Large model-based scientific research paper structure analysis method and system
By constructing a citation dependency graph of scientific research papers and breaking closed-loop paths, the loop problem in large models when parsing scientific research papers is solved, and the stability and consistency of the parsing process are improved.
Patent Information
- Application Number
- CN202610132385.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-30
- Publication Date
- 2026-05-22
AI Technical Summary
Large models suffer from problems such as repeated parsing, inconsistent output, and low parsing efficiency due to closed-loop citations when parsing scientific research papers, which is especially evident in long papers or papers with multiple cross-citations.
We construct a citation dependency graph for the paper, detect and break closed-loop paths, convert closed loops into open chains by creating expanded proxy nodes, and make corresponding modifications at the paper text level to ensure that the parsing process has a clear termination boundary.
It improves the stability and consistency of the scientific paper parsing process, avoids the repeated construction of loop contexts, and enhances parsing efficiency and output accuracy.
Smart Images
Figure CN122072820A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer technology, and more specifically, to a method and system for analyzing the structure of scientific research papers based on a large model. Background Technology
[0002] Scientific papers typically possess a strict hierarchical structure and dense cross-references. Chapters, subsections, figures, tables, formulas, and terminology definitions often guide each other through phrases such as "see section ×," "as described in section xx," and "based on the definition in section xx." In recent years, generalized large language models have been widely used for paper structure analysis and information extraction. By understanding the original text, they output structured content such as chapter functions, key points, and the correspondence between methods and results to support applications such as literature review, review generation, and knowledge retrieval enhancement. To make the analysis process more organized, the citation relationships between sections in the original text can be identified and used to organize the analysis order or the context of prompt words, enabling the model to progressively unfold reasoning along the citation chain.
[0003] However, citations in academic papers are not always unidirectional. For narrative coherence, authors often use cross-section references and mutual citations. For example, the methods section might cite the explanation of figures in the results section, while the results section cites the parameter definitions in the methods section; the introduction summarizes the conclusions, while the discussion section refers back to the research questions in the introduction; a section might first introduce a concept and then cite a later section for detailed description, and then, for rigor, cite the source of that concept again. These mutual citations form a closed-loop citation structure, making multiple sections mutually accessible in terms of citation relationships. When large models rely on these citation relationships to construct context or progressive prompts during parsing, closed-loop citations can easily lead to the same section being repeatedly pulled back, prompts constantly overlapping, and the parsing chain failing to terminate naturally, resulting in repeated parsing, inconsistent outputs, or even cyclical generation of similar content. This circularity problem is particularly amplified in long papers, those with multiple cross-citations, or those with non-standard citation identifiers, affecting parsing efficiency and stability. Therefore, a processing mechanism oriented towards the paper's citation structure is needed to ensure that the parsing process remains terminateable under complex mutual citation conditions. Summary of the Invention
[0004] The technical problem to be solved by this invention is to provide a method and system for analyzing the structure of scientific research papers based on a large model, so as to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for analyzing the structure of scientific research papers based on a large model includes the following steps: The large model extracts the citation links between segments in scientific research papers and constructs a citation dependency graph, where the nodes of the citation dependency graph are the segments of the scientific research paper and the edges of the citation dependency graph are the citation links between the segments in the scientific research paper. Closed-loop paths are detected on the paper's citation dependency graph. When a closed-loop path is detected, the starting node of the closed-loop path is determined, and the starting node is copied to create an expanded proxy node, which is an equivalent copy of the starting node. The reference links from the ending node in the closed-loop path to the starting node are cut off, and a reference link from the ending node to the expanded proxy node is established to convert the closed-loop path into an open chain. A new version of the paper citation dependency graph is formed by replacing all closed loop paths with corresponding open chains on the paper citation dependency graph; the citation structure of the research paper is modified according to the new version of the paper citation dependency graph to generate the modified research paper; prompt words are input into the large model, and the large model parses and outputs the results based on the modified research paper according to the prompt words.
[0006] Preferably, the process of modifying a research paper includes: copying and inserting the segment content corresponding to the expanded proxy node into the research paper to form a new segment, replacing the original reference content pointing to the segment corresponding to the starting node in the last segment of the corresponding closed-loop path in the research paper with the reference content pointing to the new segment, and deleting or blocking the reference content of the new segment pointing to the successor node in the corresponding closed-loop path.
[0007] Preferably, the reference content of the newly added segment pointing to the successor node in the corresponding closed-loop path includes the external reference content of the original segment corresponding to the starting node in the newly added segment, and the reference content used to point to the next node of the starting node in the closed-loop path, wherein the next node is a node in the closed-loop path that can be directly reached by the starting node along the reference link.
[0008] Preferably, when a closed-loop path is detected, if the closed-loop path intersects with other closed-loop paths, the intersecting nodes are determined as candidate starting nodes, and the candidate starting node with the most participants in the closed-loop path is selected as the starting node.
[0009] Preferably, when a closed-loop path is detected, if the closed-loop path does not have any intersecting nodes with other closed-loop paths, the node with the fewest connections to external nodes of the closed-loop path is selected as the starting node.
[0010] Preferably, when the candidate starting node is determined to be the starting node of multiple intersecting closed loop paths, an expanded proxy node is created only for the candidate starting node, and the newly added segment corresponding to the expanded proxy node is used as a shared newly added segment for multiple intersecting closed loop paths; wherein, for each intersecting closed loop path, the reference content in the last segment of the closed loop path that originally pointed to the segment corresponding to the candidate starting node is replaced with the reference content that points to the shared newly added segment.
[0011] Preferably, the Tarjan algorithm is applied to the paper citation dependency graph to detect closed-loop paths.
[0012] Preferably, the Tarjan algorithm includes: performing a depth-first traversal on the paper citation dependency graph, assigning an incrementing access number to each visited node, and recording the traceable minimum access number for each visited node; maintaining a node stack during the depth-first traversal to store nodes on the current traversal path, and setting a stack marker for nodes pushed onto the stack; when the traversal returns to a node and the minimum access number of that node is equal to its access number, popping nodes from the node stack sequentially until the node is popped to form a strongly connected node set, and determining it as a closed loop path when the strongly connected node set contains at least 2 nodes.
[0013] Preferably, the large model is a general large language model, including any one of the following: the Doubao large model, the LLaMA series large model, and the GPT series large model.
[0014] This invention also discloses a large-model-based scientific paper structure parsing system for implementing the above method, comprising: The citation link extraction unit is used to extract citation links between sections in scientific papers from a large model. A dependency graph construction unit is used to construct a paper citation dependency graph based on the citation links, wherein the nodes of the paper citation dependency graph are segments of the research paper, and the edges of the paper citation dependency graph are citation links between segments in the research paper. A closed-loop path detection unit is used to detect closed-loop paths on the paper citation dependency graph. The closed-loop unpacking unit is used to determine the starting node of the closed-loop path when a closed-loop path is detected, copy the starting node to create an unpacking proxy node, the unpacking proxy node being an equivalent copy of the starting node, and cut the reference link from the end node in the closed-loop path to the starting node, and establish the reference link from the end node to the unpacking proxy node, so as to convert the closed-loop path into an unclosed chain. The new dependency graph generation unit is used to replace all closed-loop paths with corresponding open chains in the paper citation dependency graph to form a new paper citation dependency graph. The paper citation structure modification unit is used to modify the citation structure of the research paper according to the new version of the paper citation dependency graph, and generate the modified research paper. The parsing output unit is used to input prompt words into the large model, so that the large model can parse the modified scientific paper based on the prompt words and output the results.
[0015] The advantage of this invention over existing technologies lies in its ability to address the problem of loops caused by inter-segment citations when parsing large-scale academic papers. By explicitly representing and controllably rewriting the internal citation relationships within the paper, the parsing process has a clear termination boundary. By extracting citation links between segments to construct a citation dependency graph and detecting closed-loop paths composed of multiple nodes, the citation loops causing the loops can be located at the structural level. For closed-loop paths, a starting node is selected and copied to form an expanded surrogate node. The citation link between the ending node and the starting node is severed and changed to point to the expanded surrogate node, transforming the closed loop into an open chain in the graph structure. This prevents the parsing order from repeatedly expanding back to the original starting point along citation relationships.
[0016] Furthermore, this invention not only processes loop closures on the graph but also applies the processing results to the paper text itself. Based on the new version of the paper's citation dependency graph, the citation structure of the research paper is modified. The segment content corresponding to the expanded surrogate node is copied and inserted into the research paper to form a new segment. The citation content in the last segment of the corresponding closed-loop path in the research paper, which originally pointed to the segment corresponding to the starting node, is replaced with citation content pointing to the new segment. Simultaneously, the citation content of the new segment pointing to subsequent nodes in the corresponding closed-loop path is deleted or masked. This rewrites the closed-loop citation into a terminateable citation chain at the paper text level, preventing equivalent copies from continuing to cite outwards along the original closed-loop path and forming a loop again. The resulting modified research paper has a citation structure without closed-loop citations. When any general-purpose large language model or multimodal large model parses based on this paper, it can avoid the repeated construction of loop contexts without relying on specific loop detection strategies within the model, improving the stability and consistency of the parsing process.
[0017] Furthermore, the selection rule for the starting node prioritizes intersecting nodes with a larger number of participating loops in multi-loop intersection scenarios. It also allows multiple intersecting loops to share the same expanded proxy node and newly added segment corresponding to the same intersecting node, enabling multiple loop paths to be uniformly cut at the same newly added segment, reducing structural disturbances caused by the number of newly added segments and repeated insertions. In the absence of intersections, the node with fewer external connections is selected as the starting node, reducing the impact of rewriting on external reference relationships. Employing the Tarjan algorithm and its depth-first traversal, access sequence number, minimum traceable sequence number, and node stack mechanism, it can stably identify strongly connected structures and locate loop paths. The citation identifiers of newly added segments are processed by deletion, replacement, or removal from the citation field, adapting to different paper formats and citation annotation systems, ensuring that the modified paper maintains terminating parsing while facilitating engineering implementation and deployment. Attached Figure Description
[0018] Figure 1 This is the overall flowchart of the present invention; Figure 2This is a schematic diagram of the closed-loop path of the present invention; Figure 3 This is a schematic diagram of the present invention modifying the closed-loop path into a chain; Figure 4 This is a modified schematic diagram of the paper of this invention. Detailed Implementation
[0019] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.
[0020] like Figure 1 As shown, this invention uses a large model to process the internal citation relationships of scientific papers, thus avoiding infinite loops in the parsing process. The method first uses a large model to extract citation links between sections of a scientific paper and constructs a citation dependency graph. The nodes in the graph correspond to sections of the paper, such as the abstract, methods, results, and discussion sections, while edges represent citation links between these sections. For example, a statement in the methods section mentioning "see results" forms an edge from methods to results. This design aims to transform the implicit logical structure of the paper into an explicit graph representation, facilitating the subsequent detection and handling of potential loops. This is because scientific papers often strengthen their arguments through mutual citations, but this can cause the large model to repeatedly traverse the same content during recursive context expansion, consuming computational resources and affecting output stability.
[0021] In one embodiment, when extracting citation links, the large model can scan the paper text using natural language processing techniques to identify keywords such as "see details," "based on," or "reference," and determine the correspondence between source and target segments by combining contextual semantics. For example, the large model can segment the paper into segment units and then use embedding vectors to calculate the similarity between segments to assist in link extraction and improve accuracy. In another embodiment, the extraction process can also integrate rule matching and machine learning classifiers to annotate citation phrases, thereby handling variations in different paper formats, such as PDF or LaTeX-generated documents.
[0022] After constructing a citation dependency graph, this invention detects closed-loop paths on this graph. For example... Figure 2 As shown, a closed-loop path refers to a circular chain of references consisting of multiple nodes. For example, the method section references the result section, the result section references the discussion section, and the discussion section ultimately points back to the method section, forming a closed loop. This type of closed loop is common in biological or computer science papers for cross-validation of evidence, but it can cause stack overflows during the parsing of large models.
[0023] This invention applies the Tarjan algorithm for detection. This algorithm performs a depth-first traversal of the paper citation dependency graph, assigning an incrementing access number to each visited node and recording the smallest traceable access number for that node. During the traversal, a node stack is maintained to store the nodes on the current path, and a stack marker is set for each node pushed onto the stack. When the traversal returns to a node whose smallest access number equals the current access number, nodes are popped from the stack sequentially until the current node is popped, forming a strongly connected node set. If this set contains at least two nodes, it is determined to be a closed loop path.
[0024] For example, in a graph containing segments A, B, and C, if A points to B, B points to C, and C points to A, Tarjan's algorithm starts traversing from A, assigning a visit index of 1 to A and pushing A onto the stack; then to B, index 2, pushing B onto the stack; then to C, index 3, pushing C onto the stack; backtracking from C to A, it finds that the minimum visit index of A is 1, which is equal to its visit index, so C, B, and A are popped, forming a strongly connected set containing 3 nodes, confirming a closed loop. The purpose of this algorithm is to efficiently identify all strongly connected components in a graph.
[0025] Upon detecting a closed-loop path, this invention determines the starting node of the closed-loop path and creates an expanded proxy node by copying this starting node. This expanded proxy node is an equivalent copy of the starting node, including copying its segment content and semantic information. Then, the reference links from the ending nodes in the closed-loop path to the starting node are severed, and reference links from the ending nodes to the expanded proxy node are established, thereby converting the closed-loop path into an open chain. Figure 3 As shown, this transformation turns the original circular structure into a linear sequence, avoiding repeated loops along references in the large model.
[0026] In one embodiment, if a closed-loop path intersects with other closed-loop paths, the intersecting node is identified as a candidate starting node. The candidate starting node with the most participants in the closed-loop paths is then selected as the starting node. This selection reduces processing redundancy across multiple closed loops and improves efficiency. For example, if two closed loops share a node that participates in two paths, it is prioritized. In another embodiment, if there are no intersecting nodes, the node with the fewest connections to external nodes in the closed-loop path is selected as the starting node. This design minimizes the impact of rewriting on the overall structure of the paper and avoids interference with external citations.
[0027] Furthermore, when a candidate starting node is determined to be the starting node of multiple intersecting closed-loop paths, an expanded surrogate node is created only for that candidate starting node. The newly added segment corresponding to the expanded surrogate node is then used as a shared newly added segment for the multiple intersecting closed-loop paths. Specifically, for each intersecting closed-loop path, the reference content in the last segment of that closed-loop path that originally pointed to the segment corresponding to the candidate starting node is replaced with a reference content pointing to the shared newly added segment. This sharing mechanism aims to optimize resource usage and reduce unnecessary node duplication, especially in complex papers where multiple closed loops overlap, thus preventing excessive paper size expansion.
[0028] By replacing all closed-loop paths with corresponding open chains in the paper citation dependency graph, a new version of the paper citation dependency graph is formed. This invention modifies the citation structure of research papers based on this new citation dependency graph, generating modified research papers.
[0029] like Figure 4 As shown, the modification process includes copying and inserting the segment content corresponding to the expanded proxy node into the research paper to form a new segment. It also replaces the original reference content pointing to the segment corresponding to the starting node in the last segment of the corresponding closed-loop path in the research paper with reference content pointing to the new segment. Simultaneously, it deletes or masks the reference content of the new segment pointing to subsequent nodes in the corresponding closed-loop path. This modification ensures that the new segment does not introduce a new loop because the deleted reference content includes the reference content in the new segment's external reference content to the original segment corresponding to the starting node, which points to the next node in the closed-loop path. This next node is a node directly reachable from the starting node along the reference link in the closed-loop path.
[0030] In one embodiment, deletion can be achieved by directly removing the quoted phrase; in another embodiment, masking can be handled by annotation or hidden tags to preserve the original meaning but prevent large models from following. This multi-implementation approach adapts to different paper formats, such as using deletion in Word documents and masking in HTML versions, with the aim of completely breaking potential loop paths, making the modified paper suitable for large model parsing.
[0031] Furthermore, prompt words are input into the large model, which then parses the modified research paper based on the prompt words and outputs the results. The large model in this invention is a general-purpose large language model, such as the Doubao large model, the LLaMA series large models, or the GPT series large models. These models employ a Transformer architecture, including multi-layer self-attention mechanisms and feedforward networks, with parameter sizes ranging from 7 billion to 175 billion; for example, GPT-3.5 has 175 billion parameters. They are obtained through massive pre-training on text, with the training process involving unsupervised learning to predict the next token. The optimization function is cross-entropy loss, the batch size is typically 512 to 2048, and the learning rate ranges from 1e-5 to 1e-3. During parsing, the temperature parameter can be set to 0.7 to 1.0 to balance generation diversity and determinism; the context window length is 2048 to 128000 tokens, depending on the model version. This setting aims to efficiently process the modified linear structure using pre-trained knowledge, avoiding context explosion caused by the original closed loop. In one embodiment, the prompt includes instructions for parsing the tree structure; in another embodiment, the prompt specifies the extraction of key innovations, thereby outputting a complete parsing result.
[0032] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for analyzing the structure of scientific research papers based on a large model, characterized in that, Includes the following steps: The large model extracts the citation links between segments in scientific research papers and constructs a citation dependency graph, where the nodes of the citation dependency graph are the segments of the scientific research paper and the edges of the citation dependency graph are the citation links between the segments in the scientific research paper. Closed-loop paths are detected on the paper citation dependency graph. When a closed-loop path is detected, the starting node of the closed-loop path is determined, and the starting node is copied to create an expanded proxy node, which is an equivalent copy of the starting node. Cut the reference link from the end node in the closed loop to the starting node, and establish the reference link from the end node to the expanded proxy node, so as to convert the closed loop into an unclosed chain. A new version of the paper citation dependency graph is formed by replacing all closed loop paths with corresponding open chains on the paper citation dependency graph; the citation structure of the research paper is modified according to the new version of the paper citation dependency graph to generate the modified research paper; prompt words are input into the large model, and the large model parses and outputs the results based on the modified research paper according to the prompt words.
2. The method for analyzing the structure of scientific papers based on a large model according to claim 1, characterized in that, The process of modifying a research paper includes: copying and inserting the segment content corresponding to the expanded proxy node into the research paper to form a new segment; replacing the original reference content pointing to the segment corresponding to the starting node in the last segment of the corresponding closed-loop path in the research paper with the reference content pointing to the new segment; and deleting or blocking the reference content of the new segment pointing to the successor node in the corresponding closed-loop path.
3. The method for analyzing the structure of scientific papers based on a large model according to claim 2, characterized in that, The newly added segment refers to the reference content of the successor node in the corresponding closed-loop path, including the external reference content of the original segment corresponding to the starting node in the newly added segment, and the reference content used to point to the next node of the starting node in the closed-loop path. The next node is a node in the closed-loop path that can be directly reached by the starting node along the reference link.
4. The method for analyzing the structure of scientific papers based on a large model according to claim 1, characterized in that, When a closed-loop path is detected, if the closed-loop path intersects with other closed-loop paths, the intersecting nodes are identified as candidate starting nodes, and the candidate starting node with the most participants in the closed-loop path is selected as the starting node.
5. The method for analyzing the structure of scientific papers based on a large model according to claim 1 or 4, characterized in that, When a closed-loop path is detected, if the closed-loop path does not intersect with any other closed-loop path, the node with the fewest connections to external nodes of the closed-loop path is selected as the starting node.
6. The method for analyzing the structure of scientific papers based on a large model according to claim 4, characterized in that, When the candidate starting node is determined to be the starting node of multiple intersecting closed loop paths, an expanded proxy node is created only for the candidate starting node, and the newly added segment corresponding to the expanded proxy node is used as a shared newly added segment for multiple intersecting closed loop paths; wherein, for each intersecting closed loop path, the reference content in the last segment of the closed loop path that originally pointed to the segment corresponding to the candidate starting node is replaced with the reference content that points to the shared newly added segment.
7. The method for analyzing the structure of scientific papers based on a large model according to claim 1, characterized in that, The Tarjan algorithm is applied to detect closed-loop paths on the paper citation dependency graph.
8. The method for analyzing the structure of scientific papers based on a large model according to claim 7, characterized in that, The Tarjan algorithm includes: performing a depth-first traversal on the paper citation dependency graph, assigning an incrementing access number to each visited node, and recording the traceable minimum access number for each visited node; maintaining a node stack during the depth-first traversal to store nodes on the current traversal path, and setting a stack marker for nodes pushed onto the stack; when the traversal returns to a node and the minimum access number of that node is equal to its access number, popping nodes from the node stack sequentially until the node is popped to form a strongly connected node set, and determining it as a closed loop path when the strongly connected node set contains at least 2 nodes.
9. The method for analyzing the structure of scientific papers based on a large model according to claim 1, characterized in that, The large model is a general-purpose large language model, including any one of the following: the Doubao large model, the LLaMA series large models, and the GPT series large models.
10. A large-model-based scientific paper structure parsing system for implementing the method of claim 1, characterized in that, include: The citation link extraction unit is used to extract citation links between sections in scientific papers from a large model. A dependency graph construction unit is used to construct a paper citation dependency graph based on the citation links, wherein the nodes of the paper citation dependency graph are segments of the research paper, and the edges of the paper citation dependency graph are citation links between segments in the research paper. A closed-loop path detection unit is used to detect closed-loop paths on the paper citation dependency graph. The closed-loop unpacking unit is used to determine the starting node of the closed-loop path when a closed-loop path is detected, copy the starting node to create an unpacking proxy node, the unpacking proxy node being an equivalent copy of the starting node, and cut the reference link from the end node in the closed-loop path to the starting node, and establish the reference link from the end node to the unpacking proxy node, so as to convert the closed-loop path into an unclosed chain. The new dependency graph generation unit is used to replace all closed-loop paths with corresponding open chains in the paper citation dependency graph to form a new paper citation dependency graph. The paper citation structure modification unit is used to modify the citation structure of the research paper according to the new version of the paper citation dependency graph, and generate the modified research paper. The parsing output unit is used to input prompt words into the large model, so that the large model can parse the modified scientific paper based on the prompt words and output the results.