An academic analysis method and processing system based on a large language model

Abstract

The application discloses a kind of academic analysis method and processing system based on large language model.The method comprises the following steps: obtaining multi-source academic data to construct academic hypergraph, generating high-order semantic representation by aggregating features through hypergraph neural network, inputting large language model adjusted by artificial intelligence and adapting academic scene to output preliminary analysis result;Fact anchoring verification, multi-source cross verification, illusion filtering and academic specification verification supported by artificial intelligence algorithm realize credible and anti-illusion processing;Output structured results by adapting user demand, optimize hypergraph and model parameters based on user feedback, and determine reliability level by combining three types of quantitative indicators of artificial intelligence online evaluation.The system improves the accuracy, credibility and adaptability of academic analysis through hypergraph modeling, multiple verification and closed-loop optimization, and provides an efficient and reliable artificial intelligence analysis tool for scientific research.

Description

Technical Field

[0001] This application relates to the interdisciplinary fields of natural language processing, knowledge graphs, artificial intelligence, and academic data analysis, specifically to an academic analysis method and processing system based on a large language model. Background Technology

[0002] With the rapid development of information technology, global academic data is experiencing explosive growth. Multiple sources of academic resources, including papers, patents, and experimental data, are continuously accumulating, and the analysis results have become a key decision-making basis for core tasks such as research project initiation, technological innovation, and achievement evaluation. The results of academic data analysis must simultaneously meet the requirements of knowledge completeness, conclusion credibility, and academic compliance; otherwise, they will directly affect the research direction, innovation quality, and the rigor of the results.

[0003] Currently, academic analysis often relies on traditional tools or simply large language models to generate results, which has significant limitations: First, knowledge integration depends on single data types or superficial correlation analysis, lacking effective AI-driven modeling for multimodal information and complex relationships, making it difficult to form a global knowledge view; second, it only focuses on model generation capabilities or statistical correlations, lacking AI-supported data source anchoring and logical verification of conclusions, making it impossible to avoid the risk of illusions and false associations; third, it does not adequately consider the adaptability to academic scenarios and dynamic data updates, failing to leverage the adaptive learning characteristics of AI, resulting in weak model generalization ability and difficulty in guaranteeing analytical accuracy and timeliness.

[0004] Therefore, there is an urgent need for an academic analysis method that can deeply integrate multi-source academic data and complex relationships, build a reliable verification mechanism based on artificial intelligence, and enhance the academic adaptability and dynamic optimization capabilities of the model. Summary of the Invention

[0005] The main objective of this application is to provide an academic analysis method and processing system based on a large language model, aiming to overcome the shortcomings of existing technologies, such as difficulties in integrating academic knowledge, high risk of illusion in analysis results, insufficient adaptability to academic scenarios, and lack of closed-loop optimization mechanisms. This application constructs an academic hypergraph model of multi-source academic data, combines hypergraph neural network feature aggregation with AI-driven large language model adaptation to academic scenarios, and introduces a dual processing mechanism of credibility and anti-illusion supported by AI. The system can accurately generate, reliably verify, and dynamically optimize academic analysis results, ultimately outputting authoritative, consistent, and compliant academic analysis outcomes.

[0006] To achieve the above objectives, the technical solution adopted in this application is as follows: An academic analysis method based on a large language model, characterized in that the method includes: S1: Acquire multi-source academic data and construct an academic hypergraph; S2: Based on the academic hypergraph, a high-order semantic representation is generated by aggregating hypergraph features through a hypergraph neural network. The semantic representation is then input into a large language model that has been fine-tuned by artificial intelligence and adapted to academic scenarios, and preliminary academic analysis results are output. S3: Based on the preliminary academic analysis results, perform both credibility enhancement and anti-hallucination processing to obtain the processed analysis results; S4: Based on the processed analysis results, output structured results to meet the user's academic analysis needs, and optimize the parameters of the academic hypergraph and the large language model based on user feedback.

[0007] Optionally, acquiring multi-source academic data and constructing an academic hypergraph includes: collecting papers, patents, and experimental data from academic databases, industry reports, and experimental data to generate a multi-source academic data set; parsing the multi-source academic data set to extract five types of core nodes: literature, keywords, experimental data, citations, and technical features; connecting the five types of core nodes with hyperedges to construct an academic hypergraph, while simultaneously using multimodal parsing technology to integrate text, charts, and formula semantics to configure structured feature vectors for hypergraph nodes; configuring an incremental update mechanism for the hypergraph to incorporate the latest academic data and synchronously update the hypergraph nodes, hyperedge relationships, and feature vectors.

[0008] Optionally, the preliminary academic analysis results generated based on the academic hypergraph include: based on the academic hypergraph, aggregating the node features associated with the hyperedges through hypergraph convolutional layers to generate a high-order semantic representation; using cross-disciplinary terminology corpus to perform AI-supervised fine-tuning on a general large language model to obtain a large language model adapted to academic scenarios; inputting the high-order semantic representation into the large language model, and combining it with the user's academic analysis requirements to output preliminary analysis results, wherein the academic analysis requirements include trend analysis, comparison of innovative points, and literature review writing.

[0009] Optionally, the dual processing of credibility and anti-hallucination includes fact anchoring verification: associating each conclusion of the preliminary analysis results with the academic hypergraph node, and labeling the corresponding original text fragments, citation nodes and experimental data nodes one by one; establishing a traceable relationship between the conclusion and the academic hypergraph node generated by artificial intelligence algorithm, and each conclusion is associated with an authoritative data source.

[0010] Optionally, the dual processing of credibility and anti-illusion also includes multi-source cross-validation: utilizing the multi-node association characteristics of the academic hypergraph, node information of different hyperedge associations is extracted; comparing the node information, the consistency between literature conclusions and cited experimental data, and between technical features and patent claims is verified, and an artificial intelligence consistency detection algorithm automatically detects logical conflicts and marks the conflict locations, generating a conflict analysis description.

[0011] Optionally, the dual processing of credibility and anti-hallucination also includes hallucination filtering: calculating the similarity between the preliminary academic analysis results and the higher-order semantic representation, and setting a similarity threshold; marking content below the threshold as high-risk hallucinations, automatically retrieving the academic hypergraph to supplement alternative data sources, and driving the large language model to correct the conclusions based on the node features and association information of the academic hypergraph.

[0012] Optionally, the dual processing of credibility and anti-hallucination also includes academic norm verification: calling the subject norm nodes stored in the academic hypergraph, the nodes containing the expression paradigms, terminology standards and citation norms of different disciplines; verifying the compliance of the expression and the accuracy of the terminology of the preliminary analysis results sentence by sentence, generating norm correction suggestions and optimizing the expression of the results.

[0013] Optionally, the process of adapting to users' academic analysis needs and outputting structured results and reverse optimization includes: based on the processed analysis results, outputting structured results to adapt to users' academic analysis needs; providing an interactive correction interface to receive user feedback; converting the user feedback content into optimization signals using artificial intelligence online reinforcement learning technology, and inversely updating the node weights and hyperedge relationships of the hypergraph, while simultaneously adjusting the parameters of the large language model.

[0014] Optionally, the method further includes determining the reliability level of the result: a preset reliability level range is defined, which is divided into four levels: excellent, good, qualified, and unqualified; based on the verification results of the credibility and anti-hallucination processing described in step S3, combined with the degree of support from authoritative data sources, the results of logical consistency verification, and the degree of compliance with academic norms, the artificial intelligence comprehensive scoring engine determines the reliability level corresponding to the processed analysis result; the analysis result, reliability level, and correction basis are output simultaneously, and the correction basis includes fact-anchored authoritative data source information, logical conflict correction instructions, and academic norm optimization records.

[0015] This application also provides an academic analysis and processing system based on a large language model, used to implement any of the above methods, the system comprising: Data acquisition module: used to collect papers, patents and experimental data from academic databases, industry reports and experimental data to generate multi-source academic data sets; Hypergraph construction module: used to parse the multi-source academic data set, extract five types of core nodes, construct an academic hypergraph by connecting core nodes with hyperedges, configure structured feature vectors, and perform incremental updates of the hypergraph; Hypergraph Neural Network Analysis Module: Used to generate high-order semantic representations based on the aggregated features of the academic hypergraph, driving the large language model fine-tuned by artificial intelligence to output preliminary academic analysis results; Trustworthy anti-hallucination processing module: used to perform fact anchoring verification, multi-source cross-verification, hallucination filtering and academic norm verification supported by artificial intelligence algorithms on the preliminary academic analysis results; The results output and optimization module is used to output structured results based on the processed analysis results, receive user feedback, and reverse-optimize the parameters of the hypergraph and large language model to output complete analysis results including the reliability level evaluated online by artificial intelligence.

[0016] This application proposes an academic analysis method and processing system based on a large language model. It achieves deep integration of multi-source academic knowledge through academic hypergraph modeling, enhances analysis and generation capabilities by combining hypergraph neural networks with an AI-adapted large language model for academic scenarios, ensures result credibility through an AI-driven four-fold verification mechanism, and achieves continuous iteration through user feedback and AI closed-loop optimization. This effectively overcomes the shortcomings of traditional academic analysis methods, such as weak knowledge integration, high risk of illusion, insufficient adaptability, and lack of dynamic optimization. It provides a highly efficient, accurate, and reliable academic analysis tool for scientific research, significantly improving the efficiency and quality of academic research. Attached Figure Description

[0017] Figure 1 A flowchart illustrating an academic analysis method based on a large language model, provided as an embodiment of this application; Figure 2 A block diagram of an academic analysis and processing system based on a large language model, provided as an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application.

[0018] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are for illustrative purposes only and are not intended to limit the scope of this application.

[0020] The main solution of this application embodiment is: an academic analysis method based on a large language model, which achieves in-depth and reliable analysis of multi-source academic data through a series of steps, including multi-source academic data acquisition and academic hypergraph construction, hypergraph feature aggregation and artificial intelligence adaptation to generate preliminary analysis results, artificial intelligence-supported credibility and anti-illusion dual processing, structured result output and closed-loop optimization.

[0021] Due to the limitations of existing technologies, such as weak knowledge integration capabilities, high risk of illusory analysis results, insufficient adaptability to academic scenarios, and lack of closed-loop optimization mechanisms, the academic analysis method provided in this application can integrate complex multi-source associations through hypergraph modeling, ensure the credibility of results through an AI-driven multi-verification mechanism, improve the academic adaptability of the model through AI-supervised fine-tuning, and achieve dynamic optimization through user feedback and AI reinforcement learning, thereby meeting the needs of scientific research for efficient, accurate, and credible academic analysis.

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can understand it.

[0023] Reference Figure 1 This application provides an embodiment of an academic analysis method based on a large language model, the implementation steps of which are as follows: S1. Acquire multi-source academic data and construct an academic hypergraph.

[0024] In this embodiment of the application, the multi-source academic data set is a comprehensive set of academic resources collected from academic databases, industry report repositories, and experimental data storage systems. It covers full texts of papers, patent specifications, original experimental records, and industry technical analysis reports, and includes multimodal information such as text, charts, formulas, and data tables.

[0025] In practice, standardized interfaces are used to connect to various data sources, supporting the access and parsing of PDF, TXT, DOCX, CSV, PNG, and JPG formats. Data collection covers basic research papers, applied technology patents, clinical trial data, and industry standard documents, extracting core fields such as document titles, author information, publication dates, keywords, abstracts, experimental parameters, citation lists, and technical feature descriptions. After data cleaning to remove duplicates, blanks, and corrupted files, each data entry is verified to contain a unique document identifier, core technical parameters, experimental condition descriptions, or research conclusions, generating a well-structured and complete multi-source academic data set.

[0026] This study deeply analyzes multi-source academic datasets to extract five core node categories: literature, keywords, experimental data, citations, and technical features. The literature node includes attributes such as title, author, and publication date; the keyword node includes the terminology text and its subject area; the experimental data node includes the experimental subjects, conditions, and raw data attributes; the citation node includes the cited literature identifier and citation location attributes; and the technical feature node includes feature descriptions and technical field attributes. Together, these elements construct a basic unit system of academic knowledge.

[0027] An academic hypergraph H=(V,E,X) is constructed by connecting five types of core nodes with hyperedges, where V is the set of nodes (the union of the five types of core nodes), E is the set of hyperedges, and X is the node feature matrix. Each hyperedge can connect multiple nodes of different types, capturing the complex intrinsic relationships between multiple nodes and overcoming the limitation of traditional knowledge graphs that can only model binary relationships. For example, a hyperedge can simultaneously connect the literature nodes of a paper, the 35 core keyword nodes of the paper, the two key citation nodes, the experimental data nodes supporting the core conclusions, and the technical feature nodes corresponding to the core technologies, forming a complete academic knowledge unit that comprehensively reflects the multi-dimensional relationships of the research content.

[0028] To ensure the richness and representational power of node features, a multimodal parsing technique is employed to fuse text, chart, and formula semantics, configuring structured feature vectors for hypergraph nodes. For textual data, high-dimensional semantic features are extracted using a pre-trained language model to generate text feature vectors v. text For chart data, image recognition technology is used to extract visual features, and OCR technology is combined to extract numerical and textual information from the charts, which are then fused to generate a visual feature vector v. visual For formula data, mathematical symbols, operational logic, and variable relationships are extracted using formula parsing tools to generate a formula feature vector v. formula The three types of feature vectors are weighted and fused to generate a structured feature vector with uniform dimensions, as shown in the formula: in, Let i be the structured feature vector of the i-th node. , These are the text feature vector, visual feature vector, and formula feature vector of the node, respectively. , and To integrate weights, satisfy + + =1. Based on the characteristic distribution of academic data, the preferred values ​​are 0.6, 0.2 and 0.2, which highlight the core role of textual semantics and make full use of the supplementary information from charts and formulas.

[0029] To ensure the timeliness and completeness of the academic hypergraph, an incremental update mechanism is configured. A fixed update cycle or trigger-based update conditions are set to automatically incorporate the latest published academic data. During the update process, the following operations are performed simultaneously: First, new core nodes corresponding to eligible academic data are added, supplementing node attribute information; second, based on the relationship between new nodes and existing nodes, corresponding hyperedges are added to expand the hypergraph's relational dimensions; third, according to data changes, the structured feature vectors of relevant nodes are updated, and node weights are adjusted to ensure the hypergraph dynamically reflects the development and changes in academic research.

[0030] In a specific example, 3,000 core journal articles, 500 related patents, and 200 sets of publicly available experimental data in the field of computer vision from 2020 to 2025 were collected. After parsing, 3,000 document nodes, 8,000 keyword nodes, 200 experimental data nodes, 6,000 citation nodes, and 1,200 technical feature nodes were extracted. These were then linked by hyperedges to form an academic hypergraph containing 18,400 nodes and 35,000 hyperedges, with each node configured with a 512-dimensional structured feature vector. Through an incremental update mechanism, the latest published papers and patent data in the field are automatically incorporated monthly, with an average of approximately 500 new nodes and 1,200 new hyperedges added each month, continuously maintaining the timeliness of the hypergraph and providing comprehensive and up-to-date knowledge support for subsequent analysis.

[0031] S2. Based on the academic hypergraph, the hypergraph features are aggregated through a hypergraph neural network to generate a high-order semantic representation, which is then input into a large language model that has been fine-tuned by artificial intelligence and adapted to academic scenarios, and the preliminary academic analysis results are output.

[0032] Based on the academic hypergraph H, the node features associated with hyperedges are aggregated through the hypergraph convolutional layers of a hypergraph neural network to generate a high-order semantic representation that can capture global relational information. The specific computation process consists of three steps: The first step is to calculate the hypergraph Laplacian matrix. First, construct the hypergraph adjacency matrix A, where A(i,j) indicates whether node i and node j are connected by a hyperedge (1 for connected, 0 for otherwise). Then, calculate the degree matrix D, a diagonal matrix, where D(i,i) represents the number of hyperedges connected to node i. Finally, normalize the matrix to obtain the hypergraph Laplacian matrix. , where I is the identity matrix, ensuring the numerical stability of the feature aggregation process.

[0033] The second step is to perform hypergraph convolutional feature aggregation. The hypergraph node feature matrix X is input into the hypergraph convolutional layer, and matrix operations are used to achieve the association and aggregation of features from multiple nodes. The formula is as follows: , where W is a learnable weight matrix used to adjust the feature dimension and representational power, σ is the ReLU activation function, which enhances the expressive power of the model through nonlinear transformation and generates an intermediate feature matrix H'.

[0034] The third step is to generate a global high-order semantic representation. Global average pooling is performed on the intermediate feature matrix H', fusing the aggregated features of all nodes to obtain the high-order semantic representation H of the entire academic hypergraph. high The calculation method is as follows , where n is the total number of nodes in the academic hypergraph. Let be the convolutional feature vector of the i-th node. The preferred dimension is 512 or 768, which ensures both the richness of semantic information and the control of computational complexity, and can fully represent the global correlation and core semantics of multi-source academic data.

[0035] To enhance the model's adaptability to academic scenarios, a cross-disciplinary terminology corpus was used to perform AI-supervised fine-tuning of the general-purpose large language model, resulting in a large language model adapted to academic scenarios and improving the model's ability to understand and generate academic texts. The cross-disciplinary terminology corpus covers multiple fields including natural sciences, engineering technology, and social sciences, with a total size of no less than 100GB. Specifically, it includes full-text articles from core journals, patent specifications, subject-specific dictionaries, academic writing guidelines, and research question-and-answer datasets. This cross-disciplinary terminology corpus was divided into training, validation, and test sets in an 8:1:1 ratio. The training set was used for model parameter optimization, the validation set for monitoring training effectiveness and early stopping detection, and the test set for final performance evaluation. The three datasets maintained a consistent distribution in terms of subject distribution, text type, and task scenario.

[0036] The training process focuses on academic text comprehension, professional question answering, and academic content generation as its core tasks. Supervised fine-tuning is employed to optimize model parameters. The specific training configuration is as follows: a learning rate of 5e-5, linear decay scheduling, with the last 10% of training steps decreasing to 1e-6, a batch size of 8, gradient accumulation every two micro-batches, three training epochs, and the AdamW optimizer with a weight decay coefficient of 0.01. Gradient descent is used to minimize the model's prediction error on academic tasks. During training, academic terminology accuracy, logical consistency score, and sentence structure standardization are used as evaluation metrics. An early stopping strategy is implemented, stopping the training if the validation set metrics show no improvement after three consecutive training epochs. Simultaneously, the focus is on optimizing the model's accurate understanding of professional terminology, precise grasp of academic logic, and standardized expression of complex sentences, enabling the model to adapt to the specific needs of academic analysis.

[0037] High-order semantic representation H high The system takes a large language model adapted to academic scenarios as input and, combined with the user's academic analysis needs, generates preliminary academic analysis results. The user's academic analysis needs are clearly stated through model input prompts, primarily including three core scenarios: trend analysis, innovation point comparison, and literature review writing. The prompts for different scenarios are specifically designed to adapt to the model's generation logic, including: Regarding the demand for trend analysis, the prompt explicitly requires the model to analyze the development history, current hotspots and future trends of a research field based on information such as keyword evolution, publication time distribution, technical feature iteration and citation relationship changes in high-order semantic representation, and output a trend report that includes time stage division, core direction evolution and key node events. For the comparison of innovation points, the prompt guides the model to compare the technical feature nodes, experimental data nodes and core conclusion information of different literature and patents, extract the innovation points, differences in technical paths, advantages and disadvantages of each study, and generate structured comparative analysis results. Regarding the requirements for writing literature reviews, the prompt requires the model to integrate multi-source correlation information in the hypergraph, sort out the core issues, research context, mainstream technical routes, key research results and existing challenges in the research field, and generate a review text that conforms to academic norms.

[0038] Based on structured knowledge in high-order semantic representation and its own adaptability to academic scenarios, the model outputs preliminary academic analysis results containing elements such as research context, cross-domain connections, core conclusions, and data support. The result format is adaptively adjusted according to user needs, and can be structured reports, key point lists, coherent text, etc. The full text length is controlled between 3,000 and 8,000 words according to needs to ensure the depth and readability of the analysis.

[0039] In a specific example, the academic hypergraph in the field of computer vision constructed above is used to generate a 1024-dimensional high-order semantic representation H through hypergraph convolutional layers. high The user's input is a large language model fine-tuned under AI supervision using cross-disciplinary scientific terminology corpus. The user's request is to analyze the research trends and innovations of deep learning in image recognition from 2020 to 2025. Based on information such as keyword evolution, publication time distribution, technological feature iteration, and citation relationship changes in high-order semantic representation, the model outputs preliminary analysis results including development trends, comparisons of core innovations, and future prospects. The full text is approximately 6000 words, covering 20 core conclusions, each supported by implicitly linked nodes in a hypergraph.

[0040] S3. Based on the preliminary academic analysis results, perform AI-driven credibility enhancement and anti-hallucination dual processing to obtain the processed analysis results.

[0041] The dual process of credibility enhancement and anti-hallucination is the core link in ensuring the reliability of academic analysis results. Through the coordinated processing of four links—fact anchoring verification, multi-source cross-verification, hallucination filtering, and academic norm verification—errors, unfounded conclusions, and non-standard expressions in the preliminary analysis results are completely eliminated, so that the processed analysis results have authority, consistency, and compliance.

[0042] The core objective of the fact-anchoring verification stage is to associate each conclusion of the preliminary analysis results with authoritative data sources and establish a traceable supporting relationship. Each conclusion of the preliminary analysis results is associated with an academic hypergraph node, and the corresponding original text fragment, citation node, and experimental data node are labeled one by one. The similarity between the conclusion and the feature vector of the hypergraph node is calculated using an artificial intelligence semantic similarity matching algorithm, with the following formula: Here, 'a' represents the semantic vector of the conclusion, and 'b' represents the feature vector of the hypergraph node. Nodes with the highest similarity exceeding the 80% threshold are designated as associated nodes. A traceable relationship is established between the conclusions and hypergraph nodes, generated by an artificial intelligence algorithm. This ensures that each conclusion is associated with at least one authoritative data source node, such as nodes from core journal papers, highly cited patents, or peer-reviewed experimental data, thereby guaranteeing the credibility of the conclusions from the source.

[0043] The multi-source cross-validation step utilizes the multi-node association characteristics of academic hypergraphs to extract node information associated with different hyperedges, and compares the node information to verify the consistency between literature conclusions and cited experimental data, and between technical features and patent claims.

[0044] An artificial intelligence logical consistency detection algorithm is used to construct a set of logical predicates for conclusions and node information. This is first done through formulas. Complete the formal expression of the predicate, where P is the logical predicate, T is the predicate type, A is the set of nodes in the associated hypergraph, K is the core numerical or semantic value, and S is the confidence score based on node authority.

[0045] Then, the conflict determination formula is used to identify contradictory relationships. The formula is as follows: , in, The result of the conflict determination can only be 0 or 1. When the result is 1, it is determined that there is a conflict between the two sets of predicates, and further analysis of the conflict intensity and location is required; when the result is 0, it is determined that there is no contradiction between the two sets of predicates, and the information can be considered consistent, without the need for additional conflict handling. a and P b These are two sets of logical predicates that are to be determined to be in conflict. They are predicates P a and P b Predicate types, Cosine similarity is used to calculate the semantic relevance of predicate types. For core value deviation, numerical data is categorized as follows: Calculate the relative deviation; for semantic data, assign a value of 1 or 0 based on whether they are consistent. The preset conflict threshold is set according to the predicate type and the domain prior statistics. It can be updated online. The accuracy class is set to 0.1, the technology affiliation class is set to 0, and the parameter range class is set to 0.2.

[0046] The severity of the conflict was then assessed using a conflict intensity quantification formula, which is: in, The result is used to quantify the conflict intensity, with a value range of [0,1]. 0.0-0.3 represents mild conflict, 0.3-0.6 represents moderate conflict, and 0.6-1.0 represents severe conflict. The larger the value, the more severe the conflict. They are predicates P a and P b Confidence score based on node authority.

[0047] The specific location of a conflict in the analysis results is marked using the conflict location formula, which is: in, As a unique identifier for associated nodes, This refers to text location information.

[0048] Through the aforementioned artificial intelligence algorithm, logical conflicts are automatically detected and conflict analysis explanations are generated, clarifying the type and source of the conflict. For example, if a literature conclusion states that a certain technology has an accuracy rate of 95%, but the corresponding experimental data in the citation node shows that the actual accuracy rate of the technology is 80%, the system can calculate the actual accuracy rate using a formula. The system identifies conflicts, calculates their intensity, and marks their locations to provide a basis for subsequent corrections.

[0049] The hallucination filtering step calculates the mean cosine similarity between the preliminary academic analysis results and the higher-order semantic representations, with a preset similarity threshold of 85%, which was optimally determined using F1 on the validation set. The similarity calculation employs an artificial intelligence semantic cosine similarity algorithm. The model converts the preliminary analysis results into a semantic vector 'a', which is then compared with the higher-order semantic representation vector 'b' for similarity calculation. The formula is as follows: Content below a certain threshold is marked as a high-risk illusion. The system automatically retrieves alternative data sources from the academic hypergraph, prioritizing authoritative node information such as core journal articles and highly cited patents. This drives the large language model to correct its conclusions based on the node features and association information of the hypergraph, ensuring that the corrected conclusions are consistent with the knowledge in the hypergraph.

[0050] The academic standardization verification process utilizes subject-specific standardization nodes stored in the academic hypergraph. These nodes contain expression paradigms, terminology standards, and citation norms for different disciplines. An AI-powered terminology matching and paradigm comparison algorithm is employed to verify the compliance and terminology accuracy of the preliminary analysis results sentence by sentence. This detects non-standard usage, terminology errors, and citation formatting errors, generating standardization correction suggestions and optimizing the final expression. The calculation process includes: The formula for terminology compliance matching is: Where X represents the terms to be verified in the preliminary analysis results, and Y represents the set of standard terms in the subject standardization nodes. The semantic matching degree of terms is calculated using word embedding cosine similarity. The authoritative weight of standard term y is determined by the weighting of core subject terms ≥ 0.9 and general terms 0.6-0.8. When... When the value is less than 0.8, it is determined that the terminology is not standardized and the correction mechanism is triggered.

[0051] Normal form consistency check formula: Where Q is the sentence structure vector of the expression to be verified. Let k be the standard paradigm structure vector in the subject-specific norm nodes, where k is the dimension of the structural features. This is a structural feature similarity function. When... When the value is less than 0.75, it is considered a non-compliant paradigm.

[0052] Citation format compliance scoring formula: Where E represents the citation format to be verified. This is a GB / T 7714 standard format template, where m represents the format validation dimension. This is a consistency indicator function; it returns 1 if the format is consistent and 0 if it is inconsistent. Weights are assigned to each dimension, with core dimensions such as source information having a weight of 0.3 and secondary dimensions such as punctuation format having a weight of 0.1. When If the value is less than 0.85, it is considered an incorrect citation format.

[0053] The aforementioned AI-driven quantitative verification formula can be used to quantitatively verify terminology, paradigms, and citation formats. For example, the non-standard term "deep neural network model" can be corrected to the subject-standard term "deep convolutional neural network model," and citation formats that do not conform to the GB / T 7714 standard can be corrected to the standard format, ensuring that the output results meet the requirements of academic norms.

[0054] In a specific example, the preliminary analysis results contain 20 core conclusions. After fact-anchoring verification, each conclusion is associated with 3-5 hypergraph nodes, and 150 supporting evidences are extracted. After multi-source cross-validation, three logical conflicts were detected by the artificial intelligence consistency detection algorithm: a 30% improvement in the training efficiency of a certain algorithm conflicts with the cited experimental data; a certain technology belongs to the category of deep learning conflicts with the patent claims; and an experimental conclusion conflicts with experimental data from similar studies. Detailed conflict analysis explanations are generated. After illusion filtering, it is calculated that the similarity between three conclusions and higher-order semantic representations is less than 85%, and they are marked as high-risk illusions. Five authoritative data source nodes are added by searching the hypergraph to drive the model to revise the conclusions. After academic standard verification, 20 non-standard terminology issues and 12 citation format errors are corrected, resulting in 20 revised conclusions that conform to academic standards and are supported by authoritative data.

[0055] Based on the above verification and correction results, three core quantitative indicators were further calculated to provide a basis for subsequent reliability level determination. The three core quantitative indicators include: Authoritative data source support level = number of conclusions from validly correlated nodes ÷ total number of conclusions from the processed analysis results × 100%; Logical consistency check result = (Total number of conclusions in the post-processed analysis results, number of unresolved conflicts) ÷ Total number of conclusions in the post-processed analysis results × 100%; Academic norm compliance level = number of statements that conform to academic norms ÷ total number of statements in the post-processed analysis results × 100%.

[0056] S4. Based on the processed analysis results, output structured results to meet users' academic analysis needs, and optimize the parameters of the academic hypergraph and large language model based on user feedback.

[0057] Based on the processed analysis results, the system outputs structured results tailored to users' academic analysis needs, supporting various formats such as structured reports and visualization hypergraphs. The structured reports use a standard academic format, including modules such as abstract, research background, analysis results, conclusions, and references, and can be exported in PDF and DOCX formats. The visualization hypergraph displays the nodes and relationships corresponding to the analysis results through a graphical interface, allowing users to interactively view node details and trace the basis of conclusions.

[0058] To achieve dynamic optimization, this application provides an interactive correction interface to receive user feedback. Users can evaluate and correct the output results, add new knowledge, and the system records user feedback, including corrected conclusions, supplemented knowledge, and evaluation opinions. For example, users can correct inaccurate conclusions in the analysis results, add the latest relevant published research results, and score the credibility of the conclusions.

[0059] User feedback is transformed into optimization signals using online reinforcement learning technology based on artificial intelligence, which in turn updates the parameters of the academic hypergraph and the large language model. During the hypergraph optimization process, the authority weights of relevant nodes are adjusted based on user feedback. The weights of highly credible nodes recognized by users are increased, while the weights of controversial or erroneous nodes are decreased. New nodes and hyperedges corresponding to user-supplemented knowledge are added. For example, the latest research results added by users will be parsed into new document nodes, keyword nodes, etc., and associated with existing nodes in the hypergraph. The structured feature vectors of relevant nodes are updated simultaneously to ensure that the hypergraph can reflect the latest knowledge system.

[0060] Model parameter adjustments employ an AI-based incremental fine-tuning approach. Using correct conclusions and supplementary knowledge from user feedback as training data, the model undergoes 1-2 rounds of fine-tuning with a learning rate of 2e-5 and a batch size of 4 to optimize output accuracy. During fine-tuning, the focus is on optimizing the model's understanding of user-focused areas, enabling it to better adapt to users' academic analysis habits and needs.

[0061] The method also includes a reliability level determination for the results, with four preset level ranges: excellent, good, satisfactory, and unsatisfactory. Based on the verification results of the credibility and anti-illusion processing in step S3, combined with the degree of support from authoritative data sources, the results of logical consistency verification, and the degree of compliance with academic norms, the reliability level of the processed analysis results is determined by an artificial intelligence comprehensive scoring engine. The degree of support from authoritative data sources is the proportion of conclusions associated with core authoritative nodes; the logical consistency verification result is the proportion of conflict-free conclusions; and the degree of compliance with academic norms is the proportion of statements that conform to the norms. The analysis results, reliability level, and correction basis are output simultaneously. The correction basis includes authoritative data source information anchored by facts, explanations of logical conflict corrections, and records of academic norm optimization, facilitating users to trace the source of the conclusions and the correction process. The level determination rule is: comprehensive score = degree of support from authoritative data sources × 0.4 + logical consistency verification result × 0.3 + degree of compliance with academic norms × 0.3. A comprehensive score ≥ 90% is excellent, 80%-89% is good, 60%-79% is satisfactory, and < 60% is unsatisfactory.

[0062] In a specific example, the system outputs results in two formats: a structured report and a visualized hypergraph. Users can interactively correct the wording of four conclusions and add three new research findings. The system transforms user feedback into optimization signals using online reinforcement learning technology, adjusting the authority weights of 18 relevant nodes, adding three document nodes, eight keyword nodes, and ten hyperedges. Using the user-corrected conclusions and supplementary knowledge as training data, the large language model undergoes one round of incremental fine-tuning. Combining the three indicators calculated in step S3, the final comprehensive score = 100% × 0.4 + 100% × 0.3 + 89.3% × 0.3 ≈ 96.8%, which is judged as excellent by the AI ​​comprehensive scoring engine. The final output includes the optimized analysis results, the excellent rating, and 150 corrective criteria.

[0063] Corresponding to the academic analysis method based on a large language model provided in the embodiments of the present invention, the embodiments of the present invention also provide an academic analysis and processing system based on a large language model, referring to... Figure 2 The system includes: The data acquisition module performs the data collection step in S1. It connects stably with academic databases, industry report repositories, and experimental data storage systems through standardized interfaces, comprehensively covering the collection needs of various academic resources. It supports access and parsing of common formats such as PDF, TXT, DOCX, CSV, PNG, and JPG, and is compatible with multimodal resources including full-text papers, patent specifications, and original experimental records. It can accurately extract text, charts, and formulas. During the collection process, differentiated strategies are developed for different data source characteristics, covering various academic achievements such as basic research papers and applied technology patents. After data cleaning to remove duplicates, blanks, and corrupted files, core fields such as document titles and experimental parameters are extracted. Through format standardization, a well-structured multi-source academic data set is generated. This module, through full format compatibility and refined cleaning, provides high-quality, non-redundant basic data input for subsequent Hypergraph construction.

[0064] The Hypergraph Construction Module executes the hypergraph construction step in S1, deeply analyzing multi-source academic datasets to extract five core node categories: literature, keywords, experimental data, citations, and technical features. Each node carries core content such as basic attributes, classification information, and key data, forming a basic unit system of academic knowledge. Based on these core nodes, a hyperedge association mechanism is used to construct an academic hypergraph H=(V,E,X), breaking through the limitations of traditional binary relation modeling in knowledge graphs and accurately capturing complex relationships between multiple nodes. Multimodal parsing technology is used to integrate text, charts, and formula semantics, generating structured feature vectors with unified dimensions to enhance node representation capabilities. An incremental update mechanism is also configured, incorporating the latest academic data through fixed periods or trigger-based conditions to synchronously update nodes, hyperedges, and feature vectors, ensuring the timeliness and completeness of the hypergraph. This module achieves deep integration and real-time iteration of multi-source academic knowledge through hypergraph modeling and dynamic updates.

[0065] The Hypergraph Neural Network Analysis Module executes step S2, performing feature aggregation based on the academic hypergraph through hypergraph convolutional layers. First, it constructs the adjacency and degree matrices to generate a normalized hypergraph Laplacian matrix. Then, it inputs the node feature matrices into the convolutional layers, combining them with the ReLU activation function to generate intermediate feature matrices. Finally, it uses global average pooling to generate a 512 or 768-dimensional high-order semantic representation, balancing semantic richness and computational complexity. It employs cross-disciplinary terminology corpora to perform AI-supervised fine-tuning of the general large language model, dividing the training, validation, and test sets in an 8:1:1 ratio to optimize the model's adaptability to academic scenarios. The high-order semantic representation is input into the AI-tuned model, and combined with user trend analysis, innovation point comparison, or literature review writing requirements, it outputs preliminary academic analysis results of 3,000-8,000 words. This module efficiently transforms hypergraph knowledge into accurate preliminary academic analysis results through feature aggregation and AI model fine-tuning.

[0066] The Trustworthy Anti-Hallucination Processing Module executes step S3 and incorporates AI-powered semantic matching, logical reasoning, hallucination identification, and normative verification algorithms. It works collaboratively through four sub-modules: fact anchoring verification, multi-source cross-validation, hallucination filtering, and academic normative verification. Fact anchoring verification uses AI algorithms to establish a traceable relationship between conclusions and hypergraph nodes; multi-source cross-validation uses AI consistency detection algorithms to identify conflicts, quantify their intensity, and mark their locations; hallucination filtering uses AI similarity calculations to correct low-similarity, high-risk content; and academic normative verification uses AI terminology matching and paradigm comparison algorithms to optimize the compliance of expressions. Simultaneously, it calculates three quantitative indicators: the degree of support from authoritative data sources, the results of logical consistency verification, and the degree of academic normative compliance, providing data support for reliability assessment. This module, through AI-driven, multi-dimensional verification and quantitative evaluation, comprehensively eliminates errors and non-standard content, strengthening the trustworthiness of academic analysis results.

[0067] The Results Output and Optimization Module executes step S4, outputting structured reports, visualized hypergraphs, and other deliverables based on the processed analysis results. It provides an interactive interface for receiving user feedback. Feedback is transformed into optimization signals using online reinforcement learning technology, adjusting hypergraph node weights, adding new nodes and hyperedges, and using AI to back-update the parameters of the large language model. Combining three quantitative indicators, an AI-powered comprehensive scoring engine calculates a comprehensive score, classifying it into four levels: excellent, good, satisfactory, and unsatisfactory, according to preset rules. The module simultaneously outputs the analysis results, reliability level, and correction criteria. Through diversified output and two-way AI optimization, this module achieves precise delivery and continuous iteration of academic analysis.

[0068] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram can be found in the corresponding architecture diagram. The computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor of the computer device provides computing and control capabilities, supporting the entire academic analysis process. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and a database, while the internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores core data such as multi-source academic raw data, hypergraph parameters, model training configurations, quantitative indicators, and analysis reports. The I / O interfaces are used for the processor to exchange information with external devices, such as importing academic data files and exporting analysis results. The communication interface is used to communicate with external terminals or academic data platforms via a network connection to achieve remote data transmission and sharing. When the computer program is executed by the processor, it implements the academic analysis method based on a large language model as described in this application.

[0069] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0070] In one exemplary embodiment, a computer device is also provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments, namely, to complete the entire process of multi-source academic data acquisition and hypergraph construction, high-order semantic representation generation and preliminary analysis, credibility enhancement and anti-illusion processing, structured output and closed-loop optimization.

[0071] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments, enabling the academic analysis methods to be stably reproduced.

[0072] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments, providing a convenient software implementation carrier for academic analysis based on large language models.

[0073] It should be noted that the data involved in this application (including but not limited to multi-source academic texts, charts, formula data, hypergraph node and hyperedge information, model parameters, quantitative indicators and analysis results data, etc.) are all information and data authorized by the producer or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with the relevant regulations on academic data security.

[0074] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (Read-Only Memory). Memory includes ROM, magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0075] The databases involved in the various embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the various embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., but are not limited to these.

[0076] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0077] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An academic analysis method based on a large language model, characterized in that, The method includes: S1. Acquire multi-source academic data and construct an academic hypergraph; S2. Based on the academic hypergraph, a high-order semantic representation is generated by aggregating hypergraph features through a hypergraph neural network. The semantic representation is then input into a large language model that has been fine-tuned by artificial intelligence and adapted to academic scenarios, and preliminary academic analysis results are output. S3. Based on the preliminary academic analysis results, perform both credibility enhancement and anti-hallucination processing to obtain the processed analysis results; S4. Based on the processed analysis results, output structured results to meet the user's academic analysis needs, and optimize the parameters of the constructed academic hypergraph and the large language model based on user feedback.

2. The academic analysis method based on a large language model according to claim 1, characterized in that, Step S1, which involves acquiring multi-source academic data and constructing an academic hypergraph, includes: Collect papers, patents, and experimental data from academic databases, industry reports, and experimental data to generate a multi-source academic data set; The multi-source academic data set is analyzed to extract five core nodes: literature, keywords, experimental data, citations, and technical features. The five types of core nodes are associated with hyperedges to construct an academic hypergraph. At the same time, multimodal parsing technology is used to integrate the semantics of text, charts and formulas to configure structured feature vectors for hypergraph nodes. Configure a hypergraph incremental update mechanism to incorporate the latest academic data and synchronously update the hypergraph nodes, hyperedge relationships, and feature vectors.

3. The academic analysis method based on a large language model according to claim 1, characterized in that, Step S2, which generates preliminary academic analysis results, includes: Based on the academic hypergraph, the node features associated with the hyperedges are aggregated through hypergraph convolutional layers to generate a high-order semantic representation; By using interdisciplinary terminology corpus, a general large language model is fine-tuned under AI supervision to obtain a large language model adapted to academic scenarios. The higher-order semantic representation is input into the large language model, and combined with the user's academic analysis requirements, preliminary analysis results are output. The academic analysis requirements include trend analysis, comparison of innovative points, and writing of literature reviews.

4. The academic analysis method based on a large language model according to claim 1, characterized in that, Step S3, the dual processing of credibility and anti-hallucination, includes fact anchoring verification, specifically: Each conclusion of the preliminary analysis results is associated with a node in the academic hypergraph, and the corresponding original text fragment, citation node, and experimental data node are labeled one by one. A traceable relationship is established between the conclusions and the nodes of the academic hypergraph, generated by an artificial intelligence algorithm, and each conclusion is associated with an authoritative data source.

5. The academic analysis method based on a large language model according to claim 4, characterized in that, Step S3, the dual processing of credibility and anti-hallucination, also includes multi-source cross-validation, specifically: By utilizing the multi-node association characteristics of the academic hypergraph, node information associated with different hyperedges is extracted; By comparing the node information, the consistency between the literature conclusions and the cited experimental data, and between the technical features and the patent claims is verified. The AI ​​consistency detection algorithm automatically detects logical conflicts and marks the conflict locations, and generates a conflict analysis description.

6. The academic analysis method based on a large language model according to claim 4, characterized in that, The dual processing of credibility and anti-hallucination described in step S3 also includes hallucination filtering, specifically: Calculate the similarity between the preliminary academic analysis results and the higher-order semantic representation, and preset a similarity threshold; Content below the threshold is marked as a high-risk illusion. The academic hypergraph is automatically retrieved to supplement alternative data sources, and the large language model is driven to revise the conclusion based on the node features and association information of the academic hypergraph.

7. The academic analysis method based on a large language model according to claim 4, characterized in that, Step S3, the dual processing of credibility and anti-hallucination, also includes academic standard verification, specifically: The subject specification nodes stored in the academic hypergraph are invoked, and the nodes contain the expression paradigms, terminology standards and citation specifications of different subjects; Verify the compliance of the expression and the accuracy of the terminology in the preliminary analysis results sentence by sentence, generate suggestions for standard correction, and optimize the expression of the results.

8. The academic analysis method based on a large language model according to claim 1, characterized in that, Step S4, which involves adapting to users' academic analysis needs, outputting structured results, and performing reverse optimization, includes: Based on the processed analysis results, an interactive correction interface is provided to receive user feedback. The user feedback content is transformed into optimization signals using artificial intelligence online reinforcement learning technology, which inversely updates the node weights and hyperedge relationships of the hypergraph, and simultaneously adjusts the parameters of the large language model.

9. The academic analysis method based on a large language model according to claim 8, characterized in that, The method also includes determining the reliability level of the results: A preset reliability level range is defined, which is divided into four levels: excellent, good, qualified, and unqualified. Based on the verification results of the credibility and anti-hallucination processing described in step S3, combined with the degree of support from authoritative data sources, the results of logical consistency verification, and the degree of compliance with academic norms, the artificial intelligence comprehensive scoring engine determines the reliability level of the processed analysis results. The system synchronously outputs analysis results, reliability levels, and correction criteria, which include authoritative data source information anchored to facts, explanations of logical conflict corrections, and records of academic standard optimization.

10. An academic analysis and processing system based on a large language model, characterized in that, The system is used to implement the academic analysis method based on a large language model as described in any one of claims 1-9, and the system comprises: The data acquisition module is used to collect papers, patents, and experimental data from academic databases, industry reports, and experimental data to generate a multi-source academic data set; The hypergraph construction module is used to parse the multi-source academic data set, extract five types of core nodes, construct an academic hypergraph by connecting the core nodes with hyperedges, configure structured feature vectors, and perform incremental updates of the hypergraph. The hypergraph neural network analysis module is used to generate high-order semantic representations based on the aggregated features of the academic hypergraph and drive the large language model to output preliminary academic analysis results. The credible anti-hallucination processing module is used to perform fact anchoring verification, multi-source cross-verification, hallucination filtering, and academic norm verification supported by artificial intelligence algorithms on the preliminary academic analysis results. The results output module is used to output structured results based on the processed analysis results and receive user feedback to reverse optimize the parameters of the hypergraph and large language model, and output complete analysis results including the reliability level evaluated online by artificial intelligence.

Images