Scientific research behavior monitoring and risk identification system based on large language model

Abstract

The invention discloses a scientific research behavior monitoring and risk identification system based on a large language model, and the system comprises a scientific research text collection and preprocessing module which is used for collecting an unstructured scientific research text in a scientific research activity; the scientific research semantic representation generation module is used for carrying out semantic coding on the structured input sequence based on a large language model; the risk identification module is used for receiving the scientific research semantic representation, identifying a risk behavior type and outputting a corresponding risk confidence coefficient; the risk explanation generation module is used for generating a natural language risk explanation text based on the scientific research risk identification result; the risk level division and associated text extraction module is used for dividing risk levels and extracting text paragraphs related to risk behaviors; and the scientific research behavior risk identification report generation module is used for generating a scientific research behavior risk identification report. Based on a scientific corpus large language model, scientific research behavior risk identification and explanation generation are realized, and the method has the advantages of accurate identification, clear explanation and high adaptability.

Description

Technical Field

[0001] This invention relates to the field of scientific research risk identification, and in particular to a scientific research behavior monitoring and risk identification system based on a large language model. Background Technology

[0002] In existing technologies, the identification of risks in scientific research mainly relies on manual review or rule-based detection methods. These methods typically require a large amount of manual feature template development, making them difficult to adapt to diverse and semantically complex unstructured scientific research texts. For example, risky behaviors such as text plagiarism, falsification of experimental data, or fabrication of technical pathways are highly concealed and semantically diverse, and traditional methods have significant limitations in terms of identification accuracy and interpretability. Furthermore, while some studies have introduced machine learning models for auxiliary identification, they are mostly based on shallow features and lack deep semantic modeling and contextual understanding capabilities, resulting in insufficient interpretability and generalization ability of the identification results.

[0003] To address the aforementioned issues, current technologies have not fully leveraged the semantic modeling capabilities of pre-trained language models to achieve automatic identification and causal explanation of complex risk behaviors in research contexts. There is a lack of a unified workflow capable of semantic encoding, behavior recognition, risk explanation, and report generation, and a structured research risk knowledge base has not been established to support continuous optimization training.

[0004] Therefore, how to provide a scientific research behavior monitoring and risk identification system based on a large language model is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a scientific research behavior monitoring and risk identification system based on a large language model. This invention is based on a scientific corpus and a large language model to achieve scientific research behavior risk identification and interpretation generation, and has the advantages of accurate identification, clear interpretation and strong adaptability.

[0006] A scientific research behavior monitoring and risk identification system based on a large language model according to an embodiment of the present invention includes: The scientific research text acquisition and preprocessing module is used to acquire unstructured scientific research texts from scientific research activities and generate structured input sequences. The scientific research semantic representation generation module is used to perform semantic encoding on structured input sequences based on a large language model to generate scientific research semantic representations. The risk identification module is used to receive scientific research semantic representations, identify risky behavior types, and output the corresponding risk confidence scores. The risk explanation generation module is used to generate natural language risk explanation text based on the results of scientific research risk identification. The risk level classification and associated text extraction module is used to classify risk levels and extract text paragraphs related to risky behaviors; The research behavior risk identification report generation module is used to structurally integrate risk behavior type labels, natural language risk explanation text, risk behavior type labels and related text paragraph sets to generate a research behavior risk identification report.

[0007] Optionally, modules can be integrated using the following methods: Collect unstructured research texts generated during scientific research activities, perform data preprocessing, and generate structured input sequences; The structured input sequence is fed into a large language model trained on a scientific corpus, and the structured input sequence is semantically encoded to generate the corresponding scientific research semantic representation. The scientific research semantic representation is input into the risk identification module, which identifies the risk behavior type of the current scientific research semantic representation and calculates the corresponding risk confidence level, and integrates and generates scientific research risk identification results. The results of scientific research risk identification are input into the risk explanation generation module, which generates corresponding natural language risk explanation text for the identified risky behaviors. Risk levels are categorized, semantically related paragraphs are extracted, and risk level labels and sets of related text paragraphs are generated respectively. The risk behavior type labels, scientific research semantic representations, and natural language risk explanation texts are uniformly formatted to generate a draft of the scientific research behavior risk identification report. This draft is then structurally combined with risk level labels and related text paragraph sets to generate the full scientific research behavior risk identification report.

[0008] Optionally, the unstructured scientific research text includes project proposals, research plans, draft papers, technical reports, and experimental records, and the data preprocessing includes sentence segmentation, word segmentation, stop word removal, text cleaning, normalization, and vector encoding.

[0009] Optionally, the generation of the scientific research semantic representation specifically includes: Each word in the structured input sequence is fed into the embedding mapping layer of a large language model trained on scientific corpora. The embedding mapping transforms the word into a word vector representation. The large language model trained on scientific corpora refers to a natural language processing model constructed using unstructured scientific research texts from multiple disciplines as pre-training data and a deep neural network structure. It includes three structural components: an embedding mapping layer, a multi-layer bidirectional semantic coding network, and an output mapping layer. The embedding mapping layer converts the word words in the unstructured scientific research text into word vector representations. The multi-layer bidirectional semantic coding network models the word vectors in both forward and backward contexts to generate contextual semantic representation vectors. The output mapping layer generates risk prediction labels, confidence scores, or semantic feature vectors in specific tasks. All word vector representations are input into a multi-layer bidirectional semantic coding network. The dependency relationship between each word and its context is established through forward semantic paths and backward semantic paths, respectively. Context representation vectors are generated by calculating the similarity or correlation between each word and its context. The multi-layer bidirectional semantic coding network performs context modeling through forward and backward semantic paths. The forward semantic path refers to the process in the multi-layer bidirectional semantic coding network where the model processes and transmits information step by step from the beginning of the input sequence along the sequence order, establishing the dependency relationship between each word and its corresponding forward context. The backward semantic path refers to the process in the multi-layer bidirectional semantic coding network where the model starts from the end of the input sequence and processes and transmits information step by step forward along the sequence order, establishing the dependency relationship between each word and its corresponding backward context. The context representation vector is standardized and then input into a weighted fusion of an attention mechanism based on global semantic dependencies. The attention mechanism calculates weight coefficients based on the context representation vectors of all words in the input sequence. All weighted and fused context representation vectors are concatenated in sequence and input into the output mapping layer to generate scientific semantic representations.

[0010] Optionally, the generation of the scientific research risk identification results specifically includes: The research semantic representation is input into the risk behavior type discrimination submodule of the risk identification module. The research semantic representation is mapped to the latent space representation through a fully connected mapping method. A multi-class neural network structure is introduced to output the corresponding risk behavior type label. The risk identification module includes a risk behavior type discrimination submodule and a risk confidence prediction submodule. The risk behavior type label includes text plagiarism, duplicate project applications, experimental data fabrication, fabricated technical paths, and leakage of sensitive information. The scientific research semantic representation is input into the risk confidence prediction submodule. The degree of matching between the current scientific research semantic representation and the risk behavior type label is calculated through the label matching probability function to generate the risk confidence score. By structurally integrating risk behavior type labels and corresponding risk confidence levels, research risk identification results are generated.

[0011] Optionally, the generation of the natural language risk explanation text specifically includes: The results of scientific research risk identification are input into the risk explanation generation module, which is composed of the generative language model T5-Sci. T5-Sci is a large language model that integrates scientific research context modeling capabilities and controllable text generation capabilities. It is based on the T5 architecture and is extended in a targeted manner. It uses unstructured scientific research texts from multiple disciplines as training data for supervised fine-tuning training. Structurally, it includes an input parsing encoder, a cross-attention alignment module, and a controllable text generation decoder, which supports the conversion of structured scientific research semantic representations into natural language risk explanation texts. Based on risk behavior type labels and scientific research semantic representations, semantic encoding is performed through the input parsing encoder in T5-Sci to generate encoded representation vectors; The encoded representation vector is input into the cross-attention alignment module of T5-Sci and processed through the cross-attention mechanism. A multi-head attention structure is used to extract the semantic support fragments between the risk behavior type label and the scientific research semantic representation, generating an alignment vector sequence covering semantic, structural and behavioral features. The semantic support fragments refer to the textual expressions that directly reflect or indirectly corroborate specific risk behavior types, as identified by the multi-head attention mechanism, including linguistic feature fragments in research hypotheses, descriptions of experimental processes, data citation logic, and research conclusions. The alignment vector sequence is a set of high-dimensional representation vectors calculated based on a multi-head attention mechanism; The alignment vector sequence is input into the controllable text generation decoder. Based on the risk interpretation language template built in the training phase, the decoding operation is executed step by step according to the logical order of risk behavior. In each generation step, the context state and output embedding of the current position are recorded to generate a historical generation state sequence. The risk explanation language template refers to a set of high-frequency language structure patterns constructed during the model training phase based on risk behavior explanation statements annotated by experts in scientific research behavior corpus through template extraction and induction methods. The logical order of risk behavior types refers to the language organization structure and information unfolding path set according to the composition rules of risk behavior types in the process of generating natural language risk explanation text. It includes four components in sequence: behavioral statement, semantic support, judgment basis and potential consequences, forming a risk explanation text framework with clear structure, rigorous reasoning and traceability. A gated attention mechanism is introduced, which receives an alignment vector sequence and a historical generation state sequence, calculates the weight score of each semantic feature dimension through an attention weight allocation unit, outputs a dimension adjustment coefficient through a gated control function unit, and dynamically adjusts the participation ratio of the input signal in the generation process in combination with the generation position index, thereby generating a natural language risk explanation text corresponding to the risk behavior type. The gated attention mechanism includes an attention weight allocation unit and a gated control function unit.

[0012] Optionally, the generation of the risk level labels and associated text paragraph set specifically includes: Based on the risk confidence level in the scientific research risk identification results, and combined with the preset risk level classification rules, a level calculation operation is performed to generate risk level labels corresponding to risk behavior types. Based on the context representation vector in the scientific research semantic representation, combined with the weight scores of semantic support fragments obtained through the cross-attention mechanism and the position index of risk behavior type labels, context fragments that are semantically related to the current risk behavior type are extracted from the scientific research semantic representation, mapped back to the original unstructured scientific research text, the corresponding text paragraphs are located, and a set of related text paragraphs is constructed.

[0013] The beneficial effects of this invention are: This invention provides a research behavior monitoring and risk identification system based on a large language model, overcoming the problems of low identification efficiency, poor adaptability, and weak interpretability caused by reliance on manual review, fixed rules, or shallow feature matching in existing technologies. This method collects unstructured texts such as project proposals, draft papers, and experimental records generated during research activities. Using a large language model trained on a scientific corpus, it performs word vector representation and multi-layer bidirectional semantic encoding to achieve deep semantic modeling and context-dependent understanding of research texts. This effectively identifies potentially risky research behavior types, such as plagiarism, data fabrication, and fictitious technical paths.

[0014] Building upon risk identification, this invention further constructs a generative risk explanation mechanism. Based on the T5-Sci natural language generation model with a T5 architecture, it utilizes semantic support fragments and a gated attention mechanism to perform contextual alignment of risk behavior semantics. Combined with language templates summarized by expert annotations, it achieves structured generation of natural language risk explanation text. This text not only covers the behavioral characteristics, judgment criteria, and potential consequences of risky behaviors, but also possesses the ability to analyze causal logical chains and normative comparisons, improving the understandability, verifiability, and traceability of risk identification results, and meeting the needs of scientific research review, management, and compliance governance.

[0015] Furthermore, this method classifies risk levels based on risk confidence, locates text fragments by mapping semantic vectors, constructs a set of paragraphs associated with risky behaviors, and structurally splices together risk labels, semantic interpretations, level information, and original contextual content to generate a complete research behavior risk identification report. This report is logically clear, comprehensive, and directly supports review processes and behavior monitoring tasks. In summary, this invention significantly outperforms traditional methods in terms of semantic recognition accuracy, risk interpretation capability, and structured result presentation, demonstrating high practicality and application value. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a scientific research behavior monitoring and risk identification system based on a large language model proposed in this invention. Figure 2 This is a schematic diagram of the risk interpretation generation module structure of a scientific research behavior monitoring and risk identification system based on a large language model proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1-2 A research behavior monitoring and risk identification system based on a large language model includes: The scientific research text acquisition and preprocessing module is used to acquire unstructured scientific research texts from scientific research activities and generate structured input sequences. The scientific research semantic representation generation module is used to perform semantic encoding on structured input sequences based on a large language model to generate scientific research semantic representations. The risk identification module is used to receive scientific research semantic representations, identify risky behavior types, and output the corresponding risk confidence scores. The risk explanation generation module is used to generate natural language risk explanation text based on the results of scientific research risk identification. The risk level classification and associated text extraction module is used to classify risk levels and extract text paragraphs related to risky behaviors; The research behavior risk identification report generation module is used to structurally integrate risk behavior type labels, natural language risk explanation text, risk behavior type labels and related text paragraph sets to generate a research behavior risk identification report.

[0019] In this embodiment, the modules are interconnected using the following method: Collect unstructured research texts generated during scientific research activities, perform data preprocessing, and generate structured input sequences; The structured input sequence is fed into a large language model trained on a scientific corpus, and the structured input sequence is semantically encoded to generate the corresponding scientific research semantic representation. The scientific research semantic representation is input into the risk identification module, which identifies the risk behavior type of the current scientific research semantic representation and calculates the corresponding risk confidence level, and integrates and generates scientific research risk identification results. The results of scientific research risk identification are input into the risk explanation generation module, which generates corresponding natural language risk explanation text for the identified risky behaviors. Risk levels are categorized, semantically related paragraphs are extracted, and risk level labels and sets of related text paragraphs are generated respectively. The risk behavior type labels, scientific research semantic representations, and natural language risk explanation texts are uniformly formatted to generate a draft of the scientific research behavior risk identification report. This draft is then structurally combined with risk level labels and related text paragraph sets to generate the full scientific research behavior risk identification report.

[0020] This invention provides a research behavior monitoring and risk identification system based on a large language model. By constructing an efficient research text processing system, it can accurately identify potential risky behaviors in research activities. Traditional methods often rely on manual review and rule matching, which are inefficient and have poor adaptability. This invention, however, automatically identifies risky behaviors in research texts, such as plagiarism and data fabrication, through deep semantic modeling and contextual understanding. This method significantly improves the accuracy of risk identification, avoids the limitations of rule-based reliance, and has strong adaptability, capable of handling complex research contexts. Through the combination of a multi-layer bidirectional semantic coding network and an attention mechanism, the system not only significantly improves accuracy but also comprehensively covers different types of risky behaviors, significantly improving the efficiency and effectiveness of research behavior monitoring.

[0021] In this embodiment, the unstructured scientific research text includes project proposals, research plans, draft papers, technical reports, and experimental records. The data preprocessing includes sentence segmentation, word segmentation, stop word removal, text cleaning, normalization, and vector encoding.

[0022] This invention overcomes the reliance on manual intervention in traditional methods by detailing how to preprocess unstructured scientific texts and generate structured input sequences. Traditional methods suffer from difficulties in information extraction and inconsistent text formats when processing scientific texts. This invention standardizes the processing of scientific texts through steps such as sentence segmentation, word segmentation, stop word removal, text cleaning, normalization, and vector encoding. This not only reduces manual intervention but also ensures high-quality input data, helping the subsequent semantic encoding module to more accurately understand the scientific text and thus improving the overall recognition performance. This processing step not only improves the quality of the model input but also greatly enhances the versatility and adaptability of the method.

[0023] In this embodiment, the generation of the scientific research semantic representation specifically includes: Each word in the structured input sequence is fed into the embedding mapping layer of a large language model trained on scientific corpora. The word is then converted into a word vector representation through embedding mapping. The word consists of scientific terms or language units obtained in the preprocessing step. The embedding mapping method converts the word sequence in unstructured scientific text into a sequence of real-valued vectors, based on the embedding matrix built into the pre-trained language model. Each word is represented by a low-dimensional vector by finding its corresponding position in the embedding matrix. The word vector is a fixed-dimensional real-valued vector used to express the basic semantic features of the word. The large language model trained on scientific corpora refers to a natural language processing model constructed using unstructured scientific texts from multiple disciplines as pre-training data and a deep neural network structure. It includes three structural components: an embedding mapping layer, a multi-layer bidirectional semantic encoding network, and an output mapping layer. The embedding mapping layer converts word elements in unstructured scientific text into word vector representations. The multi-layer bidirectional semantic encoding network models the word vectors for forward and backward contextualization, generating contextual semantic representation vectors. The output mapping layer generates risk prediction labels, confidence scores, or semantic feature vectors for specific tasks. All word vector representations are input into a multi-layer bidirectional semantic coding network. Dependencies between each word and its context are established through forward and backward semantic paths, respectively. Context representation vectors are generated by calculating the similarity or correlation between each word and its context. The multi-layer bidirectional semantic coding network models the context through forward and backward semantic paths. The forward semantic path refers to the model processing and transmitting information sequentially from the beginning of the input sequence, establishing dependencies between each word and its corresponding forward context. The backward semantic path refers to the model processing and transmitting information sequentially from the end of the input sequence, processing and transmitting information sequentially from the reverse order, establishing dependencies between each word and its corresponding backward context. The context representation vectors are used to represent the semantic roles of words in a scientific research context. The context representation vector is standardized and then input into a weighted fusion of an attention mechanism based on global semantic dependencies. The attention mechanism calculates weight coefficients based on the context representation vectors of all words in the input sequence. All weighted and fused context representation vectors are concatenated in sequence and input into the output mapping layer to generate a scientific research semantic representation, which is then passed as output to the risk identification module to support subsequent risk behavior type discrimination and confidence prediction operations.

[0024] The risk identification module of this invention is based on the semantic representation of scientific research using a large language model. It can accurately identify potential risky behaviors in scientific research texts and output risk confidence scores. Traditional methods rely on manual labeling or rule-based judgments during risk identification, resulting in low accuracy and susceptibility to human bias. This invention, however, automatically identifies risky behavior types and calculates corresponding confidence scores through risk identification based on scientific research semantic representations, avoiding the limitations of rules and enhancing the robustness and automation of risk identification. The application of multi-classification neural networks further improves the ability to identify complex scientific research behaviors, thereby effectively improving the accuracy and efficiency of scientific research behavior monitoring.

[0025] In this embodiment, the generation of the scientific research risk identification result specifically includes: The research semantic representation is input into the risk behavior type discrimination submodule of the risk identification module. The research semantic representation is mapped to the latent space representation through a fully connected mapping method. A multi-class neural network structure is introduced to output the corresponding risk behavior type label. The risk identification module includes a risk behavior type discrimination submodule and a risk confidence prediction submodule. The risk behavior type label includes text plagiarism, duplicate project applications, experimental data fabrication, fabricated technical paths, and leakage of sensitive information. The scientific research semantic representation is input into the risk confidence prediction submodule. The matching degree between the current scientific research semantic representation and the risk behavior type label is calculated through the label matching probability function to generate the risk confidence. The risk confidence is a real value between zero and one, which is used to characterize the credibility of the risk behavior type label. By structurally integrating risk behavior type labels and corresponding risk confidence levels, research risk identification results are generated.

[0026] This invention automatically identifies potential risk behaviors in research texts, such as plagiarism and data fabrication, by inputting the research semantic representation into the risk behavior type discrimination submodule of the risk identification module and combining fully connected mapping with a multi-classification neural network structure. Compared to traditional methods that rely on manual rules, this invention significantly improves the accuracy and intelligence of risk identification through deep learning and contextual modeling. By inputting the research semantic representation into the risk confidence prediction submodule, this invention generates accurate risk confidence scores, characterizing the credibility of risk behavior type labels. This design provides a quantitative basis for risk assessment. Finally, by combining risk behavior type labels and confidence scores, this invention generates a complete research risk identification report through structured integration, which not only improves the accuracy and credibility of risk identification but also provides efficient and scientific decision support for research compliance management.

[0027] In this embodiment, the generation of the natural language risk explanation text specifically includes: The results of scientific research risk identification are input into the risk explanation generation module, which is composed of the generative language model T5-Sci. T5-Sci is a large language model that integrates scientific research context modeling capabilities and controllable text generation capabilities. It is based on the T5 architecture and is extended in a targeted manner. It uses unstructured scientific research texts from multiple disciplines as training data for supervised fine-tuning training. Structurally, it includes an input parsing encoder, a cross-attention alignment module, and a controllable text generation decoder. It supports the conversion of structured scientific research semantic representations into natural language risk explanation texts. It has high language controllability, cross-risk behavior type generalization ability, and adaptability to complex scientific research contexts. It is especially suitable for the intelligent generation of causal and compliance explanation texts in fields such as scientific research behavior monitoring and risk identification. Based on risk behavior type labels and scientific research semantic representations, semantic encoding is performed through the input parsing encoder in T5-Sci to generate encoded representation vectors; The encoded representation vector is input into the cross-attention alignment module of T5-Sci and processed through the cross-attention mechanism. A multi-head attention structure is used to extract the semantic support fragments between the risk behavior type label and the scientific research semantic representation, generating an alignment vector sequence covering semantic, structural and behavioral features. The semantic support fragments refer to the textual expressions that directly reflect or indirectly corroborate specific risk behavior types, as identified by the multi-head attention mechanism, including linguistic feature fragments in research hypotheses, descriptions of experimental processes, data citation logic, and research conclusions. The alignment vector sequence is a set of high-dimensional representation vectors calculated based on the multi-head attention mechanism. It reflects the inherent relationship between risk behavior type labels and scientific research semantic representation at multiple semantic levels, such as semantic logic, text structure and research behavior characteristics. As the input state of the generative language model, it is used to drive the generation process of natural language risk explanation text. The alignment vector sequence is input into the controllable text generation decoder. Based on the risk interpretation language template built in the training phase, the decoding operation is executed step by step according to the logical order of risk behavior. In each generation step, the context state and output embedding of the current position are recorded to generate a historical generation state sequence. The aforementioned risk explanation language templates refer to a set of high-frequency language structure patterns constructed during the model training phase based on expert-annotated risk behavior explanation statements in scientific research behavior corpora through template extraction and induction methods. These templates cover typical language expression frameworks for multiple types of scientific research risk behaviors. Each template clearly defines the generation order, grammatical structure, and logical connection method. Specific components include risk description, behavioral characteristics, supporting evidence, comparative analysis, and conclusion judgment, which are used to express risk behavior types, describe behavioral patterns, provide semantic evidence, compare normative standards, and output analytical conclusions, respectively. As a structural control mechanism for generative text decoders, these templates provide language structure guidance and content organization constraints for the generation process of risk explanation texts, thereby ensuring the professionalism and consistency of the final generated content. The logical order of risk behavior types refers to the language organization structure and information unfolding path set according to the composition rules of risk behavior types in the process of generating natural language risk explanation text. It includes four components in sequence: behavioral statement, semantic support, judgment basis and potential consequences, forming a risk explanation text framework with clear structure, rigorous reasoning and traceability. The behavioral statement section clearly describes the type of risky behavior currently identified; The semantic support component extracts keywords, logical chains, or contextual information from unstructured scientific research texts to construct evidence support for risky behavior types. The judgment criteria section analyzes the rationality of risky behavior types by comparing them with scientific research norms or historical cases; The section on potential consequences describes the negative impacts of current risky behaviors on scientific research practice; A gated attention mechanism is introduced, which receives an alignment vector sequence and a historical generation state sequence, calculates the weight score of each semantic feature dimension through an attention weight allocation unit, outputs a dimension adjustment coefficient through a gated control function unit, and dynamically adjusts the participation ratio of the input signal in the generation process in combination with the generation position index, thereby generating a natural language risk explanation text corresponding to the risk behavior type. The gated attention mechanism includes an attention weight allocation unit and a gated control function unit.

[0028] This invention generates natural language risk explanation text through a risk explanation generation module. Based on the generative language model T5-Sci, it can generate risk explanations with clear structure and causal logic. Compared with traditional methods, this module overcomes the limitations of generating explanations solely through manual interpretation or simple templates. By using the T5-Sci model, combined with research context modeling capabilities and controllable text generation capabilities, it automatically generates explanation texts that conform to research norms, greatly improving the readability and auditability of risk identification results. Simultaneously, the high linguistic controllability and generalization ability across risk behavior types during the generation process ensure the accuracy and broad adaptability of the generated text, optimizing the automation and intelligence level of research compliance management.

[0029] In this embodiment, the generation of the risk level label and the associated text paragraph set specifically includes: Based on the risk confidence level in the scientific research risk identification results, and combined with the preset risk level classification rules, a level calculation operation is performed to generate risk level labels corresponding to risk behavior types. Based on the context representation vector in the scientific research semantic representation, combined with the weight scores of semantic support fragments obtained through the cross-attention mechanism and the position index of risk behavior type labels, context fragments that are semantically related to the current risk behavior type are extracted from the scientific research semantic representation, mapped back to the original unstructured scientific research text, the corresponding text paragraphs are located, and a set of related text paragraphs is constructed.

[0030] This invention generates a research behavior risk identification report by classifying risk levels and extracting related text paragraphs, combined with risk behavior type tags and research semantic representation. Compared to traditional methods, this invention can more accurately classify research risks and extract related paragraphs, automating the arduous manual review process in traditional methods. By combining semantic representation and contextual understanding, the system can extract research text paragraphs closely related to risky behaviors, ensuring that the report's content is more comprehensive and accurate. This method not only improves the quality of the report but also provides reliable data for subsequent risk monitoring and decision support, enhancing the report's systematicity and practicality. Example

[0031] To verify the feasibility of this invention in practice, it was applied to the research compliance review work of a university's research management department. This department routinely processes a large volume of documents submitted by faculty and researchers, including research proposals, research reports, draft papers, and experimental records. These documents are complex and highly specialized, making manual review difficult, time-consuming, and prone to issues such as overlooking risk behaviors and unclear interpretations. This is especially true when dealing with interdisciplinary research content, where traditional rule matching and keyword comparison methods are insufficient for the demands of refined compliance management.

[0032] In practical applications, staff first import the unstructured research text to be reviewed into the research behavior monitoring and risk identification system proposed in this invention. The system automatically completes data preprocessing operations, including text cleaning, sentence segmentation, stop word removal, and vector encoding, thereby converting the original text into a structured input sequence, facilitating subsequent semantic modeling. Subsequently, the system calls a large language model trained on scientific corpora to perform semantic encoding operations on the structured input sequence, outputting a representation vector that expresses the deep semantics of the research text. This semantic representation not only preserves the semantic features inherent in the research text but also captures the contextual logical relationships between words, improving the model's ability to understand complex scientific language.

[0033] Subsequently, the system feeds the aforementioned semantic representations into the risk identification module. Combining this with the semantic features of various research non-compliance behaviors acquired during model training, the system automatically identifies high-risk behaviors such as plagiarism, fabricated experimental data, duplicate submissions of research plans, or fabricated technical pathways, and provides corresponding confidence scores. Based on the identification results, the system further invokes the risk explanation generation module. Combining risk behavior labels and research semantic representations, and leveraging cross-attention and gated attention mechanisms, it accurately extracts semantic support content highly correlated with the risk behavior, automatically generating a clear causal and structurally complete natural language risk explanation. The explanations generated by the system not only clearly indicate the type of identified risk behavior but also analyze the potential causes and consequences of the behavior through contextual semantic chains, greatly enhancing the comprehensibility and persuasiveness of the review results.

[0034] To facilitate subsequent review and management, the system automatically classifies risk levels based on the confidence level in the risk identification results and extracts semantic context paragraphs related to risky behaviors from the original research texts to construct a set of semantically related paragraphs. Finally, the system integrates risk behavior labels, research semantic representations, natural language interpretations, risk level labels, and related text paragraphs to generate a well-structured and comprehensive research behavior risk identification report. This report is submitted to management for review and archiving, improving the standardization and normalization of the overall review process.

[0035] By applying this system in this scenario, research management departments have significantly reduced the time spent manually reviewing research texts word by word, while simultaneously improving the coverage of high-risk research behaviors. Previously overlooked semantic risks can now be accurately revealed with the support of this system, enhancing the institution's ability to control research compliance. Furthermore, the natural language explanations generated by the system are clear and easy to understand, allowing managers to quickly grasp the issues without requiring extensive professional background and make reasonable judgments. This process not only improves the efficiency and quality of research behavior review but also provides strong technical support for promoting research integrity and standardizing the research process.

[0036] Table 1. Performance Comparison of the Invention and Traditional Methods in Scientific Research Behavior Risk Identification Tasks

[0037] As can be clearly seen from Table 1, the overall performance of this invention in the task of identifying risks in scientific research behavior shows a significant advantage over traditional methods.

[0038] First, in terms of accuracy and recall for risk identification, this invention achieved 92.6% and 89.7% respectively, while traditional methods achieved 78.3% and 72.5%. This fully demonstrates that this invention offers a dual improvement in both accuracy and coverage of risk behavior identification. Most traditional methods are based on manual rules or shallow machine learning, which struggles to adapt to complex semantic contexts, leading to frequent misjudgments or omissions. In contrast, this method employs a multi-layer bidirectional semantic coding network combined with a global attention mechanism, effectively understanding long-distance dependencies and latent semantic associations in research texts, thereby achieving higher-quality risk identification.

[0039] Secondly, the comparison of expert review intervention rates shows that the present invention has an intervention rate of only 13.4%, while the traditional method has an intervention rate as high as 39.6%. This means that the present invention significantly reduces the frequency of manual review and alleviates the review burden considerably. This effect is due to the fact that the risk interpretation generated by the present method has a causal logical structure and semantic support chain, which facilitates rapid verification and review and enhances the reliability of the system results.

[0040] In terms of processing efficiency, this invention reduces the processing time for a single document to 15.2 seconds, while the traditional method takes an average of 48.7 seconds, demonstrating a significant improvement in efficiency. This difference reflects the superiority of the end-to-end automated processing flow of this invention. From unstructured text to structured report generation, the process is highly integrated and optimized in parallel, greatly shortening the overall analysis time and laying the foundation for real-time monitoring of large-scale scientific literature.

[0041] In terms of risk interpretation and comprehensibility scoring, this invention achieved a score of 9.1, significantly higher than the 6.3 score of traditional methods. This advantage stems from the introduction of the generative language model T5-Sci and its semantically controllable generation capabilities. Through guided template generation, cross-attention mechanisms, and gating adjustment strategies, the generated content not only conforms to grammatical norms but also exhibits clear organization and logical consistency, thereby enhancing the end-user's understanding and acceptance of the system's output.

[0042] Furthermore, this invention also demonstrates significant advantages in multidisciplinary text adaptability and contextual understanding depth scoring, achieving scores of 9.3 and 8.8 respectively, while traditional methods score 5.4 and 5.1 respectively. This indicates that traditional solutions often perform well for texts in specific domains, but are prone to problems such as fuzzy recognition or fragmented interpretation in cross-disciplinary scenarios. In contrast, the large language model trained on cross-disciplinary scientific research corpora in this invention possesses stronger semantic generalization ability and deeper text understanding, maintaining high recognition quality and stability across different scientific research fields.

[0043] 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 research behavior monitoring and risk identification system based on a large language model, characterized in that, include: The scientific research text acquisition and preprocessing module is used to acquire unstructured scientific research texts from scientific research activities and generate structured input sequences. The scientific research semantic representation generation module is used to perform semantic encoding on structured input sequences based on a large language model to generate scientific research semantic representations. The risk identification module is used to receive scientific research semantic representations, identify risky behavior types, and output the corresponding risk confidence scores. The risk explanation generation module is used to generate natural language risk explanation text based on the results of scientific research risk identification. The risk level classification and associated text extraction module is used to classify risk levels and extract text paragraphs related to risky behaviors; The research behavior risk identification report generation module is used to structurally integrate risk behavior type labels, natural language risk explanation text, risk behavior type labels and related text paragraph sets to generate a research behavior risk identification report.

2. The scientific research behavior monitoring and risk identification system based on a large language model according to claim 1, characterized in that, The modules are connected in the following way: Collect unstructured research texts generated during scientific research activities, perform data preprocessing, and generate structured input sequences; The structured input sequence is fed into a large language model trained on a scientific corpus, and the structured input sequence is semantically encoded to generate the corresponding scientific research semantic representation. The scientific research semantic representation is input into the risk identification module, which identifies the risk behavior type of the current scientific research semantic representation and calculates the corresponding risk confidence level, and integrates and generates scientific research risk identification results. The results of scientific research risk identification are input into the risk explanation generation module, which generates corresponding natural language risk explanation text for the identified risky behaviors. Risk levels are categorized, semantically related paragraphs are extracted, and risk level labels and sets of related text paragraphs are generated respectively. The risk behavior type labels, scientific research semantic representations, and natural language risk explanation texts are uniformly formatted to generate a draft of the scientific research behavior risk identification report. This draft is then structurally combined with risk level labels and related text paragraph sets to generate the full scientific research behavior risk identification report.

3. The scientific research behavior monitoring and risk identification system based on a large language model according to claim 2, characterized in that, The unstructured scientific research texts include project proposals, research plans, draft papers, technical reports, and experimental records. The data preprocessing includes sentence segmentation, word segmentation, stop word removal, text cleaning, normalization, and vector encoding.

4. The scientific research behavior monitoring and risk identification system based on a large language model according to claim 2, characterized in that, The generation of the scientific research semantic representation specifically includes: Each word in the structured input sequence is fed into the embedding mapping layer of a large language model trained on scientific corpora. The embedding mapping transforms the word into a word vector representation. The large language model trained on scientific corpora refers to a natural language processing model constructed using unstructured scientific research texts from multiple disciplines as pre-training data and a deep neural network structure. It includes three structural components: an embedding mapping layer, a multi-layer bidirectional semantic coding network, and an output mapping layer. The embedding mapping layer converts the word words in the unstructured scientific research text into word vector representations. The multi-layer bidirectional semantic coding network models the word vectors for forward and backward context and generates contextual semantic representation vectors. The output mapping layer generates risk prediction labels, confidence scores, or semantic feature vectors in specific tasks. All word vector representations are input into a multi-layer bidirectional semantic coding network. The dependency relationship between each word and its context is established through forward semantic paths and backward semantic paths, respectively. Context representation vectors are generated by calculating the similarity or correlation between each word and its context. The multi-layer bidirectional semantic coding network performs context modeling through forward and backward semantic paths. The forward semantic path refers to the process in the multi-layer bidirectional semantic coding network where the model processes and transmits information step by step from the beginning of the input sequence along the sequence order, establishing the dependency relationship between each word and its corresponding forward context. The backward semantic path refers to the process in the multi-layer bidirectional semantic coding network where the model starts from the end of the input sequence and processes and transmits information step by step forward along the sequence order, establishing the dependency relationship between each word and its corresponding backward context. The context representation vector is standardized and then input into a weighted fusion of an attention mechanism based on global semantic dependencies. The attention mechanism calculates weight coefficients based on the context representation vectors of all words in the input sequence. All weighted and fused context representation vectors are concatenated in sequence and input into the output mapping layer to generate scientific semantic representations.

5. A scientific research behavior monitoring and risk identification system based on a large language model according to claim 2, characterized in that, The generation of the scientific research risk identification results specifically includes: The research semantic representation is input into the risk behavior type discrimination submodule of the risk identification module. The research semantic representation is mapped to the latent space representation through a fully connected mapping method. A multi-class neural network structure is introduced to output the corresponding risk behavior type label. The risk identification module includes a risk behavior type discrimination submodule and a risk confidence prediction submodule. The risk behavior type label includes text plagiarism, duplicate project applications, experimental data fabrication, fabricated technical paths, and leakage of sensitive information. The scientific research semantic representation is input into the risk confidence prediction submodule. The degree of matching between the current scientific research semantic representation and the risk behavior type label is calculated through the label matching probability function to generate the risk confidence score. By structurally integrating risk behavior type labels and corresponding risk confidence levels, research risk identification results are generated.

6. The scientific research behavior monitoring and risk identification system based on a large language model according to claim 2, characterized in that, The generation of the natural language risk explanation text specifically includes: The results of scientific research risk identification are input into the risk explanation generation module, which is composed of the generative language model T5-Sci. T5-Sci is a large language model that integrates scientific research context modeling capabilities and controllable text generation capabilities. It is based on the T5 architecture and is extended in a targeted manner. It uses unstructured scientific research texts from multiple disciplines as training data for supervised fine-tuning training. Structurally, it includes an input parsing encoder, a cross-attention alignment module, and a controllable text generation decoder, which supports the conversion of structured scientific research semantic representations into natural language risk explanation texts. Based on risk behavior type labels and scientific research semantic representations, semantic encoding is performed through the input parsing encoder in T5-Sci to generate encoded representation vectors; The encoded representation vector is input into the cross-attention alignment module of T5-Sci and processed through the cross-attention mechanism. A multi-head attention structure is used to extract the semantic support fragments between the risk behavior type label and the scientific research semantic representation, generating an alignment vector sequence covering semantic, structural and behavioral features. The semantic support fragments refer to the textual expressions that directly reflect or indirectly corroborate specific risk behavior types, as identified by the multi-head attention mechanism, including linguistic feature fragments in research hypotheses, descriptions of experimental processes, data citation logic, and research conclusions. The alignment vector sequence is a set of high-dimensional representation vectors calculated based on a multi-head attention mechanism; The alignment vector sequence is input into the controllable text generation decoder. Based on the risk interpretation language template built in the training phase, the decoding operation is executed step by step according to the logical order of risk behavior. In each generation step, the context state and output embedding of the current position are recorded to generate a historical generation state sequence. The risk explanation language template refers to a set of high-frequency language structure patterns constructed during the model training phase based on risk behavior explanation statements annotated by experts in scientific research behavior corpus through template extraction and induction methods. The logical order of risk behavior types refers to the language organization structure and information unfolding path set according to the composition rules of risk behavior types in the process of generating natural language risk explanation text. It includes four components in sequence: behavioral statement, semantic support, judgment basis and potential consequences, forming a risk explanation text framework with clear structure, rigorous reasoning and traceability. A gated attention mechanism is introduced, which receives an alignment vector sequence and a historical generation state sequence, calculates the weight score of each semantic feature dimension through an attention weight allocation unit, outputs a dimension adjustment coefficient through a gated control function unit, and dynamically adjusts the participation ratio of the input signal in the generation process in combination with the generation position index, thereby generating a natural language risk explanation text corresponding to the risk behavior type. The gated attention mechanism includes an attention weight allocation unit and a gated control function unit.

7. A scientific research behavior monitoring and risk identification system based on a large language model according to claim 2, characterized in that, The generation of the risk level labels and associated text paragraph sets specifically includes: Based on the risk confidence level in the scientific research risk identification results, and combined with the preset risk level classification rules, a level calculation operation is performed to generate risk level labels corresponding to risk behavior types. Based on the context representation vector in the scientific research semantic representation, combined with the weight scores of semantic support fragments obtained through the cross-attention mechanism and the position index of risk behavior type labels, context fragments that are semantically related to the current risk behavior type are extracted from the scientific research semantic representation, mapped back to the original unstructured scientific research text, the corresponding text paragraphs are located, and a set of related text paragraphs is constructed.

Images