A subject proposal form assisted generation method and system based on multi-source scientific research portraits

Abstract

The present application belongs to the field of artificial intelligence and scientific research management information technology, specifically to a project proposal assisted generation method and system based on multi-source scientific research portrait, wherein the method comprises: collecting multi-source heterogeneous scientific research data and constructing a structured knowledge graph; constructing a multi-dimensional scientific research portrait through a graph attention network and a dynamic topic model; matching the portrait with the semantic of the proposal template to drive a large language model to generate a preliminary draft of the proposal by modules; injecting literature evidence based on paragraph-level intent recognition; quantitatively evaluating the proposal and giving improvement suggestions; finally, learning the user's writing preferences by using a direct preference optimization algorithm, and incrementally updating the scientific research portrait to form a closed-loop iterative optimization. The present application realizes the intelligent and personalized generation and evaluation of the project proposal, significantly improving the writing efficiency, the quality of the proposal and the objectivity of the innovation evaluation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and scientific research management information technology, and in particular to a method and system for assisting in the generation of project application forms based on multi-source scientific research profiles. Background Technology

[0002] Currently, technologies related to the generation of project proposals mainly fall into three categories: (1) Researcher profiling – constructing a representation of a researcher's abilities and interests by analyzing research data such as papers and projects; (2) Retrieval-enhanced text generation (RAG) – using external knowledge bases to retrieve relevant literature to assist large language models in generating high-quality text; and (3) Novelty search and evaluation of research projects – evaluating the novelty of research projects through literature retrieval and comparative analysis. However, existing technologies have the following obvious shortcomings: First, the profiles are too narrow in scope and lack targeted modeling for project application generation. For example, patent application CN111241283B discloses a rapid representation method for researchers' profiles based on TF-IDF and clustering, which only extracts interest tags at the keyword level from paper abstracts. Patent application CN117332297A discloses a static / dynamic tagging profile method based on language models, which, although introducing a dynamic tagging mechanism, remains at the tag classification level. Patent application CN121436762A discloses a researcher competitiveness profile system based on multi-source data fusion, which, while constructing a multi-dimensional competitiveness profile, mainly serves career development path planning. None of these methods can construct a comprehensive profile encompassing multiple dimensions such as research capabilities, collaboration networks, domain expertise, and development trends, and therefore cannot directly serve the automatic generation of each chapter of the project application.

[0003] Second, there is a disconnect between profile creation and text generation, lacking end-to-end capabilities for assisting in the generation of research application documents. Existing technologies treat the construction of research profiles and text generation as independent tasks. Profile creation technology focuses on the representation and analysis of researcher characteristics, but does not address how to transform profile information into application document text. Existing search-enhanced document generation technologies (such as the search-enhanced automatic patent document generation method disclosed in patent application CN119886098A) have achieved automatic document generation based on RAG, but they are mainly geared towards specific scenarios such as patent documents and do not address the structured generation needs of research application documents, nor can they generate personalized content based on the applicant's research background. Summary of the Invention

[0004] This invention proposes a method and system for assisting in the generation of project proposals based on multi-source scientific research profiles, aiming to solve the technical problems of single-dimensional scientific research profiles and disconnect between profiles and text generation in existing technologies.

[0005] In a first aspect, the present invention provides a method for assisting in the generation of research proposals based on multi-source scientific research profiles, including: S1 collects heterogeneous scientific research data from multiple sources, performs data cleaning, entity disambiguation and format unification, and constructs a structured scientific research knowledge graph. S2. Based on the scientific research knowledge graph, construct a multi-dimensional scientific research profile that includes a research capability profile, a collaboration network profile, a domain expertise profile, and a development trend profile. S3, perform semantic matching between the multi-dimensional scientific research profile and the preset project application structure template, drive the module content generator, and generate the first draft of the application. S4. Perform paragraph-level semantic segmentation and writing intent identification on the initial draft of the application, implement differentiated hybrid retrieval strategies for paragraphs with different intents, retrieve relevant literature evidence from the scientific literature database, and use a large language model to perform retrieval enhancement and generative rewriting on the original paragraphs to generate a refined application. S5. Conduct a multi-dimensional quantitative evaluation of the refined application, including novelty, feasibility, cutting-edge nature and team fit, and generate a comprehensive innovation score through weighted fusion. S6. Receive user feedback on the application form through the human-computer collaborative editing interface. Based on the feedback, fine-tune the large language model using the direct preference optimization algorithm to learn the user's writing preferences, and update the multi-dimensional research profile using the implicit signal increment extracted from the feedback.

[0006] The technical advantages of the proposed method for generating research proposals based on multi-source research profiles disclosed in this invention are as follows: By collecting heterogeneous research data from multiple sources to construct a structured knowledge graph, a multi-dimensional research profile is built, encompassing research capabilities, collaborative networks, domain expertise, and development trends. This profile is then semantically matched with the proposal template to drive a large language model to generate initial drafts in modules. Further refinement is achieved through a differentiated hybrid retrieval enhancement generation technique driven by paragraph-level intent recognition. Finally, a quantitative evaluation is conducted across four dimensions: novelty, feasibility, cutting-edge research, and team fit, with iterative optimization based on feedback. This method achieves seamless integration between profile generation and proposal creation, resulting in highly personalized and academically credible content. The evaluation dimensions are comprehensive and objective, and the feedback learning mechanism enables continuous system evolution, significantly improving the efficiency and quality of research proposal writing.

[0007] Furthermore, S1 specifically includes: Using a heterogeneous data acquisition engine, raw scientific research data is collected from one or more sources, including paper databases, patent databases, funding project databases, and scholar homepages, through one or more modes such as API calls, web crawlers, and structured document parsing. Through the data cleaning and standardization module, the original data is precisely deduplicated based on unique identifiers, and the original data is fuzzily deduplicated based on the similarity between title and author using Jaccard similarity. The format of date and institution name is unified, and entity disambiguation is performed on scholar names using a graph neural network-based disambiguation model. The cleaned data is converted into a unified representation in JSON-LD format using a unified data schema converter and then stored in the document database.

[0008] Furthermore, S2 specifically includes: The BERT-BiLSTM-CRF model is used to extract research topic entities and method entities from the titles and abstracts of scientific research data. Dependency parsing is combined to extract semantic relationships between entities. A scientific research knowledge graph centered on scholars and including four types of nodes and their relationships, including scholars, papers, research fields and institutions, is constructed and stored in a graph database. Research capability is calculated from four quantitative dimensions: research output, academic influence, collaboration network degree, and funding acquisition capability, and then normalized by Min-Max to form a capability vector. A dynamic topic model is used to analyze the topic evolution of scholars' paper abstracts over the years with the year as the time window. The topic distribution vector of each time window is output, and the KL divergence of topic distribution in adjacent windows is calculated to quantify the intensity of research direction shift. A graph attention network is used to learn the representation of scholar nodes in the scientific research knowledge graph to obtain graph structure embedding vectors. Then, the graph structure embedding vectors, the capability vectors and the topic distribution vectors are aligned and fused through a cross-modal attention mechanism to generate the multi-dimensional scientific research profile vector.

[0009] Furthermore, S3 specifically includes: A pre-built template library contains application form structure templates for various fund types, with each template defining chapter structure, word count requirements, and content specifications; The application guide text is encoded into a semantic vector through a semantic matching engine. The cosine similarity between the vector and the multidimensional scientific research profile vector is calculated to assess the degree of matching. A suitable application form structure template is retrieved from the template library, and a mapping relationship is established between the fields in the template and the corresponding components of the scientific research profile vector. The modular content generator includes at least a research background generation module, a research content generation module, a technical route generation module, and an expected outcome generation module. Each module uses a dedicated prompt template to generate the text of the corresponding chapter based on a large language model. The generation process adopts a chapter-by-chapter serial strategy to maintain the coherence of the entire text.

[0010] Furthermore, S4 specifically includes: The paragraph-level semantic segmentation and intent recognition module divides the initial draft of the application into natural paragraphs and identifies the writing intent category of each paragraph through an intent classifier fine-tuned based on RoBERTa. The hybrid retrieval module performs a two-stage retrieval: the first stage uses FAISS for dense vector retrieval and Elasticsearch for BM25 sparse retrieval to coarsely recall Top-K candidate documents and take the union; the second stage uses a cross-encoder based on SciBERT fine-tuning to reorder the candidate documents in the union and select the most relevant documents as the final references. The large language model refinement generation module receives the original paragraph text and the final references, and performs enhanced rewriting according to the preset prompt template. During the rewriting process, reference marks are automatically inserted and supplemented with data support. The rewritten text is then cross-validated by the fact verification module for key data and references.

[0011] Furthermore, S5 specifically includes: The refined application form is encoded into a semantic vector. The most similar Top-K documents are retrieved from the scientific literature vector database, and the average semantic distance is calculated as the novelty score. ; Where K is the number of similar documents returned by the search. To refine the semantic vector of the application form, Let i be the semantic vector of the i-th similar document. The cosine similarity function; Furthermore, the weighted fusion in S5 specifically includes: The application document is analyzed using a natural language reasoning model to detect logical inconsistencies between its chapters, and the completeness and resource feasibility of the technical approach are examined, with a feasibility score output. ; The cutting-edge assessment module uses a dynamic topic model to calculate the overlap between the research topic in the application and recent academic hot topics, and combines the time series trends of emerging keywords to determine the stage of the research direction, outputting a cutting-edge score. ; The team matching assessment module calculates the cosine similarity between the comprehensive research profile vector of the applicant and team members and the research requirement vector extracted from the application, thus obtaining the team matching score. ; The weighted fusion scoring engine dynamically adjusts the weight vector W=[ based on the fund type.] , , , ] Calculate the comprehensive innovation score ; in, , , , The weights are assigned to four dimensions: novelty, feasibility, cutting-edge nature, and team fit. + + + =1.

[0012] Furthermore, S6 specifically includes: By comparing the changes in the topic distribution of the text before and after user modifications, the incremental vector of the profile update for each dimension of the multidimensional scientific research profile is calculated. The multidimensional scientific research profile is updated using an exponential moving average strategy. ; in, This is the updated multidimensional scientific research profile vector; The vector of the multidimensional scientific research profile before the update; This is the attenuation coefficient.

[0013] Secondly, the present invention provides a project proposal generation system based on multi-source scientific research profiling, the system being used to execute the method, the system comprising: The multi-source data acquisition and fusion module is used to collect multi-source heterogeneous scientific research data, perform data cleaning, entity disambiguation and format unification, and construct a structured scientific research knowledge graph. The multi-dimensional profile building module is used to construct a multi-dimensional scientific research profile that includes a research capability profile, a collaboration network profile, a domain expertise profile, and a development trend profile. The structured generation module is used to semantically match the multi-dimensional scientific research profile with the preset project application structure template, drive the sub-module content generator, and generate the first draft of the application. The retrieval-enhanced refined writing module is used to perform paragraph-level semantic segmentation and writing intent identification on the initial draft of the application. Differentiated hybrid retrieval strategies are implemented for paragraphs with different intents to retrieve relevant literature evidence from the scientific literature database. The original paragraphs are then rewritten using a large language model to enhance retrieval and generate a refined application. The innovation quantitative assessment module is used to conduct a multi-dimensional quantitative assessment of the innovation of the refined application. The assessment dimensions include novelty, feasibility, cutting-edge nature and team fit, and a comprehensive innovation score is generated through weighted fusion. The human-computer collaboration and iterative optimization module is used to receive user feedback on the application form through the human-computer collaborative editing interface. Based on the feedback, the module uses the direct preference optimization algorithm to fine-tune the large language model to learn the user's writing preferences and uses the implicit signal increment extracted from the feedback to update the multi-dimensional scientific research profile.

[0014] The technical effects of the system disclosed in this invention are as follows: This system can automatically construct a detailed profile of researchers from multi-source heterogeneous data, intelligently generate application forms that meet the requirements of different types of grants, provide interpretable multi-dimensional innovation evaluation reports, and continuously learn user preferences through a human-computer collaborative editing interface, becoming smarter with use, and providing researchers with a one-stop intelligent assistance service for grant application forms.

[0015] Thirdly, the present invention also provides a computer storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the steps of the method described. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for assisting in the generation of research proposals based on multi-source scientific research profiles, as proposed in an embodiment of the present invention. Figure 2 This is a schematic diagram of the multi-dimensional scientific research profile construction process provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a refined writing process based on RAG provided for an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The technical problem this invention aims to solve is to address the comprehensive deficiencies of existing technologies, such as the single dimension of scientific research profiles, the disconnect between profile generation and text generation, the lack of retrieval enhancement and refined writing mechanisms for scientific research scenarios, the lack of systematic and innovative quantitative evaluation methods, and the lack of human-machine collaborative iterative optimization mechanisms. This invention provides a method and system for assisting in the generation of project proposals based on multi-source scientific research profiles, achieving intelligent assistance throughout the entire process, from multi-source data to multi-dimensional profiles, to personalized proposal generation, retrieval enhancement and refined writing, multi-dimensional quantitative evaluation, and human-machine collaborative closed-loop iteration.

[0019] Based on this, embodiments of the present invention provide a method for assisting in the generation of research proposals based on multi-source scientific research profiles. This method drives the personalized generation of proposal content by constructing a multi-dimensional, detailed profile of the researcher, and is supplemented by search enhancement and quantitative evaluation, forming a closed-loop optimization through human-machine collaboration. Figure 1As shown, this method mainly includes six core steps: Step S1: Multi-source scientific research data collection and heterogeneous fusion; Step S2: Construction of multi-dimensional scientific research profiles; Step S3: Structured template matching and content generation for project proposals; Step S4: Refined writing based on RAG-based retrieval enhancement; Step S5: Multi-dimensional quantitative evaluation of innovation; Step S6: Human-machine collaborative iterative optimization and feedback learning. Specifically, it includes the following steps: Step S1: Multi-source scientific research data collection and heterogeneous fusion. This step aims to provide a comprehensive, accurate, and structured data foundation for subsequent profile construction. The specific implementation method is as follows: First, the system deploys a distributed data acquisition cluster, with each acquisition node responsible for one type of data source. The heterogeneous data acquisition engine supports three acquisition modes: for data sources providing standard APIs (such as Scopus and ORCID), data is acquired in batches via RESTful interfaces; for data sources without public APIs (such as some scholars' homepages), customized web crawlers are used, along with anti-crawling strategies such as User-Agent rotation and IP proxies; for user-uploaded local files (such as Excel, PDF, and BibTeX), information is extracted using the corresponding structured parser. Acquisition tasks are scheduled via message queues (such as RabbitMQ), supporting both incremental acquisition (based on the last acquisition timestamp, only acquiring new and updated records) and full acquisition modes.

[0020] The collected raw data enters a four-stage preprocessing pipeline: (1) Deduplication and noise reduction: First, precise deduplication is performed based on globally unique identifiers such as Digital Object Identifier (DOI) and patent number. For records lacking unique identifiers, fuzzy matching is performed using the combination of title and author information. The title similarity is calculated using Jaccard similarity. When the similarity is ≥0.85, it is determined to be a duplicate record and is removed. Jaccard similarity is defined as the ratio of the size of the intersection to the size of the union of two sets.

[0021] (2) Format standardization: Convert dates from different sources to ISO8601 format (e.g., “2023-10-26”); unify institutional names through a pre-built dictionary of institutional aliases (e.g., normalize “Chinese Academy of Sciences” and “Chinese Academy of Sciences” to “CAS”); journal names are also normalized through a dictionary of journal aliases.

[0022] (3) Entity disambiguation: To address the issue of duplicate names among researchers, a triple binding verification strategy of "name-institution-ORCID" is adopted. For scholars lacking ORCID information, the consistency of their collaborator network and research direction is utilized to input a graph neural network (GNN)-based disambiguation model for probabilistic disambiguation. This model takes a heterogeneous graph of scholar-paper-institution as input, learns the embedded representation of nodes through a message passing mechanism, and finally outputs a scholar ID with a unique identifier.

[0023] (4) Time alignment: All time-related fields (such as publication year, project start and end time) are unified to ISO8601 format and support alignment at three granularities: annual, quarterly and monthly, which facilitates subsequent time series analysis.

[0024] After preprocessing, the unified data schema converter transforms the cleaned data into a unified representation of JSON-LD (a lightweight LinkedData format based on JSON). The JSON-LD format effectively preserves the semantics and relational information of the data. Finally, each metadata record, including its source identifier, collection time, and confidence score, is stored in a MongoDB document database, preparing it for step S2.

[0025] Step S2: Constructing a multi-dimensional scientific research profile.

[0026] Based on the unified data output in step S1, this step constructs a multi-dimensional profile vector that comprehensively characterizes the researcher's features and serves the purpose of application generation. For example... Figure 2 As shown, the specific implementation methods include: First, a research knowledge graph is constructed. Using JSON-LD data as input, a BERT-BiLSTM-CRF model is employed to automatically extract research topic entities (such as "deep learning," "remote sensing image segmentation") and method entities (such as "convolutional neural network," "attention mechanism") from paper titles and abstracts. The BERT (Bidirectional Encoder Representation from Transformer) model is used to generate context-sensitive word embeddings, BiLSTM (Bidirectional Long Short-Term Memory Network) is used to capture sequence features, and CRF (Conditional Random Field) is used for globally optimal label sequence decoding. Simultaneously, dependency parsing is used to extract semantic relationships between entities (such as "A is applied to B," "C improves D"). Based on this, a scholar-centric research knowledge graph G=(V,E) is constructed and stored in the Neo4j graph database. The node set V contains four types of entities: scholars, papers, research fields, and institutions; the edge set E contains relationship types such as "published," "collaborated," "belongs to," and "funded."

[0027] Secondly, the multidimensional capability vector calculation module quantifies research capabilities from four dimensions: Research output: This is assessed by comprehensively considering indicators such as the number of papers, journal impact factor (IF) weighted publication volume, and H-index.

[0028] Academic Influence: Based on the PageRank algorithm, the influence score of scholars is calculated on the paper citation network, and a self-citation rate correction factor is introduced to avoid the bias caused by self-citation.

[0029] Collaborative network degree: Calculates the number of collaborators, collaboration diversity (the proportion of cross-institutional and cross-disciplinary collaborations), and the network centrality index of nodes in the graph (such as betweenness centrality).

[0030] Funding Acquisition Capability: Statistics on the number of funded projects led or participated in by the researcher, the total amount, and the project level (national, provincial, or ministerial level, etc.).

[0031] The quantification results of the above dimensions are then normalized using Min-Max to form a dense capability vector.

[0032] Then, the research direction time series modeling module employs a Dynamic Topic Model (DTM). Using the year as a time window, it performs topic evolution analysis on the abstracts of scholars' published papers over the years. The model outputs a topic distribution vector for each scholar within each time window. The intensity of research direction shifts is quantified by calculating the KL divergence of topic distributions between adjacent time windows, and future research direction trends are predicted using an exponentially weighted moving average (EWMA).

[0033] Finally, the multimodal profile fusion engine fuses all features into a single profile vector. First, a Graph Attention Network (GAT) is used to learn the representation of scholar nodes in the Neo4j knowledge graph. GAT effectively captures the complex contextual information of scholars in the research network by assigning different attention weights to neighboring nodes, ultimately outputting a graph structure embedding vector. Then, a cross-modal attention mechanism aligns and fuses the graph structure embedding vector, the capability vector, and the topic distribution vector. This mechanism learns the contribution weights of different modal features to the final task (application generation). The final output is a comprehensive multidimensional research profile vector containing information on four sub-profiles: research capability, collaborative network, domain expertise, and development trends. This profile vector has 512 dimensions and is stored in the Milvus vector database to support efficient similarity retrieval.

[0034] The alignment and fusion methods for the three modal vectors are as follows: First, three independent linear transformation matrices are used to project the three vectors to the same dimension. (e.g., 256-dimensional): ; ; ; Where Q is the query matrix, used to calculate the attention score with the key; K is the key matrix, used to match with the query; and V is the value matrix, used for weighted summation. To query the projection matrix; The key projection matrix; The projection matrix is ​​a value. Embedded vectors for graph structures; This is a capability vector; This represents the topic distribution vector.

[0035] Secondly, using Q (projected from the graph embedding) as the query and K (projected from the capability vector) as the key, the attention score is calculated: ; Where is the transpose of K; This is a scaling factor used to prevent the dot product from becoming too large; It is a normalized exponential function.

[0036] Next, the attention weights are used to perform a weighted summation of V (topic distribution projection) to obtain the initial fusion vector: .

[0037] Finally, $F$ is concatenated with the original $V_{graph}$ along the dimension, and then compressed to the target dimension through a fully connected layer: ; in, This is a vector concatenation operation. For fully connected layers, P is the final output multidimensional scientific research profile vector.

[0038] Step S3: Matching and generating structured templates for project proposals.

[0039] This step utilizes the research profile generated in step S2 to drive the automatic generation of the application content. The specific implementation method is as follows: The system comes pre-loaded with a library of application templates, stored in YAML format. The template library covers various types of projects, including National Natural Science Foundation of China (NSFC) general programs, youth programs, provincial-level funds, and industry-funded projects. Each template defines a chapter-based tree structure (e.g., "Basis for Project Establishment - Research Content - Technical Route - Expected Outcomes"), word count constraints for each chapter, and a list of required content elements.

[0040] The semantic matching engine first receives the application guideline text or grant name specified by the user. The engine uses a BERT encoder to encode the application guideline text into a semantic vector, then calculates its cosine similarity to the user's research profile vector to assess the degree of match between the user profile and the application direction, providing early warning feedback to the user. Simultaneously, the engine retrieves the most matching structural template from its template library based on the grant type. Subsequently, the engine establishes a mapping relationship, logically associating each field in the template (such as "research background" and "research content") with the corresponding component of the user's research profile vector.

[0041] The profile vector is manually defined using a hard-coded method, with the first 128 dimensions representing domain expertise, the middle 256 dimensions representing research capabilities, and the last 128 dimensions representing development trends. Template fields are mapped to corresponding dimension segments based on experience.

[0042] The modular content generator is based on a Large Language Model (LLM), which has been fine-tuned using numerous publicly available, high-quality grant application examples (rigorously anonymized). The generator comprises four dedicated modules, each equipped with a carefully designed prompt template: Research Background Generation Module: The Prompt template injects the domain expertise and development trend components from the profile, and, in accordance with the application guidelines, instructs the LLM to generate background text that includes a review of the current state of research at home and abroad, existing problems, and an explanation of the significance of the research.

[0043] Research Content Generation Module: The Prompt template, with research capability components at its core, guides LLM to generate specific research objectives, core research content, and key scientific questions to be addressed, based on the user's initial research ideas.

[0044] The technology roadmap generation module: The Prompt template integrates the collaboration network component and the capability component, requiring LLM to generate a description that includes research methods, experimental design, key technologies and their logical relationships, and finally outputs a text description that can be used to draw flowcharts.

[0045] Expected Outcomes Generation Module: The Prompt template integrates all profile components to guide LLM in generating expected outcomes that match the applicant's capabilities, such as the expected number and level of published papers, proposed patents, software copyrights, talent development plans, etc.

[0046] The generation process employs a chapter-by-chapter sequential strategy. For example, the "Research Background" section is generated first, and then the results of that section are used as context to generate the "Research Content," and so on, ensuring the logical and stylistic coherence of the entire text. Finally, the generated Markdown text is converted into a Word document that meets the application requirements using a template engine (such as Pandoc).

[0047] Step S4: RAG-based search enhancement for refined writing.

[0048] To enhance the academic credibility and timeliness of the generated content, this step employs search-enhanced generation technology to refine the initial draft. For example... Figure 3 As shown, the specific implementation method is as follows: First, the paragraph-level semantic segmentation module divides the initial draft of the application output from step S3 into natural paragraphs. For each paragraph, the system uses an intent classifier fine-tuned based on a RoBERTa pre-trained model to identify its writing intent. This classifier supports eight intent categories: research status review, problem statement, method description, theoretical analysis, experimental design, feasibility demonstration, innovation point description, and expected results.

[0049] The paragraph-level intent classifier supports the following eight intent categories. Based on the intent type, the system dynamically selects a retrieval strategy (dense vector retrieval, BM25 sparse retrieval, or hybrid retrieval) and determines the query construction method. The specific correspondences are as follows:

[0050] Next, the hybrid retrieval module dynamically constructs retrieval queries and executes a two-stage retrieval based on the identified intent category.

[0051] Phase 1 (Coarse Recall): For paragraphs in the "Research Status Overview" category, key terms (such as place names, algorithm names, and dataset names) are extracted to construct Boolean queries; for paragraphs in the "Method Description" category, the semantic vector of the entire paragraph is used directly for retrieval. The system performs two types of searches in parallel: dense search using FAISS in a pre-built scientific literature vector database (containing approximately 50 million Chinese and English paper abstracts from the past 10 years, encoded as 768-dimensional vectors using Sentence-BERT, and indexed with IVF-PQ) to recall the Top-100 documents; simultaneously, sparse search using Elasticsearch on the same database using BM25 to recall the Top-100 documents. The results from both recalls are then merged and deduplicated.

[0052] The second stage (fine-tuning): Each candidate document from the coarse recall stage is input into a cross-encoder (paragraph text, document abstract) that is finely tuned based on SciBERT. This encoder directly outputs a relevance score between 0 and 1. The system sorts all candidate documents in descending order based on this score and selects the Top-10 as the final references.

[0053] Finally, the LLM refinement generation module receives the original paragraph text and the retrieved Top-10 references. The system uses a specially designed Prompt template to guide the LLM in enhancing the rewriting. This template requires the LLM to: (1) insert reference citations (e.g., [Author, Year]) in appropriate positions in the text; (2) supplement and support the original arguments using specific data, charts, or conclusions from the references; (3) correct any factual errors in the original text based on authoritative literature; and (4) ensure that the rewritten text maintains logical coherence with the preceding and following paragraphs. After the rewriting is completed, an independent fact-checking module scans the generated text, extracts numerical statements (e.g., "accuracy reaches 95%)" and key citations, and cross-validates them with the retrieved original literature content. For content with a confidence level below a preset threshold (e.g., 0.8), the system will mark it and prompt the user for manual review.

[0054] Step S5: Multi-dimensional quantitative assessment of innovation.

[0055] This step involves an objective and quantitative evaluation of the innovativeness of the refined application. The specific implementation method is as follows: The system is evaluated from four dimensions: (1) Novelty assessment: The entire refined application text is encoded into a semantic vector using Sentence-BERT. In step S4, retrieve the top-50 most similar documents (K=50) from the scientific literature vector database, and calculate the average semantic distance as the novelty score: ; Where K is the number of similar documents returned by the search. To refine the semantic vector of the application form, Let i be the semantic vector of the i-th similar document. This is the cosine similarity function.

[0056] A higher novelty score indicates a greater difference between the application and existing research, and thus a higher level of novelty. The system also identifies new concepts and methods proposed in the application, performs precise matching, and uses the percentage of no matching items as a supplementary indicator.

[0057] (2) Feasibility Assessment: The assessment is conducted from three sub-dimensions. ① Completeness of Technical Route: A rule-based approach is used to check whether the application contains a complete logical chain of "problem definition → method design → experimental verification → result analysis." Points are deducted for any missing link. ② Logical Consistency: A DeBERTa-v3 Natural Language Inference (NLI) model, fine-tuned on the MultiNLI and SciNLI datasets, is used to check for logical contradictions (implication, contradiction, neutrality) between different sections of the application (e.g., "research objectives" and "technical route"). ③ Resource Feasibility: The required experimental conditions, datasets, and computing resources are matched with existing resources in the applicant's profile to assess resource gaps. The final feasibility score is output. .

[0058] (3) Frontier Assessment: Using the output of the dynamic topic model in step S2, calculate the overlap between the research topic in the application and the academic hot topics represented by the ESI (Essential Science Indicators) hot papers and highly cited papers in the past three years. At the same time, by analyzing the frequency of emerging keywords (keywords that have suddenly appeared and grown rapidly in the past year) in the application, determine whether the research direction is in an upward, stable, or declining phase, and assign corresponding weights to them, finally outputting the frontier score. .

[0059] (4) Team matching assessment: The cosine similarity between the comprehensive scientific research profile vector of the applicant and team members (which can be obtained by aggregating multiple member profile vectors) and the topic requirement vector extracted from the application (obtained by encoding the research content and technical route) is used as the team matching score.

[0060] Finally, the weighted fusion scoring engine dynamically adjusts the weights based on the fund type. For example, for NSFC general projects, which emphasize novelty and cutting-edge research, a weight vector W can be set as follows: , , , = [0.35, 0.15, 0.30, 0.20]; For enterprise-level horizontal research projects, focusing on feasibility and team fit, W=[0.15, 0.35, 0.20, 0.30]. The overall innovation score is: ; Ultimately, the scoring results are visualized in the form of a radar chart, and a text-based list of improvement suggestions is generated to accurately pinpoint the weaknesses in each dimension, providing users with clear directions for modification.

[0061] Step S6: Human-machine collaborative iterative optimization and feedback learning.

[0062] This step aims to enable the system to learn from and continuously evolve based on user modification behavior through human-machine collaboration. The specific implementation method is as follows: The human-computer collaborative editing interface is an interactive environment built on a web technology stack (React frontend, FastAPI backend). It presents the innovation assessment report (Step S5) and the detailed application form (Step S4) side-by-side to the user. The interface supports three interaction modes: Paragraph-level editing: Users can directly modify any paragraph in the WYSIWYG (what you see is what you get) editor, and the system records the text differences before and after the modification in real time via WebSocket.

[0063] Annotation feedback: Users can add text annotations to specific paragraphs (such as "This needs to be supplemented with the latest review from 2023"). Based on the annotation annotation map, the system calls LLM to automatically generate one or more modification proposals for the user to confirm and adopt.

[0064] Version Comparison: The system employs a Git-like difference storage strategy to maintain a complete version history of the application. Users can review and compare the differences between any two versions at any time, and one-click rollback is supported.

[0065] The preference learning module collects user modification behaviors from all the aforementioned interactions to construct a preference dataset. Based on this dataset, the underlying LLM is fine-tuned using the Direct Preference Optimization (DPO) algorithm. The DPO algorithm does not require training a separate reward model; instead, it optimizes the policy directly from the preference data using a closed-form solution, making it more stable and efficient. To reduce computational costs, the fine-tuning process employs Low-Rank Adaptation (LoRA) technology, updating only the low-rank adapter parameters injected into the LLM. A single fine-tuning session takes approximately 30 minutes on a single A100 GPU. After fine-tuning, the model can learn users' writing style preferences (e.g., whether they prefer short or long sentences), terminology usage habits (e.g., a preference for "attention mechanisms" rather than "deep learning algorithms"), and argumentation logic preferences.

[0066] The user profile incremental update module also extracts implicit feedback signals from modification behavior. For example, if a user significantly modifies the description of "Generative Adversarial Networks" in the "Research Background" section, the system analyzes the changes in the topic distribution of the text before and after the modification. If the modified text is more biased towards "Diffusion Model," it indicates that the weight of "Generative Adversarial Networks" in the user profile is too high, while the weight of "Diffusion Model" needs to be increased. The module calculates the incremental profile update vector for each dimension of the profile. And use an exponential moving average strategy for smooth updates: ; in, This is the updated multidimensional scientific research profile vector; The vector of the multidimensional scientific research profile before the update; This is the decay coefficient, used to balance the weights of historical profiles and new feedback. The default value is 0.8, ensuring that profile updates respond to current feedback while maintaining historical stability. The updated profile vectors are synchronously written back to the Neo4j graph database and the Milvus vector database and will take effect in the next generation round.

[0067] In addition, by monitoring every text modification operation by the user through the front-end editor, the text fragments before and after the modification are automatically extracted to form a comparison pair of (original text, modified text).

[0068] For each comparison pair, the system performs the following transformation steps: Using the same encoder as in step S2, the original text and the modified text are encoded into semantic vectors respectively. and ; Calculate the vector difference ΔV = - This difference represents the direction of semantic preference shift reflected by the user's modification; The difference vector ΔV is used as the image update increment vector ΔP for subsequent image updates.

[0069] Using the above method, any editing action of the user (including replacement, addition, deletion, and reordering) is uniformly converted into a calculable semantic offset signal, eliminating the need to design special rules for different editing types.

[0070] The entire system forms a closed-loop optimization process of "generation → evaluation → feedback → profile update → regeneration". The system is set to a maximum of 5 iterations. When the change in innovation score is less than 2% for two consecutive rounds, the system automatically determines that the result has converged and terminates the iteration.

[0071] The invention is illustrated below through a specific application scenario. Dr. Wang, a university professor, logs into the system to apply for a National Natural Science Foundation of China (NSFC) general program, uploading his personal information, including a list of 30 papers exported from CNKI (China National Knowledge Infrastructure), information on three funded projects he has participated in, and a brief CV. The system automatically collects and integrates the data: it obtains detailed information about Dr. Wang's papers by accessing the CNKI API and retrieves details of his project participation from the research funding database, organizing this scattered data into a unified format. Based on the collected data, a detailed profile of Dr. Wang is generated: his research ability is above average; his areas of expertise are computer vision and deep learning; his collaborative network shows frequent collaborations with researchers from Tsinghua University and Peking University; and his research trends indicate a shift from traditional image processing to AI generation in the past three years.

[0072] After Dr. Wang selected the corresponding application template, the system began generating a draft application. Based on 50 million articles, the system selected the 100 most relevant ones, wrote a review article in the research background section and inserted citations; designed research objectives according to Dr. Wang's areas of expertise; drew a flowchart of the technical roadmap from data collection to experimental verification; and predicted that the expected results might include 3 SCI papers and 1 software copyright. After the draft was generated, the system immediately provided an innovation assessment report: novelty 82 points, feasibility 75 points, cutting-edge nature 88 points, team fit 90 points, and an overall score of 84 points. The system also suggested supplementing the research background with a discussion of new methods developed in the last two years.

[0073] Dr. Wang carefully reviewed the initial draft and found a point that needed modification. He changed "deep learning algorithm" to "attention mechanism algorithm" and added his own perspective after a certain paragraph. After he made the changes and saved the document, the system learned his preferences: it remembered his preference for using the term "attention mechanism" and his habit of adding personal perspectives after each method, so that these styles could be automatically applied when generating the next application. After a period of continuous use, the system gradually adapted to Dr. Wang's writing style, and the generated applications became increasingly similar to his own writing, requiring fewer and fewer manual modifications.

[0074] Based on the same inventive concept, embodiments of the present invention also provide a project proposal generation assistance system based on multi-source scientific research profiling, the system being used to execute the method, the system comprising: The multi-source data acquisition and fusion module is used to collect multi-source heterogeneous scientific research data, perform data cleaning, entity disambiguation and format unification, and construct a structured scientific research knowledge graph. The multi-dimensional profile building module is used to construct a multi-dimensional scientific research profile that includes a research capability profile, a collaboration network profile, a domain expertise profile, and a development trend profile. The structured generation module is used to semantically match the multi-dimensional scientific research profile with the preset project application structure template, drive the sub-module content generator, and generate the first draft of the application. The retrieval-enhanced refined writing module is used to perform paragraph-level semantic segmentation and writing intent identification on the initial draft of the application. Differentiated hybrid retrieval strategies are implemented for paragraphs with different intents to retrieve relevant literature evidence from the scientific literature database. The original paragraphs are then rewritten using a large language model to enhance retrieval and generate a refined application. The innovation quantitative assessment module is used to conduct a multi-dimensional quantitative assessment of the innovation of the refined application. The assessment dimensions include novelty, feasibility, cutting-edge nature and team fit, and a comprehensive innovation score is generated through weighted fusion. The human-computer collaboration and iterative optimization module is used to receive user feedback on the application form through the human-computer collaborative editing interface. Based on the feedback, the module uses the direct preference optimization algorithm to fine-tune the large language model to learn the user's writing preferences and uses the implicit signal increment extracted from the feedback to update the multi-dimensional scientific research profile.

[0075] Based on the same inventive concept, embodiments of the present invention also provide a computer storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the steps of the method described.

[0076] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A method for assisting in the generation of research proposals based on multi-source scientific research profiles, characterized in that, include: S1 collects heterogeneous scientific research data from multiple sources, performs data cleaning, entity disambiguation and format unification, and constructs a structured scientific research knowledge graph. S2. Based on the scientific research knowledge graph, construct a multi-dimensional scientific research profile that includes a research capability profile, a collaboration network profile, a domain expertise profile, and a development trend profile. S3, perform semantic matching between the multi-dimensional scientific research profile and the preset project application structure template, drive the module content generator, and generate the first draft of the application. S4. Perform paragraph-level semantic segmentation and writing intent identification on the initial draft of the application, implement differentiated hybrid retrieval strategies for paragraphs with different intents, retrieve relevant literature evidence from the scientific literature database, and use a large language model to perform retrieval enhancement and generative rewriting on the original paragraphs to generate a refined application. S5. Conduct a multi-dimensional quantitative evaluation of the refined application, including novelty, feasibility, cutting-edge nature and team fit, and generate a comprehensive innovation score through weighted fusion. S6. Receive user feedback on the application form through the human-computer collaborative editing interface. Based on the feedback, fine-tune the large language model using the direct preference optimization algorithm to learn the user's writing preferences, and update the multi-dimensional research profile using the implicit signal increment extracted from the feedback.

2. The method according to claim 1, characterized in that, S1 specifically includes: Using a heterogeneous data acquisition engine, raw scientific research data is collected from one or more sources, including paper databases, patent databases, funding project databases, and scholar homepages, through one or more modes such as API calls, web crawlers, and structured document parsing. Through the data cleaning and standardization module, the original data is precisely deduplicated based on unique identifiers, and the original data is fuzzily deduplicated based on the similarity between title and author using Jaccard similarity. The format of date and institution name is unified, and entity disambiguation is performed on scholar names using a graph neural network-based disambiguation model. The cleaned data is converted into a unified representation in JSON-LD format using a unified data schema converter and then stored in the document database.

3. The method according to claim 1, characterized in that, S2 specifically includes: The BERT-BiLSTM-CRF model is used to extract research topic entities and method entities from the titles and abstracts of scientific research data. Dependency parsing is combined to extract semantic relationships between entities. A scientific research knowledge graph centered on scholars and including four types of nodes and their relationships, including scholars, papers, research fields and institutions, is constructed and stored in a graph database. Research capability is calculated from four quantitative dimensions: research output, academic influence, collaboration network degree, and funding acquisition capability, and then normalized by Min-Max to form a capability vector. A dynamic topic model is used to analyze the topic evolution of scholars' paper abstracts over the years with the year as the time window. The topic distribution vector of each time window is output, and the KL divergence of topic distribution in adjacent windows is calculated to quantify the intensity of research direction shift. A graph attention network is used to learn the representation of scholar nodes in the scientific research knowledge graph to obtain graph structure embedding vectors. Then, the graph structure embedding vectors, the capability vectors and the topic distribution vectors are aligned and fused through a cross-modal attention mechanism to generate the multi-dimensional scientific research profile vector.

4. The method according to claim 1, characterized in that, S3 specifically includes: A pre-built template library contains application form structure templates for various fund types, with each template defining chapter structure, word count requirements, and content specifications; The application guide text is encoded into a semantic vector through a semantic matching engine. The cosine similarity between the vector and the multidimensional scientific research profile vector is calculated to assess the degree of matching. A suitable application form structure template is retrieved from the template library, and a mapping relationship is established between the fields in the template and the corresponding components of the scientific research profile vector. The modular content generator includes at least a research background generation module, a research content generation module, a technical route generation module, and an expected outcome generation module. Each module uses a dedicated prompt template to generate the text of the corresponding chapter based on a large language model. The generation process adopts a chapter-by-chapter serial strategy to maintain the coherence of the entire text.

5. The method according to claim 1, characterized in that, S4 specifically includes: The paragraph-level semantic segmentation and intent recognition module divides the initial draft of the application into natural paragraphs and identifies the writing intent category of each paragraph through an intent classifier fine-tuned based on RoBERTa. The hybrid retrieval module performs a two-stage retrieval: the first stage uses FAISS for dense vector retrieval and Elasticsearch for BM25 sparse retrieval to coarsely recall Top-K candidate documents and take the union; the second stage uses a cross-encoder based on SciBERT fine-tuning to reorder the candidate documents in the union and select the most relevant documents as the final references. The large language model refinement generation module receives the original paragraph text and the final references, and performs enhanced rewriting according to the preset prompt template. During the rewriting process, reference marks are automatically inserted and supplemented with data support. The rewritten text is then cross-validated by the fact verification module for key data and references.

6. The method according to claim 1, characterized in that, S5 specifically includes: The refined application form is encoded into a semantic vector. The most similar Top-K documents are retrieved from the scientific literature vector database, and the average semantic distance is calculated as the novelty score. ; Where K is the number of similar documents returned by the search. To refine the semantic vector of the application form, Let i be the semantic vector of the i-th similar document. This is the cosine similarity function.

7. The method according to claim 1, characterized in that, The weighted fusion in S5 specifically includes: The application document is analyzed using a natural language reasoning model to detect logical inconsistencies between its chapters, and the completeness and resource feasibility of the technical approach are examined, with a feasibility score output. ; The cutting-edge assessment module uses a dynamic topic model to calculate the overlap between the research topic in the application and recent academic hot topics, and combines the time series trends of emerging keywords to determine the stage of the research direction, outputting a cutting-edge score. ; The team matching assessment module calculates the cosine similarity between the comprehensive research profile vector of the applicant and team members and the research requirement vector extracted from the application, thus obtaining the team matching score. ; The weighted fusion scoring engine dynamically adjusts the weight vector W=[ based on the fund type.] , , , ] Calculate the comprehensive innovation score ; in, , , , The weights are assigned to four dimensions: novelty, feasibility, cutting-edge nature, and team fit. + + + =1.

8. The method according to claim 1, characterized in that, S6 specifically includes: By comparing the changes in the topic distribution of the text before and after user modifications, the incremental vector of the profile update for each dimension of the multidimensional scientific research profile is calculated. The multidimensional scientific research profile is updated using an exponential moving average strategy. ； in, This is the updated multidimensional scientific research profile vector; The vector of the multidimensional scientific research profile before the update; This is the attenuation coefficient.

9. A project proposal generation system based on multi-source scientific research profiling, characterized in that, The system is used to perform the method according to any one of claims 1-8, the system comprising: The multi-source data acquisition and fusion module is used to collect multi-source heterogeneous scientific research data, perform data cleaning, entity disambiguation and format unification, and construct a structured scientific research knowledge graph. The multi-dimensional profile building module is used to construct a multi-dimensional scientific research profile that includes a research capability profile, a collaboration network profile, a domain expertise profile, and a development trend profile. The structured generation module is used to semantically match the multi-dimensional scientific research profile with the preset project application structure template, drive the sub-module content generator, and generate the first draft of the application. The retrieval-enhanced refined writing module is used to perform paragraph-level semantic segmentation and writing intent identification on the initial draft of the application. Differentiated hybrid retrieval strategies are implemented for paragraphs with different intents to retrieve relevant literature evidence from the scientific literature database. The original paragraphs are then rewritten using a large language model to enhance retrieval and generate a refined application. The innovation quantitative assessment module is used to conduct a multi-dimensional quantitative assessment of the innovation of the refined application. The assessment dimensions include novelty, feasibility, cutting-edge nature and team fit, and a comprehensive innovation score is generated through weighted fusion. The human-computer collaboration and iterative optimization module is used to receive user feedback on the application form through the human-computer collaborative editing interface. Based on the feedback, the module uses the direct preference optimization algorithm to fine-tune the large language model to learn the user's writing preferences and uses the implicit signal increment extracted from the feedback to update the multi-dimensional scientific research profile.

10. A computer storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1-8.

Images