text based on semantic understanding. This evidence may be scattered throughout methodological descriptions, experimental results, data charts, theoretical derivations, or relevant literature citations. For each potential piece of evidence located, the model needs to perform multi-dimensional analysis: assessing its semantic relevance to the argument node (whether it directly addresses the same topic), directness (whether it is a direct statement or an indirect implication), and sufficiency (whether the evidence itself is sufficiently strong, such as whether the experimental data is statistically significant or whether the theoretical derivation is complete). The system evaluates the evidence set for each argument node according to pre-defined evaluation criteria (e.g., classifying evidence into levels such as "strong direct support," "weak indirect support," and "no support or contradiction"). Finally, by synthesizing the strength of evidence support for all argument nodes (e.g., calculating the proportion of nodes with "strong direct support" and combining it with the score of the weakest supporting element), an overall confidence score is generated. This score quantifies the degree to which the reasoning logic chain is rooted and solid within the original text. The scoring process for these three evaluation dimensions together constitutes the core technical step in automating and refining the quality verification of the initial reasoning logic chain.

[0129] To obtain the confidence score for each of the above evaluation dimensions, this method further includes a detailed set of score generation steps.

[0130] For each evaluation dimension, a corresponding verification prompt template is constructed. For example, the semantic compliance prompt template is as follows:

[0131] "You are a high-quality academic reviewer. Now you have a full dissertation and a core line of reasoning that you claim to have extracted from it."

[0132] Full text of the paper: {Full_Text};

[0133] The thought chain to be evaluated: {L};

[0134] Please evaluate this thought chain from the perspective of semantic conformity, and give a judgment on whether the content of the full text and the thought chain correspond and are semantically consistent.

[0135] Semantic consistency: Whether the content of the thought chain is highly consistent with the meaning expressed in the whole paper, and whether there are any factual errors or omissions of core arguments.

[0136] Please make a judgment based on the given information. First, output True / False, and then explain your reasoning.

[0137] The prompt templates for other dimensions can be built using this framework.

[0138] Specifically, for any evaluation dimension, the system provides a clear instruction to the long-window large language model, guiding it to judge the input content. After the model generates output, the system precisely extracts the original Logits values ​​corresponding to the preset positive judgments (e.g., "True" or "compliant") and negative judgments (e.g., "False" or "incompatible") from the output layer of the long-window large language model. When calling the large model interface to obtain the output, the system extracts the Logits values ​​of the "True" and "False" token positions in the lm_head layer of the first generated token. After normalization using the Softmax function, the probability distributions of "True" and "False" are obtained, and the probability value of "True" is used as the confidence score of the data in this dimension. Specifically, for each evaluation dimension, its confidence score is calculated. Calculated using the following formula: ,in, and These represent the Logit values ​​corresponding to "True" and "False" respectively. This function represents the score for a specific dimension. It transforms the raw, unbounded Logits values ​​into a probability distribution between 0 and 1.

[0139] After normalization, the probability value of a positive judgment is formally adopted as the final confidence score for that evaluation dimension. After obtaining the confidence scores for all evaluation dimensions, the system calculates a comprehensive evaluation score based on these independent confidence scores. The comprehensive evaluation score... The calculation formula is: Where N is the number of evaluation dimensions, and n is the index variable. It is the sum of confidence scores across N dimensions.

[0140] The calculated overall evaluation score is compared with a pre-set empirical quality threshold. If the overall evaluation score of a preliminary inference logic chain is lower than the quality threshold, the logic chain will be judged as low-quality data and automatically filtered out by the system. Only those preliminary inference logic chains whose overall evaluation scores reach or exceed the quality threshold can pass the screening and be output as the final, verified high-quality inference logic chains.

[0141] To more fully disclose the implementation process of this method, the following describes the process of automating the construction of the thought chain in an example doctoral dissertation, "Research on Remote Sensing Image Target Detection Algorithm Based on Deep Learning," using a specific application example.

[0142] First, the PDF document of the thesis was parsed to extract the thesis title, table of contents, and abstract. The parsing yielded the thesis title, "Research on Target Detection Algorithm for Remote Sensing Images Based on Deep Learning," and an abstract text containing the core challenges and objectives. Structured table of contents information was then extracted. This table of contents was converted into a JSON data object in the following format: [{"chapter_id": "1", "title": "Introduction"}, {"chapter_id": "2", "title": "Relevant Theoretical Basis"}, {"chapter_id": "3", "title": "Target Detection Model Based on Multi-Scale Feature Enhancement"}, {"chapter_id": "4", "title": "Detection Head Optimization Based on Attention Mechanism"}, {"chapter_id": "5", "title": "Experiments and Analysis"}, {"chapter_id": "6", "title": "Summary and Outlook"}].

[0143] The paper's title, table of contents, and abstract are input into a first-level language model to generate and determine the key scientific question. The first-level language model outputs: "The core scientific question of this paper is how to effectively improve the detection accuracy of multi-scale, small targets in remote sensing images under complex scenes." After post-processing, the key scientific question K is determined to be: "How to effectively improve the detection accuracy of multi-scale, small targets in remote sensing images under complex scenes?"

[0144] Subsequently, the key scientific question K and the chapter list are input into the second large language model to generate the core reasoning framework. The second large language model outputs a structured JSON object that defines the mapping relationship between reasoning steps and paper chapters. An example output is shown below:

[0145] {

[0146] "reasoning_chain_outline": [

[0147] {"step_id": 1, "reasoning_step": "Clarifying the difficulties and shortcomings of existing methods in remote sensing image target detection", "relevant_chapters": ["1"]},

[0148] {"step_id": 2, "reasoning_step": "Establish the theoretical foundation for deep learning and object detection","relevant_chapters": ["2"]},

[0149] {"step_id": 3, "reasoning_step": "Design and implement a multi-scale feature enhancement module to improve the feature representation of small targets", "relevant_chapters": ["3"]},

[0150] {"step_id": 4, "reasoning_step": "Design and introduce an attention mechanism to optimize the detection head and improve localization and classification accuracy", "relevant_chapters": ["4"]},

[0151] {"step_id": 5, "reasoning_step": "Comparative experiments verify the superiority of the proposed model in multi-scale, small-object detection tasks", "relevant_chapters": ["5"]} ]

[0153] }

[0154] For each reasoning step in the core reasoning framework, detailed reasoning content is filled in. Based on the mapping relationship, the complete text content of the corresponding chapter is extracted from the thesis. Taking Chapter 3, which corresponds to the third step, as an example, the text content of Chapter 3 and the key scientific question K are input into the third language model to generate the detailed content of this step.

[0155] The detailed content generated by the third language model is as follows: "This chapter addresses the problem of easy loss of features for small targets by proposing a cross-level feature fusion pyramid structure. This structure integrates deep semantic features with shallow detail features by introducing a bidirectional path. Preliminary experiments on the COCO remote sensing dataset show that this module improves the average detection accuracy (AP_s) of small targets by about 5%." After traversing all inference steps and completing content generation, the steps are integrated in sequence to form a preliminary inference logic chain L.

[0156] A long-window large language model is used to verify and screen the quality of the preliminary inference logic chain L. The inference logic chain L and the full text of the thesis are input into the long-window large language model, and evaluated based on three preset dimensions: semantic compliance, logical coherence, and reasoning rigor. For each evaluation dimension, the corresponding "True" and "False" Logits are extracted from the output of the long-window large language model. In this embodiment, the confidence scores for the three dimensions are calculated to be 0.92, 0.89, and 0.85, respectively. Then, a comprehensive evaluation score is calculated based on the confidence scores of each dimension. The preset quality threshold was set to 0.7. Since the comprehensive evaluation score of 0.887 was higher than the threshold, the reasoning logic chain passed the verification and was output as the final thought chain data.

[0157] In addition, as an advanced implementation, the preliminary reasoning logic chain and the full text of the paper can be input into a pre-trained discriminative classification model, which can then directly make a binary decision of "keep" or "delete" to complete the screening.

[0158] This application example fully demonstrates the entire process of the method, from document parsing to the final output of high-quality thought chain data. Through these steps, the implicit reasoning logic in dissertations can be automatically extracted and verified with high fidelity.

[0159] like Figure 2 As shown, a system for constructing thought chain data from dissertations according to a second embodiment of the present invention is used to implement a method for constructing thought chain data from dissertations. The system includes:

[0160] The data processing module is configured to parse the document of the thesis to extract the thesis title, table of contents, and abstract;

[0161] The paper title, chapter list, and abstract are input into the first language model to generate text containing key scientific questions, and the key scientific questions are identified from the text.

[0162] The scientific question identification module is configured to input the key scientific questions and chapter directory into the second language model to generate a core reasoning framework, which includes multiple reasoning steps and the corresponding paper chapters for each reasoning step.

[0163] The thought chain extraction module is configured to extract the text content of the corresponding thesis chapters from the thesis for each reasoning step in the core reasoning framework, input the text content and the key scientific question into the third language model, generate detailed content for the reasoning step, and integrate the detailed content of all reasoning steps into a preliminary reasoning logic chain.

[0164] The quality assessment module is configured to use a long-window, large-language model to compare and verify the preliminary inference logic chain with the full text of the thesis, and to filter based on the verification results, outputting the final inference logic chain. Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related explanations of the method described above can be found in the corresponding processes of the foregoing system embodiments, and will not be repeated here.

[0165] It should be noted that the system for constructing thought chain data from dissertations provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the various modules or steps and are not considered as an improper limitation of the present invention.

[0166] A device according to a third embodiment of the present invention includes:

[0167] At least one processor;

[0168] and a memory communicatively connected to at least one of the processors;

[0169] The memory stores instructions that can be executed by the processor to implement the above-described method for constructing thought chain data from a thesis.

[0170] A fourth embodiment of the present invention provides a computer-readable storage medium storing computer instructions for execution by a computer to implement the above-described method for constructing thought chain data from a thesis.

[0171] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the storage device and processing device described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0172] The following is for reference. Figure 3 It shows a schematic diagram of the structure of a computer system for implementing embodiments of the systems, methods, and electronic devices of this application. Figure 3 The server shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0173] like Figure 3As shown, the computer system includes a Central Processing Unit (CPU) 301, which can perform various appropriate actions and processes based on programs stored in Read Only Memory (ROM) 302 or programs loaded from storage section 308 into Random Access Memory (RAM) 303. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An Input / Output (I / O) interface 305 is also connected to the bus 304.

[0174] The following components are connected to I / O interface 305: an input section 306 including a keyboard, mouse, etc.; an output section 307 including a cathode ray tube (CRT), liquid crystal display (LCD), and speakers, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card and a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0175] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0176] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0177] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0178] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0179] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0180] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for constructing thought chain data from dissertations, characterized in that, include: The document of the thesis is parsed to extract the thesis title, table of contents and abstract; The paper title, chapter list, and abstract are input into the first language model to generate text containing key scientific questions, and the key scientific questions are identified from the text. The key scientific questions and chapter list are input into the second language model to generate a core reasoning framework, which includes multiple reasoning steps and the corresponding chapters of the paper for each reasoning step. For each reasoning step in the core reasoning framework, the text content of the corresponding thesis chapter is extracted from the thesis, and the text content and the key scientific question are input into the third language model to generate detailed content for the reasoning step. The detailed content of all reasoning steps is then integrated into a preliminary reasoning logic chain. A long-window, large-language model is used to compare and verify the preliminary reasoning logic chain with the full text of the thesis. Based on the verification results, the final reasoning logic chain is output.

2. The method according to claim 1, characterized in that, At least two of the first, second, and third language models are the same large language model.

3. The method according to claim 1, characterized in that, The core reasoning framework includes a mapping relationship between each reasoning step and at least one chapter of the paper; the method also includes: When generating the detailed content of the reasoning steps, the text content input to the third language model is limited to the text content of the mapped paper chapters based on the mapping relationship.

4. The method according to claim 1 or 3, characterized in that, The steps for generating detailed information for this reasoning step include: For each reasoning step, an independent context is created and input into the third language model. The independent context includes only the text content of the paper chapter corresponding to the reasoning step and the key scientific question.

5. The method according to claim 3, characterized in that, The core reasoning framework is generated and stored in the form of structured data objects.

6. The method according to claim 5, characterized in that, The structured data object is a JSON format object.

7. The method according to claim 1, characterized in that, The screening steps include: The preliminary reasoning logic chain and the full text of the thesis are input into the long window large language model to evaluate multiple preset evaluation dimensions and obtain the confidence score of each evaluation dimension. A comprehensive evaluation score is calculated based on the confidence scores of each of the aforementioned evaluation dimensions; The initial reasoning logic chain that scores below the preset quality threshold is filtered out.

8. The method according to claim 7, characterized in that, The multiple evaluation dimensions include at least: semantic conformity, logical coherence, and reasoning rigor; The semantic conformity degree is used to characterize the degree of consistency between the content of the reasoning logic chain and the meaning expressed in the full text of the thesis; The logical coherence is used to characterize the logical progression relationship between multiple reasoning steps in the reasoning logic chain; The rigor of reasoning is used to characterize the degree to which the argument of the reasoning logic chain is supported by the corresponding evidence in the full text of the thesis.

9. The method according to claim 7 or 8, characterized in that, The steps to obtain confidence scores for each evaluation dimension include: For each evaluation dimension, the Logits values ​​corresponding to preset positive and negative judgments are extracted from the output of the long-window large language model; The Logits values ​​are normalized using the Softmax function to obtain the probability value of a positive judgment, and this probability value is used as the confidence score for this evaluation dimension.

10. A system for constructing thought chain data from dissertations, used to implement the method for constructing thought chain data from dissertations as described in any one of claims 1-9, characterized in that, The system includes: The data processing module is configured to parse the thesis document to extract the thesis title, table of contents, and abstract; The paper title, chapter list, and abstract are input into the first language model to generate text containing key scientific questions, and the key scientific questions are identified from the text. The scientific question identification module is configured to input the key scientific questions and chapter directory into the second language model to generate a core reasoning framework, which includes multiple reasoning steps and the corresponding chapters of the paper for each reasoning step. The thought chain extraction module is configured to extract the text content of the corresponding thesis chapters from the thesis for each reasoning step in the core reasoning framework, input the text content and the key scientific question into the third language model, generate detailed content for the reasoning step, and integrate the detailed content of all reasoning steps into a preliminary reasoning logic chain. The quality assessment module is configured to use a long-window, large-language model to compare and verify the preliminary reasoning logic chain with the full text of the thesis, and to filter based on the verification results, outputting the final reasoning logic chain.

Images