A scientific research topic deduplication auxiliary decision system and method fusing a large language model and vector retrieval

Abstract

This invention discloses a research topic deduplication auxiliary decision-making system and method integrating large language model and vector retrieval. The system comprises a five-layer architecture: a user interaction layer, an application service layer, an intelligent analysis layer, a retrieval engine layer, and a data storage layer. The method includes: acquiring the text of the research topic to be searched and performing structured extraction; performing multi-path parallel retrieval and large language model re-ranking; calculating the comprehensive duplication rate based on an internal candidate set of results; calculating the comprehensive novelty score based on an external candidate set of patents; and outputting project approval suggestions based on the comprehensive duplication rate and comprehensive novelty score using a four-quadrant decision matrix. This invention solves the technical problems of insufficient semantic understanding ability, weak decision support ability, and lack of multi-dimensional analysis ability in existing technologies by using a multi-algorithm fusion similarity calculation method, a Facet-based novelty assessment method, a multi-level duplication determination model, and a four-quadrant matrix decision framework. It achieves quantitative assessment of the novelty of research topics and intelligent assistance in project approval decisions.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence-assisted scientific research management technology, specifically involving an intelligent decision support system that deeply integrates large language models and vector retrieval technology for the novelty search of scientific research projects. Background Technology

[0002] Novelty search for research projects is a crucial step in research management. Its core task is to determine whether the proposed project is novel compared to existing research results and whether there is a risk of duplicate funding, thereby providing a basis for decision-making on the optimal allocation of research resources.

[0003] Currently, novelty searches for scientific research projects mainly rely on the following traditional methods.

[0004] The first method is manual novelty search. This method involves professionals searching for relevant literature using keywords and reading and comparing each article, ultimately judging the novelty of the topic based on subjective experience. However, with the exponential growth in the number of documents, manual novelty search is extremely inefficient; more importantly, different reviewers have significant differences in their judgments of similarity, resulting in highly subjective and inconsistent evaluation results.

[0005] The second type is keyword-based retrieval systems. Existing systems, such as search engines like Elasticsearch, use the BM25 algorithm or TF-IDF cosine similarity for text matching. These methods can only handle literal similarity and cannot understand technical content that is essentially the same but expressed differently, such as "large-span steel structure" and "large-span spatial structure." For repetitive topics that use synonym substitution, near-synonym substitution, or sentence variation, these systems cannot accurately identify them, leading to frequent missed and false positives for repetitive topics.

[0006] The third type is a general-purpose large language model question-answering system. Although large language models have certain intelligent question-answering and semantic understanding capabilities, existing applications are mostly limited to the general interaction level. They lack specific design for the particular scenario of deduplication in scientific research topic selection and cannot output quantitative evaluation results that can be used for project approval decisions.

[0007] Furthermore, Chinese invention patent application CN121234913A discloses a method for detecting the duplication of marine engineering research projects. This method segmentes words by constructing a marine engineering terminology dictionary, introduces semantic similarity compensation, and performs duplication detection based on a combination of text similarity and attribute similarity. This method, like this invention, belongs to the field of novelty searching for research project initiation and has high domain relevance. However, it has the following shortcomings: First, the core innovation of this method is still concentrated on the duplication detection level, only outputting the determination result of the degree of duplication, without involving quantitative assessment of novelty, let alone establishing a decision-making mechanism that integrates novelty and duplication; Second, although this method introduces semantic similarity compensation in similarity calculation, it is essentially still based on a single similarity calculation path and does not adopt a multi-algorithm weighted fusion calculation strategy; Third, this method does not link the duplication detection results with project initiation decision recommendations, and cannot answer key decision-making questions such as "to what extent is the duplication" and "whether to recommend project initiation."

[0008] In addition, Chinese invention patent application CN119166746A discloses a method and apparatus for evaluating the novelty credibility of search points under multi-model collaboration. This method proposes to use a large language model to expand the topic and keywords to construct a search formula, use the search point as a reference template to extract relevant steps from the screened documents and summarize relevant points, and then calculate the similarity distance and novelty score between the search point and relevant points in each dimension. This scheme has the following shortcomings: (1) The similarity calculation of this method depends on the single path of the large language model, and there is no verification and correction mechanism for the possible deviation of the single model; (2) The dimensional decomposition of this method is based on the text structure of the search point itself, which leads to the lack of comparability and standardization of the evaluation results between different search points; (3) It has not formed a quantitative indicator and its integrated decision-making framework that can be directly used for project establishment decision-making, and it has not provided a four-quadrant matrix-style visualization decision-making tool.

[0009] Therefore, there is an urgent need in this field for a solution that can automatically, quantitatively, and multidimensionally assess the novelty of research topics, comprehensively determine the degree of duplication with existing results, and output clear suggestions for project establishment decisions, in order to overcome the technical defects of the existing technologies, such as insufficient semantic understanding ability, weak decision support ability, lack of multidimensional analysis ability, and strong subjectivity of evaluation results. Summary of the Invention

[0010] The purpose of this invention is to provide a scientific research topic deduplication auxiliary decision-making system and method that integrates large language models and vector retrieval, so as to solve the technical problems existing in the prior art, such as insufficient semantic understanding ability, weak decision support ability, lack of multi-dimensional analysis ability, and strong subjectivity of evaluation results.

[0011] To achieve the above-mentioned objectives, the present invention adopts the following technical solution.

[0012] Option 1: This invention provides a research topic deduplication auxiliary decision-making method that integrates large language models and vector retrieval, the method comprising: Obtain the text input of the topic to be investigated, preprocess the text of the topic to be investigated and extract structured information to obtain structured topic features; Perform multi-path parallel retrieval, which includes keyword-based retrieval, semantic vector-based retrieval, and knowledge graph-based retrieval. After fusing the retrieval results, obtain an internal result candidate set and an external patent candidate set, and use a large language model to reorder the candidate results. The comprehensive duplication score is calculated based on the internal candidate set of results. The calculation of the comprehensive duplication score includes: decomposing the topic to be investigated and the internal results into multiple preset dimensions, calculating the comprehensive similarity of each dimension, and obtaining the comprehensive similarity score by weighted fusion of the results of multiple text similarity algorithms. Then, the similarity scores of each dimension are weighted and summed to obtain the comprehensive duplication score. The comprehensive novelty score is calculated based on an external patent candidate set. The calculation of the comprehensive novelty score includes: using a large language model to decompose the innovation points of the subject to be investigated into multiple core dimensions, calculating the similarity between each dimension and the external patent candidate set, and then calculating the comprehensive novelty score based on the similarity of each dimension. Based on the overall repeatability score and the overall novelty score, a project initiation decision recommendation is output through a four-quadrant decision matrix.

[0013] Furthermore, the multiple text similarity algorithms include at least two of the following: word frequency-inverse document frequency cosine similarity, BM25 score, and Jaccard similarity coefficient; the multiple preset dimensions include at least two of the following: research content dimension, technical route dimension, expected results dimension, and innovation point dimension; the multiple core dimensions include at least two of the following: research purpose dimension, technical solution dimension, verification method dimension, and application scenario dimension.

[0014] Furthermore, the weighted fusion of the comprehensive similarity adopts configurable weight parameters, and the formula for calculating the comprehensive repetition is the weighted sum of the similarity of each dimension and the corresponding dimension weight. The dimension weights are configurable according to the focus of the novelty search task.

[0015] Furthermore, the comprehensive novelty score is calculated using the maximum value method.

[0016] Furthermore, the novelty level is determined based on the comparison between the comprehensive novelty score and the preset threshold; and / or, the duplication type is determined based on the comparison between the comprehensive duplication score and the preset threshold.

[0017] Furthermore, the four-quadrant decision matrix uses the overall repetition rate as the horizontal axis and the overall novelty score as the vertical axis. It maps the topics to four quadrants according to the comparison results of the two dimensions with their respective intermediate thresholds. Each quadrant corresponds to different project decision recommendations.

[0018] Option 2: This invention provides a research topic deduplication auxiliary decision-making system that integrates large language models and vector retrieval, the system comprising: The user interaction layer is used to obtain input on the research topic and display the decision results; The application service layer is used to schedule the novelty search process, generate reports, and provide decision-making suggestions. The intelligent analysis layer is used to execute the research topic deduplication auxiliary decision-making method as described in any one of claims 1-6; The search engine layer efficiently retrieves candidate documents related to the research topic from external patent databases and internal results databases, and performs fine sorting on the candidate set. The data storage layer persistently stores all the data required for the system to operate.

[0019] Furthermore, the user interaction layer includes a topic upload entry, a text paste input box, a novelty search report viewing panel, and a project proposal display area, and the user interaction layer presents a four-quadrant decision matrix in a visual manner.

[0020] Furthermore, the intelligent analysis layer deploys a multi-dimensional similarity calculation module, which encapsulates the weighted fusion calculation logic of TF-IDF cosine similarity, BM25 score and Jaccard similarity coefficient.

[0021] Furthermore, the intelligent analysis layer is equipped with a novelty quantification evaluation module and a repeatability analysis module; The novelty quantification assessment module is configured to work in conjunction with the multi-dimensional similarity calculation module to calculate the overall novelty score using the maximum value method; The repeatability analysis module is configured to work in conjunction with the multi-dimensional similarity calculation module to calculate the overall repeatability using a weighted sum formula.

[0022] Compared with the prior art, the solution provided by the present invention has the following beneficial effects: (1) This invention employs a multi-algorithm fusion similarity calculation method, comprehensively utilizing various text similarity calculation algorithms and combining the results of different algorithms based on a weighted fusion approach. This effectively avoids the limitations of a single algorithm and significantly improves the accuracy and recall of novelty searches. Compared to existing technologies that only use a single algorithm, this invention can comprehensively evaluate the degree of text similarity from multiple perspectives, such as literal matching, word frequency distribution, and set overlap. It has a higher recognition capability for repetitive tasks that use synonym replacement, sentence transformation, and other means to evade detection.

[0023] (2) This invention uses a novelty assessment method based on Facet to decompose the innovation points of scientific research projects into four core dimensions: research purpose, technical solution, verification method, and application scenario, and conducts quantitative assessments for each dimension. It also uses the maximum value method to determine the comprehensive novelty score, which can accurately identify the innovative value of the project in specific dimensions and avoid the information loss problem caused by binary judgment in the prior art.

[0024] (3) This invention uses a multi-level repetition determination model to perform weighted fusion calculations from four dimensions: research content, technical route, expected results, and innovation points. It can accurately quantify the degree of repetition between the research topic and existing internal results, and distinguish four different levels of repetition types: complete repetition, core repetition, partial repetition, and reference. This provides a refined and quantifiable basis for scientific research management.

[0025] (4) The invention uses a four-quadrant matrix decision-making framework to comprehensively consider two dimensions: the degree of duplication between the research topic and existing results and the degree of novelty of the research topic itself. It maps research topics to four quadrants: leading innovation, risky innovation, exploratory research, and not recommended, and outputs differentiated project decision-making suggestions respectively. This effectively solves the problem that existing technologies only output a single binary judgment, and helps decision-makers to more accurately identify the innovative value and potential risks of the research topic.

[0026] (5) This invention has been specifically optimized for the construction industry. By loading a custom dictionary for the construction industry and a standardized mapping of domain terms, the application effect of the system in the construction industry scientific research and novelty search scenario has been improved. Attached Figure Description

[0027] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0028] Figure 1 This is the overall architecture diagram of the scientific research topic deduplication auxiliary decision-making system that integrates large language model and vector retrieval in this invention.

[0029] Figure 2 This is a flowchart illustrating the workflow of the research topic deduplication auxiliary decision-making method that integrates large language models and vector retrieval in this invention.

[0030] Figure 3 This is a flowchart of the novelty assessment process in this invention.

[0031] Figure 4 This is a diagram of the multi-level repeatability determination model in this invention.

[0032] Figure 5 This is the four-quadrant matrix decision diagram in this invention. Detailed Implementation

[0033] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific illustrations.

[0034] To address the technical problems existing in current research topic selection technologies, such as insufficient semantic understanding, weak decision support capabilities, lack of multi-dimensional analysis capabilities, and strong subjectivity in evaluation results, this invention constructs a comprehensive research topic deduplication auxiliary decision-making system that integrates large language models and vector retrieval.

[0035] This invention innovatively employs a multi-algorithm fusion strategy for similarity calculation. Traditional novelty search systems rely solely on single algorithms such as BM25 or TF-IDF cosine similarity for text matching, each with inherent limitations. This invention weightedly fuses TF-IDF cosine similarity, BM25 score, and Jaccard similarity coefficient, enabling the system to comprehensively evaluate the similarity between two texts from three different mathematical perspectives: word frequency distribution, vector direction, and set overlap. This significantly reduces the false negative rate for duplicate topics.

[0036] This invention constructs a multi-level, configurable, dimensional weighted model, decomposing the repetition determination into four dimensions: research content, technical route, expected results, and innovative points, and configuring adjustable weight coefficients for each dimension. This allows the system to adapt to the diverse needs of different research institutions and disciplines.

[0037] See Figure 2 The diagram shows the specific implementation workflow of the scientific research topic deduplication auxiliary decision-making method that integrates large language models and vector retrieval provided by the present invention.

[0038] Based on the illustration, the research topic deduplication auxiliary decision-making method presented in this invention is mainly implemented through the following steps: Step S101: The user submits a research topic to be investigated.

[0039] This step is set to receive research project application materials input by the user, and to receive and parse the original text data submitted by the user through text pasting or file upload, and transmit the parsed plain text data and its task identifier to step S102.

[0040] Step S102, text preprocessing.

[0041] This step is set to clean and standardize the original text. For the original plain text data passed in step S101, operations such as format cleaning, character standardization and domain terminology segmentation are performed. The cleaned and standardized text is then transmitted to step S103, and the segmentation results are transmitted to step S104.

[0042] Step S103: Extraction of structured information.

[0043] This step is designed to transform unstructured natural language text into a structured description of the research topic's features. This step uses a large language model to perform named entity recognition and key information extraction on the standardized text input in step S102, extracting structured fields such as the research topic name, technical approach, innovative points, and research objectives. The structured feature description object is then transmitted to steps S104 and S106.

[0044] Step S104: Multi-path parallel retrieval.

[0045] This step is configured to retrieve candidate documents related to the topic to be searched from external patent databases and internal results databases. Based on the word segmentation results passed in step S102 and the keyword features passed in step S103, this step performs external patent keyword retrieval, internal results BM25 retrieval, internal results semantic vector retrieval, and internal results knowledge graph retrieval in parallel, and transmits the candidate set formed by fusing the three retrieval results to step S105.

[0046] Step S105: Large model reordering.

[0047] This step is designed to perform a fine-grained sorting of the candidate set and select the top 10 most relevant results. For the candidate set input in step S104, this step first uses a lightweight dual-tower model for coarse sorting, and then uses a large language model as a cross-encoder for fine sorting. The sorted top 10 external patents and top 10 internal results are then transferred to the novelty evaluation sub-process and the duplication determination sub-process in step S106, respectively.

[0048] Step S106, comprehensive evaluation.

[0049] This step is designed to calculate the novelty score and duplication score of the subject matter under investigation. This step executes two sub-processes in parallel: the novelty assessment sub-process calculates a comprehensive novelty score N based on the top 10 external patents input in step S105 through dimensional decomposition and basic similarity calculation; the duplication determination sub-process calculates a comprehensive duplication score R based on the top 10 internal results input in step S105 through dimensional weighted fusion. This step then transmits the calculated N and R values ​​to step S107.

[0050] Step S107, four-quadrant decision.

[0051] This step is configured to output project initiation decision recommendations based on the quantitative values ​​of novelty and repetition. Taking the N and R values ​​input in step S106, this step constructs a four-quadrant matrix with R as the horizontal axis and N as the vertical axis. Based on threshold values, it determines the quadrant to which the project belongs, generates corresponding decision recommendations and adjustment direction suggestions, and transmits the decision results to step S108.

[0052] Step S108: Report generated.

[0053] This step is designed to integrate the results of the entire evaluation process into a structured novelty search report. This step takes all intermediate results and final decisions generated in steps S101 to S107, including topic information, search abstracts, detailed similarity data for each dimension, N-value, R-value, novelty level, duplication type, quadrant determination results, and decision recommendations, and populates them into the report template to generate a structured novelty search report, which is then pushed to the user.

[0054] The following section provides a further explanation of the specific implementation schemes for each step in this method.

[0055] In some embodiments of the present invention, step S101 in the scientific research topic deduplication auxiliary decision-making method includes the following sub-steps: Sub-step S101-1: Input method adaptation. The system provides two input entry points on the front-end interface: a text paste input box and a file upload component. The text paste input box allows users to directly enter or paste text content from project application materials, while the file upload component allows users to upload document files in common formats.

[0056] Sub-step S101-2: Text extraction and format recognition. After the user submits a request, the system first determines the input type. If it is a file upload, the system uses the corresponding parsing library to extract the text based on the file extension. If it is text pasting, the system directly obtains the content of the input area.

[0057] Sub-step S101-3: Task initialization. The system generates a globally unique task identifier for this novelty search request and records the submission timestamp and user ID. The original text and the task identifier are stored together in a temporary storage area, and the task identifier is simultaneously written to the processing queue.

[0058] Sub-step S101-4: Start scheduling. The system calls the new query scheduling service, passing the task identifier and the original text path as parameters to the processing function in step S102, triggering the subsequent process.

[0059] This implementation scheme supports multiple input formats and enables automatic file parsing without requiring manual text preprocessing by the user; at the same time, the generated task identifier is used throughout all steps of the process for log association and result tracking, while the original plain text data serves as the sole input for step S102.

[0060] In some embodiments of the present invention, step S102 in the scientific research topic deduplication auxiliary decision-making method includes the following sub-steps: Sub-step S102-1: Format cleaning. This step receives the original text string from step S101 and performs the following cleaning operations: remove consecutive extra spaces at the beginning and end of the text and within the text; delete special characters, including bullet points, decorative characters, tabs, and invisible control characters; remove header and footer areas; convert full-width letters, numbers, and punctuation marks to half-width characters; and convert traditional Chinese characters to simplified Chinese characters.

[0061] Sub-step S102-2: Terminology Standardization. This step loads a custom dictionary file for the construction industry. This dictionary is stored with one entry per line, covering terms from professional categories such as earthwork engineering, foundation engineering, main structure, decoration and renovation, and electromechanical installation. The entries in the dictionary take precedence over the word segmentation results of the default dictionary.

[0062] Sub-step S102-3: Word segmentation. The system initializes the word segmentation tool, loads the aforementioned custom dictionary, and then performs precise word segmentation on the cleaned text to ensure that complex terms such as "large-span spatial grid structure" are recognized as a whole.

[0063] Sub-step S102-4: Output assembly. The system saves the cleaned, standardized text as an independent field, and also saves the word list after segmentation, both of which are associated with the current task identifier.

[0064] This step performs domain-adaptive word segmentation optimization for the construction industry. By using a custom dictionary, it solves the problem of inaccurate segmentation of professional terms by general word segmenters, and provides high-fidelity term sequences for subsequent TF-IDF vectorization, BM25 retrieval, and semantic encoding.

[0065] In some embodiments of the present invention, step S103 in the scientific research topic deduplication auxiliary decision-making method includes the following sub-steps: Sub-step S103-1: Prompt word template construction. This step is set to read a prompt word template suitable for scientific research topic parsing from the configuration library. The template includes three parts: task description, output format constraints, and field definitions, explicitly requiring the extraction of the topic name, research background, research objectives, technical route, expected results, key innovation points, and key problems to be solved.

[0066] Sub-step S103-2: Model Invocation. This step is set up to fill the specified position of the prompt word template with the standardized text output from step S102, construct a complete request message, and then invoke the large language model service via API.

[0067] Sub-step S103-3: Response Parsing and Validation. This step is configured to receive a string returned by the model and attempt to parse it into a JSON object. Upon successful parsing, it checks that each required field exists and is not empty. If a field is missing or parsing fails, the system triggers a fallback strategy: using regular expression-based rule extraction as an alternative.

[0068] Sub-step S103-4: Structured Output Assembly. This step is configured to integrate the successfully parsed fields and the downgraded supplemented fields into a structured task feature description object, which includes the aforementioned fields as well as the original task identifier.

[0069] In some embodiments of the present invention, step S104 in the scientific research topic deduplication auxiliary decision-making method is implemented by dividing it into two links: external patent retrieval and internal achievement retrieval, and is achieved through the following sub-step scheme: Sub-step S104-1: External Patent Search. This step is set up to extract the core terms from the topic name, keywords, and technical routes from the structured features output in step S103, and construct a Boolean search expression; based on this, a search request is initiated to an external patent database via API, and after receiving the returned results, the metadata such as the title, abstract, and application number of each patent is encapsulated into external patent candidate objects.

[0070] Internal results retrieval is based on a three-way parallel recall strategy: Sub-step S104-2: Internal Results BM25 Keyword Recall. This step is set up to use the word segmentation results output in step S102 as query terms to initiate a search request to the full-text database of internal results, using BM25 as the default scoring algorithm for the query statement.

[0071] Sub-step S104-3: Internal result semantic vector retrieval. This step is configured to use a semantic encoding model to encode the standardized text output from step S102 into semantic vectors, and then initiate an approximate nearest neighbor search request to the vector database, returning a list of results with the smallest distance to the query vector.

[0072] Sub-step S104-4: Internal outcome knowledge graph retrieval. This step is configured to extract a list of technical element nouns from the structured features output in step S103, and then perform a query in the pre-built knowledge graph to return all outcome nodes that satisfy the association relationship.

[0073] Sub-step S104-5: Fusion and Ranking. This step is configured to merge the results from BM25 recall, semantic vector recall, and knowledge graph recall, and calculate a weighted inverse fusion score for each candidate result. The weights of the three retrievers are equal by default, and the results are sorted in descending order of fusion score. The top few candidate results are selected as the internal result candidate set.

[0074] As an example, the weighted inverse fusion score here can be calculated using the following formula: Score(c) = Σ_{i=1}^{3} w_i / (rank_i(c) + 1); Where rank_i(c) is the rank number of candidate result c in the i-th retrieval result (starting from 1), and w_i is the weight of the i-th retrieval machine, which is 1 / 3 by default.

[0075] This step-by-step approach, through the complementarity of three retrieval strategies, greatly improves the comprehensiveness of the recall.

[0076] In some embodiments of the present invention, step S105 in the scientific research topic deduplication auxiliary decision-making method includes the following sub-steps: Sub-step S105-1: Coarse Ranking. This step is set to load a lightweight dual-tower model. The standardized text of the topic to be investigated and the text of each candidate result output from step S104 are respectively input into the dual-tower model to obtain sentence vectors and calculate the similarity as the coarse ranking score. The results are sorted in descending order of the coarse ranking score, and the top few candidate results are selected for the fine ranking stage. This sub-step uses the dual-tower model for fast calculation and can quickly narrow down the scope of fine ranking.

[0077] Sub-step S105-2: Fine-ranking input construction. This step is set to concatenate the standardized text of the topic to be searched with the text of each fine-ranking candidate result, separated by delimiters, to construct an input format suitable for large language models.

[0078] Sub-step S105-3: Fine-tuning of the large language model. This step involves inputting the concatenated text into the large language model one by one, requiring the model to output a relevance score between 0 and 1 to evaluate the degree of matching between the two in terms of research content, technical methods, and innovative points.

[0079] Sub-step S105-4: Sorting Output. This step is set to sort the candidate results in descending order according to the relevance scores output by the large language model, and take the top 10 as the Top 10 external patents and the Top 10 internal results.

[0080] The resulting S105 scheme utilizes a large language model as a cross-encoder, enabling it to simultaneously process the interactive information between the research topic and candidate results, achieving significantly better ranking results than the dual-tower model. The two-stage "coarse-fine ranking" strategy further reduces computational costs while maintaining ranking quality.

[0081] In some embodiments of the present invention, step S106 in the scientific research topic deduplication auxiliary decision-making method includes two parallel sub-processes: a novelty evaluation sub-process and a duplication determination sub-process, both of which rely on a reusable basic similarity calculation unit. The specific implementation of the basic similarity calculation unit will be described first, followed by descriptions of the two sub-processes.

[0082] The basic similarity calculation unit takes two text segments (text A and text B) as input and generates a fusion similarity score as output. This is achieved through the following scheme: First, word segmentation is performed. A word segmentation tool is used, and a custom dictionary for the construction industry is loaded to accurately segment the two input text segments.

[0083] Next, TF-IDF cosine similarity is calculated: a corpus containing two text segments is constructed, and the term frequency TF(t,d)=freq(t,d) / max_len(d) and inverse document frequency IDF(t)=log(N / n(t)) are calculated. Based on these, TF-IDF feature vectors are constructed. TF-IDF(t,d) = TF(t,d) × IDF(t).

[0084] Convert the TF-IDF feature vectors to text vectors, calculate the TF-IDF feature vectors, and then calculate the cosine similarity S_cosine, that is, calculate the cosine similarity between text A and text B: S_cosine = (A · B) / (||A|| × ||B||) Next, the BM25 score is calculated. The BM25 algorithm is used to saturate the term frequencies and consider document length to calculate the relevance score between the query and the document. S_bm25 = Σ IDF(qi) × fi × (k1+1) / (fi + k1 × (1-b + b × dl / avgdl)); The typical value of parameter k1 is 1.2, and the typical value of b is 0.75.

[0085] Next, the Jaccard similarity coefficient is calculated, which measures the degree of overlap between the two sets. Specifically, the term lists of the two texts are converted into sets, and the ratio S_jaccard of the intersection to the union is calculated: S_jaccard = |A ∩ B| / |A ∪ B|.

[0086] Next, the similarity is fused. The results of the three algorithms are weighted and fused to obtain the comprehensive similarity score S_fusion: S_fusion = w_cosine × S_cosine + w_bm25 × S_bm25 + w_jaccard × S_jaccard; Where w_cosine, w_bm25, and w_jaccard are the fusion weights of each algorithm, satisfying: w_cosine + w_bm25 + w_jaccard = 1; The determination of this weight configuration is based on two levels: the first level is the basis for the initial weight setting, which, based on research findings in the field of information retrieval, defaults to w_cosine equaling 0.4, w_bm25 equaling 0.4, and w_jaccard equaling 0.2; the second level is the weight optimization method, which utilizes historical novelty search data to perform supervised learning optimization of the weights during the scheme's operation.

[0087] As an example, the similarity fusion weight can be updated periodically based on historical novelty search data through the offline weight optimization module.

[0088] As a further example, the offline weight optimization module can update the similarity fusion weights through the following steps: (1) Training data construction: 2,000 pairs of topics that have been confirmed to be duplicates or novel were extracted from the historical operation records of the system. Each data pair contains a text pair of topic A and topic B, as well as the similarity level (highly similar, partially similar, and dissimilar) manually labeled.

[0089] (2) Feature extraction: For each topic pair, calculate the three basic similarity values: TF-IDF cosine similarity, BM25 score, and Jaccard similarity coefficient to form a three-dimensional feature vector [x1, x2, x3].

[0090] (3) Model training: The LambdaMART algorithm was used as the learning ranking model. The loss function was defined as cross-entropy loss, and the evaluation index was NDCG@10 (normalized depreciation cumulative gain). The dataset was divided into training set and validation set in an 8:2 ratio. The optimal weight combination was trained to make NDCG@10 on the validation set reach more than 0.85.

[0091] (4) Weight update: The optimal weights obtained during training (e.g., w_cosine=0.35, w_bm25=0.50, w_jaccard=0.15) are automatically written to the system configuration center and take effect in real time in subsequent new search tasks.

[0092] Using the methods described above, this system can adapt to the data distribution across different disciplines and continuously improve the accuracy of similarity calculation.

[0093] By invoking the aforementioned basic similarity calculation unit, this step uses the novelty assessment sub-process to calculate the novelty score N of the subject matter under investigation relative to externally published patent literature. Novelty assessment determines the innovative value of the subject matter by comparing the degree of difference between the subject matter under investigation and similar externally retrieved patent literature across various dimensions.

[0094] Specifically, the novelty assessment sub-process in this step is implemented through the following sub-step scheme: Sub-step S106-N-1: Perform external similar patent screening, and take all 10 patents from the Top 10 external patents output in step S105 as the comparison set.

[0095] Sub-step S106-N-2: Perform innovation point dimension decomposition, call the same innovation point extraction module as step S103, decompose the innovation points of the research topic to be investigated into four core dimensions: research purpose, technical solution, verification method and application scenario, and extract the text of each dimension of the comparative patent.

[0096] Specifically, this step uses a large language model to decompose the innovative points of the research topic into four core dimensions (Facet): research purpose (Purpose), technical solution (Mechanism), verification method (Evaluation), and application scenario (Application).

[0097] For each dimension, extract the text of the research topic in that dimension (denoted as I_purpose, I_mechanism, I_evaluation, I_application) and the text of the comparative patent in that dimension (denoted as P_purpose, P_mechanism, P_evaluation, P_application).

[0098] Sub-step S106-N-3: Perform similarity calculations for each dimension. For each dimension of the research topic and each comparative patent, call the basic similarity calculation unit to calculate the comprehensive similarity of that pair in that dimension: S_purpose_i, S_mechanism_i, S_evaluation_i, S_application_i. Then, take the maximum value of the similarity of all comparative patents in the same dimension as the final similarity of that dimension: that is, S_purpose = max_i S_purpose_i. Similarly, obtain S_mechanism, S_evaluation, and S_application.

[0099] Taking the research purpose dimension as an example, S_purpose_i is calculated for the i-th comparative patent, and then max(S_purpose_1, ..., S_purpose_10) is taken as the final similarity S_purpose for this dimension.

[0100] The formulas for calculating similarity in each dimension are as follows: S_purpose = max(S_purpose_i) for i in Top10_patents; S_mechanism = max(S_mechanism_i) for i in Top10_patents; S_evaluation = max(S_evaluation_i) for i in Top10_patents; S_application = max(S_application_i) for i in Top10_patents; Where S_purpose_i equals the comprehensive similarity between the research topic to be investigated and the i-th patent in the research purpose dimension, calculated by calling the basic similarity calculation unit.

[0101] Sub-step S106-N-4: Derivation of novelty for each dimension, and calculation of novelty score for each dimension based on similarity. Novelty is calculated based on the principle that "the higher the similarity to external patents, the less novel the topic is in that dimension." Specifically, the novelty score for each dimension is calculated according to the formula N_dim = 1 - S_dim.

[0102] Specifically, the formulas for calculating the novelty scores for each dimension in this step are as follows: N_purpose = 1 - S_purpose; N_mechanism = 1 - S_mechanism; N_evaluation = 1 - S_evaluation; N_application = 1 - S_application.

[0103] Sub-step S106-N-5: Calculate the overall novelty score. This step integrates the novelty scores from four dimensions and uses the maximum value method to calculate the overall novelty score.

[0104] Specifically, this step is based on the principle that "as long as there is a substantial difference from existing external patents in any core dimension, the topic has novelty value," and calculates the comprehensive novelty score N using the maximum value method. The corresponding calculation formula is as follows: N = 1 - max(S_purpose, S_mechanism, S_evaluation, S_application).

[0105] Sub-step S106-N-6: Determine the novelty level. This step determines the novelty level based on the calculated comprehensive novelty score N.

[0106] Specifically, the rules for determining the novelty level are as follows: If N is greater than 0.8, the novelty level is novel. If N is greater than 0.6 and less than or equal to 0.8, the novelty level is significant improvement. If N is greater than 0.3 and less than or equal to 0.6, the novelty level is slight improvement. If N is less than or equal to 0.3, the novelty level is known technology application.

[0107] This novelty assessment subprocess ensures comparability between different topics through standardized decomposition of fixed four dimensions, while the maximum value method can keenly identify the high degree of overlap between the topic and existing technologies in a certain dimension.

[0108] Similarly, by calling the basic similarity calculation unit mentioned above, this step uses the duplication determination subprocess to calculate the overall duplication rate R between the topic under investigation and existing internal achievements. The duplication determination assesses whether there is a risk of duplicate project approval by comparing the similarity between the topic under investigation and existing internal achievements in various dimensions.

[0109] For details, see Figure 4 The repetition determination sub-process in this step is implemented through the following sub-step scheme: Sub-step S106-R-1: Perform internal similarity result screening. This step takes all 10 articles from the Top 10 internal results output in step S105 as the comparison set.

[0110] Sub-step S106-R-2: Perform dimensional text extraction, decompose the research topic to be investigated and each comparative result into four dimensions: research content, technical route, expected results and innovation points. The text of the research topic to be investigated in the research content dimension is denoted as T_content, and the text of the comparative results in the research content dimension is denoted as P_content. The other dimensions are treated similarly.

[0111] Sub-step S106-R-3: Perform similarity calculation for each dimension. Call the basic similarity calculation unit for each dimension of the topic to be investigated and each comparison result, calculate the comprehensive similarity S_content_i, S_method_i, S_output_i, and S_innov_i for that dimension, and take the maximum value as the final similarity S_content = max_i S_content_i for that dimension. Similarly, obtain S_method, S_output, and S_innov.

[0112] Taking the research content dimension as an example, S_content_i is calculated for the i-th comparison result, and then max(S_content_1, ..., S_content_10) is taken as the final similarity S_content for this dimension.

[0113] The formula for calculating the similarity fusion across all dimensions is as follows: S_content = max(S_content_i) for i in Top10_internal; S_method = max(S_method_i) for i in Top10_internal; S_output = max(S_output_i) for i in Top10_internal; S_innov = max(S_innov_i) for i in Top10_internal.

[0114] Sub-step S106-R-4: Calculate the overall similarity score. This step calculates the overall similarity score R based on the similarity across four dimensions using a weighted sum.

[0115] The overall repeatability R is configured to be calculated based on the following formula: R = α × S_content + β × S_method + γ × S_output + δ × S_innov; Where α, β, γ, and δ are the weight coefficients for each dimension, satisfying α + β + γ + δ = 1. The setting of dimension weights is based on the general principles in the field of scientific research novelty search, with research content and technical route being the core dimensions for judging the degree of duplication of a topic. As an example, the default weight configuration is α = 0.3, β = 0.3, γ = 0.2, and δ = 0.2.

[0116] Sub-step S106-R-5: Determine the repetition type. This step determines the repetition type based on the comparison between the overall repetition degree R and the preset threshold.

[0117] Specifically, the rules for determining duplicate types are as follows: If R is greater than or equal to 0.8, the duplication type is complete duplication. If R is greater than or equal to 0.6 and less than 0.8, the repetition type is core repetition. If R is greater than or equal to 0.4 and less than 0.6, the repetition type is partial repetition. If R is less than 0.4, the repetition type is either reference or completely novel.

[0118] The repetition determination subprocess based on the above scheme allows for weight adjustment to adapt to different management needs, thereby quantifying the degree of repetition and providing operational repetition type labels.

[0119] The implementation of step S106 decouples novelty assessment and duplication determination into two independent sub-processes. These two processes can be executed in parallel without interference and share the basic similarity calculation unit, improving code reusability and execution efficiency. The comprehensive novelty score N and comprehensive duplication score R output from this step are then passed to step S107.

[0120] In some embodiments of the present invention, step S107 of the scientific research topic deduplication auxiliary decision-making method integrates the novelty score N (from the novelty assessment process) and the duplication rate R (from the duplication determination process) to output a project establishment decision suggestion.

[0121] This step determines the position of the research topic in the four-quadrant matrix based on the values ​​of N and R calculated in step S106. See also... Figure 5 In this invention, the four-quadrant matrix uses the overall repeatability R as the horizontal axis (X-axis) and the overall novelty score N as the vertical axis (Y-axis), with the midpoint thresholds of R_mid = 0.5 and N_mid = 0.5, respectively. These two threshold lines divide the plane into four quadrants.

[0122] The resulting four-quadrant matrix's coordinate axis division and decision implications are as follows: Quadrant 1 (Innovation Leading, N ≥ 0.5 and R < 0.5): This topic has high novelty and low repetition with existing results. It belongs to a highly innovative and independent research direction and is recommended for priority approval.

[0123] Quadrant 2 (Innovation with risks, N ≥ 0.5 and R ≥ 0.5): The topic has some novelty but has a high degree of duplication with existing results. It is necessary to focus on demonstrating the differences from existing research. It is recommended to adjust the research focus and re-evaluate.

[0124] The third quadrant (exploratory research, N less than 0.5 and R less than 0.5): This topic has low novelty and low repetition with existing results. It may be an exploratory but not clearly innovative study. It is recommended to further refine the innovative points.

[0125] Quadrant 4 (Not recommended, N < 0.5 and R > 0.5): This topic is highly repetitive with existing results and lacks novelty. It is considered repetitive or low-value research and is not recommended for funding.

[0126] Based on this, the four-quadrant decision-making process in step S107 is as follows: Sub-step S107-1: Threshold comparison. This step is set to receive the N and R values ​​passed in from step S106, and map the topic to the corresponding position in the constructed four-quadrant matrix according to the values ​​of N and R.

[0127] Sub-step S107-2: Quadrant I determination. When N ≥ 0.5 and R < 0.5, the topic falls in the first quadrant: Innovation Leading. This area represents a topic with high novelty (N above the median) and low overlap with existing results (R below the median), making it the most ideal candidate for project approval. The output decision is "Recommend Project Approval," with an additional explanation in the novelty search report: "This topic is highly innovative and independent of existing research; it is recommended to prioritize its approval and increase resource investment." Sub-step S107-3: Second Quadrant Judgment. When N ≥ 0.5 and R ≥ 0.5, the topic falls in the second quadrant: Innovation is at risk. This area represents a topic that, while possessing some novelty, also has a high degree of overlap with existing research, posing a risk that its innovativeness will be overshadowed by existing studies. The output decision is "Cautious Project Approval," with the following explanation: "This topic possesses some innovativeness, but it has a high degree of overlap with existing research. It is necessary to focus on demonstrating the differences from existing research. It is recommended to adjust the research focus and re-evaluate." Sub-step S107-4: Third Quadrant Determination. When N < 0.5 and R < 0.5, the topic falls into the third quadrant: exploratory research. This area represents a topic with low novelty but also low repetition with existing results. It may be exploratory research, but the innovative point is not clear enough. The output decision is "Cautious Project Approval," with the following explanation: "This topic has low novelty and low repetition with existing results. It may be exploratory research, but the innovative point is not clearly defined. It is recommended to further refine the innovative point before further evaluation." Sub-step S107-5: Quadrant IV determination. When N < 0.5 and R ≥ 0.5, the research topic falls in the fourth quadrant: Not recommended. This area represents a research topic that highly overlaps with existing findings and lacks novelty, belonging to typical repetitive or low-value research. The output decision is "Not recommended for project initiation," with the following explanation: "This research topic highly overlaps with existing findings and lacks novelty, belonging to repetitive or low-value research, and is not recommended for project initiation." Sub-step S107-6: Generation of Personalized Adjustment Suggestions. This step generates additional adjustment direction suggestions based on the similarity contribution of each dimension. The specific rules are as follows: if S_content in the repetition assessment is greater than 0.7, an additional suggestion is added: "It is recommended to clearly distinguish the research content from existing results"; if S_method is greater than 0.7, an additional suggestion is added: "It is recommended to seek differentiated breakthroughs in technical routes"; if S_application in the novelty assessment is greater than 0.7, an additional suggestion is added: "It is recommended to adjust the application scenario positioning to enhance novelty." Similar rules apply to other dimensions.

[0128] The aforementioned four-quadrant decision-making process transforms two opposing quantitative indicators into a synergistic decision-making basis. The four-quadrant division gives the decision space a clear semantic meaning, effectively converting abstract numerical results into intuitive and operable management instructions, reducing the cognitive burden and decision-making bias of research managers. The quadrant determination results, decision suggestions, and personalized prompts output in this step are then transmitted to step S108.

[0129] In some embodiments of the present invention, step S108 in the scientific research topic deduplication auxiliary decision-making method includes the following sub-steps: Sub-step S108-1: Data Aggregation. This step obtains basic information about the research topic from step S103, including the topic name and submission time. It obtains a summary of the search results from steps S104 and S105, including the titles, abstracts, and relevance scores of the top 5 external patents and the top 5 internal research results. Step S106 obtains detailed similarity data for each dimension, the overall duplication rate (R), the overall novelty score (N), the novelty level, and the duplication type. Step S107 obtains the quadrant determination results, decision suggestions, and personalized adjustment prompts.

[0130] Sub-step S108-2: Template Rendering. This step reads a pre-made HTML report template file. For example, this template uses Jinja2 syntax to define the report's structure and style. The template contains SVG code for a four-quadrant matrix diagram, marking the topic's location on the diagram with red dots based on the quadrant determination results. The template also includes a loop structure for a similarity comparison table, filling the table with similarity data for each dimension, and highlighting values ​​exceeding a threshold of 0.7 with a red background. In the original text key segment comparison section, the sentence most similar to the topic to be investigated is extracted from the comparison results, and the text segment with the highest similarity is located using a similarity backtracking algorithm.

[0131] Sub-step S108-3: PDF Generation. This step takes the rendered HTML report as input and calls the wkhtmltopdf command-line tool for conversion. Conversion parameters are set to A4 page size, 15mm margins on all sides, and JavaScript support is enabled to correctly render charts. The generated PDF file is saved to the specified output directory.

[0132] Sub-step S108-4: Storage and Push. This step stores the generated HTML and PDF files on a file server, using the naming convention of task identifier + underscore + timestamp + extension. The system pushes the report download link and summary information to the user through the user interaction layer, while persisting all intermediate data and final results of the entire novelty search task to a structured database.

[0133] Based on the report generation steps outlined above, users are provided with transparent and traceable evaluation data. The structured report format facilitates archiving and subsequent review and analysis. Highlighting and text-to-speech functions enable users to quickly locate key evidence of duplication or innovation, enhancing the credibility of their decisions.

[0134] The research topic deduplication and decision support method integrating large language models and vector retrieval presented in this example can be developed into a corresponding software program to form a research topic deduplication and decision support system integrating large language models and vector retrieval. When running, this software program will execute the aforementioned research topic deduplication and decision support method process integrating large language models and vector retrieval, and store the results in a suitable storage medium for the processor to retrieve and execute.

[0135] like Figure 1 As shown, the scientific research topic deduplication auxiliary decision-making system integrating large language model and vector retrieval provided by the present invention has an overall architecture comprising five components, which are, from top to bottom, a user interaction layer 100, an application service layer 200, an intelligent analysis layer 300, a retrieval engine layer 400, and a data storage layer 500.

[0136] The user interaction layer 100 in this system is configured to receive texts of research topics submitted by users and display evaluation results and novelty search reports to them. This layer is directly aimed at research administrators or project applicants, providing multiple input methods and a visual output interface.

[0137] The application service layer 200 is configured to schedule the execution order of various tasks in the entire novelty search process, manage the task queue, and generate a structured novelty search report and decision recommendations after evaluation. This layer acts as an intermediate coordination layer, receiving requests from the user interaction layer 100, calling lower-level modules to complete specific processing, and returning the results to the user interaction layer 100.

[0138] The intelligent analysis layer 300 is configured to execute the core semantic understanding and quantitative evaluation algorithms, including extracting structured features from unstructured text, calculating multi-dimensional similarity, and evaluating novelty and repetition.

[0139] The retrieval engine layer 400 is configured to efficiently retrieve candidate documents related to the research topic from external patent databases and internal results repositories, and to perform refined ranking of the candidate set. This layer ensures the comprehensiveness of the recall and the accuracy of the ranking through multi-path retrieval and re-ranking strategies.

[0140] The data storage layer 500 is configured to persistently store various data required for the system's operation, including semantic vectors from the vector database, raw text from the full-text database, and user information, task records, and configuration parameters from the structured database. This layer provides a data access interface to the retrieval engine layer 400 and provides data persistence support to the application service layer 200.

[0141] Based on this, the research topic deduplication auxiliary decision-making system consists of a user interaction layer 100 initiating a novelty search request by calling the API interface of the application service layer 200. The application service layer 200 then calls the functional modules of the intelligent analysis layer 300 and the retrieval engine layer 400 according to the request type. The innovation point extraction module in the intelligent analysis layer 300 needs to call the large language model API, and the multi-dimensional similarity calculation module needs to call the retrieval results of the retrieval engine layer 400. The multi-path recall module of the retrieval engine layer 400 needs to access the full-text database, vector database, and knowledge graph in the data storage layer 500. The data storage layer 500 provides data read and write services to the other layers. Overall, a clear call chain is formed: "user interaction → application scheduling → intelligent analysis / retrieval engine → data storage".

[0142] The following section provides a further explanation of the specific configuration of each logical functional layer in this system solution.

[0143] In some embodiments of the present invention, the user interaction layer 100 in the system mainly consists of three functional units: an input unit, a display unit, and a configuration unit.

[0144] The input unit includes a topic upload entry 110 and a text paste input box 120; the display unit includes a novelty search report viewing panel 130 and a project proposal display area 140; and the configuration unit is the administrator backend.

[0145] Furthermore, the input unit supports parsing .docx and .pdf files, and the front end uses a file parsing library to extract text content.

[0146] The exhibition units are presented in a visual way. Figure 5 The diagram shows a four-quadrant matrix, with different colored markers indicating the location of the current topic. The report viewing panel supports online HTML preview and PDF download. The configuration unit, visible only to administrators, is used to adjust various weight parameters and thresholds in the system.

[0147] Based on this, the user interaction layer is preferably designed with responsiveness, which can adapt to different screen sizes (PC, tablet, mobile phone). All user input is validated in the front-end (such as non-empty validation and text length limit) before submission. This allows complex quantitative evaluation results to be presented intuitively through a four-quadrant graph, reducing the user's understanding threshold.

[0148] In some embodiments of the present invention, the application service layer 200 of the system is equipped with a novelty search scheduling service 210, a report generation service 220, and a decision suggestion service 230.

[0149] The novelty search scheduling service 210 is implemented based on an asynchronous task queue. Internally, it maintains a directed acyclic graph (DAG) definition, which clearly defines the dependencies between the aforementioned steps S101 to S108 (for example, S104 must wait for S103 to complete, and S106 must wait for S105 to complete). When a user submits a request, the service generates a unique task ID, enqueues the task, and schedules the execution of each subtask according to the topological order of the DAG.

[0150] The report generation service 220 is configured to render HTML reports using the Jinja2 template engine and call the wkhtmltopdf tool to generate PDF versions; the service also implements the function of customizing report templates, allowing companies to add their own logos and report styles.

[0151] The decision suggestion service 230 maintains a suggestion text library. Based on the quadrant judgment results and the similarity contribution of each dimension, it selects basic suggestion texts from the library and dynamically generates personalized suggestions through a rule engine (adding a "high degree of duplication of research content" prompt if S_content > 0.7). The innovation of the application service layer lies in decoupling the three concerns of task scheduling, report generation, and decision suggestion, so that each part can evolve independently (for example, changing the report template does not affect the scheduling logic).

[0152] In some embodiments of the present invention, the intelligent analysis layer 300 of the system deploys four core modules: an innovation point extraction module 310, a multi-dimensional similarity calculation module 320, a novelty quantification evaluation module 330, and a repeatability analysis module 340.

[0153] Among them, the innovation point extraction module 310 internally maintains multiple prompt word templates for large language models (which may have slight differences for different disciplines), and implements a retry mechanism and degradation strategy for model calls (if the large language model API times out, rule-based keyword extraction is used as an alternative).

[0154] The multi-dimensional similarity calculation module 320 encapsulates the aforementioned basic similarity calculation unit and can execute the basic similarity calculation process, including word segmentation, TF-IDF cosine similarity calculation, BM25 score calculation, Jaccard similarity coefficient calculation, and comprehensive similarity fusion. It also implements a cache of calculation results internally (LRU strategy, maximum capacity 1000 records). For repeated calculations of the same text pair, the cached results are returned directly, which greatly improves the efficiency of batch processing.

[0155] The novelty quantification assessment module 330 is configured to work in conjunction with the multi-dimensional similarity calculation module 320. By calling the multi-dimensional similarity calculation module 320 to execute the specific steps of sub-steps S106-N-1 to S106-N-6 in the novelty assessment sub-process, the comprehensive novelty score N is calculated and evaluated.

[0156] The repeatability analysis module 340 is configured to work in conjunction with the multi-dimensional similarity calculation module 320. By calling the multi-dimensional similarity calculation module 320 to execute the specific steps of sub-steps S106-R-1 to S106-R-5 in the repeatability determination sub-process, the comprehensive repeatability R is calculated and evaluated.

[0157] The novelty quantification assessment module 330 and the repeatability analysis module 340 both rely on the multi-dimensional similarity calculation module 320 for operation, but they independently manage dimension definitions and weight configurations.

[0158] The resulting intelligent analysis layer uses standardized JSON objects for the input and output of all modules, facilitating data transfer and logging between modules. Through modular decoupling between the novelty quantification evaluation module 330 and the repeatability analysis module 340, when the dimensions of novelty evaluation need to be increased from four to five or more, only the code of the novelty quantification evaluation module 330 needs to be modified, without affecting the normal operation of other modules.

[0159] In some embodiments of the present invention, the retrieval engine layer 400 in the system includes a multi-path recall module 410, a fusion sorting module 420, and a re-sorting module 430.

[0160] The multi-path retrieval module 410 internally encapsulates three retrieval tools: the BM25 retrieval tool (calling Elasticsearch's match query), the vector retrieval tool (calling Milvus's approximate nearest neighbor search interface, using the HNSW index), and the knowledge graph retrieval tool (calling Neo4j's Cypher query). These three retrieval tools execute in parallel, each returning a Top 200 result.

[0161] The fusion sorting module 420 implements a weighted inverse fusion algorithm. Each retrieval has a default weight of 1 / 3, and the first 100 results are taken after the fusion score is calculated.

[0162] The reordering module 430 implements the aforementioned two-level strategy of "coarse ranking - fine ranking": the coarse ranking stage uses the MiniLM dual-tower model (accelerated by ONNX runtime), and the fine ranking stage calls the large language model (each call simultaneously inputs the text of the topic to be searched and the text of the candidate results).

[0163] The resulting retrieval engine layer employs a multi-stage pipeline design, where coarse-ranking results are directly used as input for fine-ranking, reducing unnecessary calls to large language models. This effectively balances the relationship between retrieval recall, efficiency, and accuracy.

[0164] In some embodiments of the present invention, the data storage layer 500 in the system includes a vector database 510 (Milvus), a full-text database 520 (Elasticsearch), and a structured database 530 (PostgreSQL).

[0165] Among them, the vector database 510 stores the semantic vectors (768-dimensional floating-point numbers) of all internal results, as well as the corresponding metadata (result ID, title, abstract).

[0166] Two types of indexes have been established in the full-text database 520: internal full-text indexes of research results and external full-text indexes of patents. Both of their analyzers are configured with custom dictionaries for the construction industry.

[0167] The structured database 530 stores user information tables, task record tables, configuration parameter tables (including fields such as w_cosine, w_bm25, w_jaccard, α, β, γ, δ, R_mid, N_mid, T1, T2, T3, etc.) and log tables.

[0168] The update operations (such as adding internal results) of the vector database and full-text database in the data storage layer 500 formed by this are carried out in an asynchronous batch manner to avoid blocking the main process; the modification of the configuration parameter table is synchronized to the cache of the application service layer in real time through triggers; at the same time, this data storage layer 500 has customized a full-text search analyzer for professional terms in the construction industry, which can significantly improve the accuracy of keyword search.

[0169] The research topic deduplication auxiliary decision-making system based on the above scheme, which integrates large language model and vector retrieval, mainly includes four stages during operation: input, retrieval, evaluation, and decision.

[0170] During the input phase, users submit the text of the topic to be searched through the topic upload portal 110 or the text paste input box 120 in the user interaction layer 100. After performing front-end format validation on the input, the user interaction layer 100 sends the text content along with the task identifier to the application service layer 200. Upon receiving the request, the novelty search scheduling service 210 of the application service layer 200 generates a unique task ID for this task and pushes the task into an asynchronous queue. According to the predefined DAG dependency relationship, the novelty search scheduling service 210 first triggers the subtask of text preprocessing in step S102.

[0171] During the retrieval phase, the novelty search scheduling service 210 sequentially triggers step S104 for multi-path parallel retrieval and step S105 for large model re-ranking. The multi-path recall module 410 of the retrieval engine layer 400 accesses the full-text database 520 (BM25 retrieval), the vector database 510 (semantic vector retrieval), and the knowledge graph (technical element retrieval) in parallel. The three-path recall results are sent to the fusion ranking module 420 for weighted inverse fusion, and the top 100 candidate results are selected. Subsequently, the re-ranking module 430 starts a two-level strategy of coarse ranking and fine ranking: first, a lightweight dual-tower model is used to quickly score the 100 candidate results, and the top 30 are selected for fine ranking; then, the large language model (Cross-Encoder) is called one by one for these 30 candidate results to output relevance scores, and finally, the Top 10 external patents and Top 10 internal results are output in descending order of scores.

[0172] During the evaluation phase, the novelty search scheduling service 210 triggers two sub-processes in step S106 in parallel: The novelty assessment sub-process is executed by the novelty quantification assessment module 330: the innovation point extraction module 310 first decomposes the innovation points of the research topic into four core dimensions: research purpose, technical solution, verification method, and application scenario. Then, it extracts the text of the corresponding dimension for each external patent, calls the multi-dimensional similarity calculation module 320 to calculate the similarity of each dimension, takes the maximum value, and obtains the comprehensive novelty score N according to the formula N = 1 - max(S_dim), and determines the novelty level accordingly.

[0173] The duplication analysis module 340 executes the duplication determination sub-process: the topic to be investigated and the top 10 internal results are decomposed into four dimensions: research content, technical route, expected results, and innovation points. The multi-dimensional similarity calculation module 320 is called to calculate the similarity of each dimension. After taking the maximum value, the comprehensive duplication score R is calculated according to the weighted sum formula R = α×S_content + β×S_method + γ×S_output + δ×S_innov, and the duplication type is determined according to the threshold.

[0174] During the decision-making phase, decision recommendation service 230 determines the value of N and R. Figure 5 The four quadrant positions shown are used to select basic suggestion texts from the suggestion text library and generate personalized adjustment prompts by combining the similarity contribution of each dimension. The report generation service 220 then aggregates all intermediate results and final decisions from each module, renders an HTML report and converts it to PDF format, finally stores the report on the file server, and pushes the download link to the novelty report viewing panel 130 of the user interaction layer 100.

[0175] Throughout the entire process, data is transferred between layers in a standardized JSON format. The application service layer 200 records the time consumption and intermediate results of all steps in the log table of the structured database 530, which facilitates subsequent performance analysis and model optimization.

[0176] The following three specific embodiments illustrate the repeatability determination process, novelty assessment process, and four-quadrant decision-making process of the present invention.

[0177] Example 1: Repeatability Determination Process Taking the research topic "Research on key technologies for construction of large-span steel structures in super high-rise buildings" and the existing internal results "Research on construction technology of large-span spatial steel structure buildings" as examples, the process of determining the repeatability of this invention is explained in detail.

[0178] First, the system performs multi-path parallel retrieval. The multi-path recall module simultaneously performs keyword retrieval based on the BM25 algorithm, similarity retrieval based on semantic vectors, and retrieval based on the technical element knowledge graph index. After the three retrieval results are merged and sorted, the top 10 internal results are output. From these, the result most similar to the topic to be searched is selected as the comparison result, namely the result numbered INT-0782, "Research on Construction Technology of Large-Span Spatial Steel Structure Building".

[0179] Next, the system performs dimensional decomposition. The duplication analysis module decomposes the research topic to be investigated and the comparative results into research content dimension, technical route dimension, expected results dimension, and innovation point dimension. Taking the research content dimension text of the research topic to be investigated, "Research on key technologies for construction of large-span steel structures in super high-rise buildings," and the research content dimension text of the comparative results, "Research on construction technology of large-span spatial steel structure buildings," as an example.

[0180] Then, the system calculates the comprehensive similarity S_content across the research content dimension. The multi-dimensional similarity calculation module calculates the following three indicators respectively: TF-IDF cosine similarity S_cosine_content = 0.75; BM25 score S_bm25_content = 0.82; The Jaccard similarity coefficient S_jaccard_content = 0.50.

[0181] Using the default fusion weights (w_cosine = 0.4, w_bm25 = 0.4, w_jaccard = 0.2), the weighted fusion is calculated as follows: S_content = 0.4 × 0.75 + 0.4 × 0.82 + 0.2 × 0.50 = 0.728.

[0182] Similarly, the system calculates the comprehensive similarity across other dimensions: The technical approach dimension S_method = 0.65; Expected output dimension S_output = 0.70; The innovation dimension S_innov = 0.55.

[0183] Next, the system calculates the overall repetition score R. Using the default dimension weights (α = 0.3, β = 0.3, γ = 0.2, δ = 0.2), the score is calculated according to the weighted sum formula: R = 0.3 × 0.728 + 0.3 × 0.65 + 0.2 × 0.70 + 0.2 × 0.55 = 0.2184 + 0.195 + 0.14 + 0.11 = 0.663.

[0184] Finally, the system determines the type of duplication based on preset thresholds. The preset thresholds are T1 = 0.8, T2 = 0.6, and T3 = 0.4. Since R = 0.663 falls between T2 (0.6) and T1 (0.8), the system classifies the duplication type as "core duplication." This determination indicates that the research topic and existing internal findings have a high degree of overlap in core content, requiring close attention.

[0185] Example 2: Novelty Assessment Process The user's submitted project application materials stated that the core innovation was "a method for dynamic optimization of construction progress based on digital twins". The following section uses this as an example to explain in detail the novelty assessment process of this invention.

[0186] First, the system calls the innovation point extraction module and uses a large language model to decompose the innovation point into four core dimensions: Research objective: To achieve real-time dynamic optimization of construction progress; Technical solution (Mechanism): Virtual-real mapping based on digital twin technology; Validation method (Evaluation): Validation through practical engineering applications; Application scenario: Construction progress management of super high-rise buildings.

[0187] Then, the system calculates the comprehensive similarity between the research topic and the top 10 external patents across each dimension, and takes the maximum value as the final similarity for that dimension. The calculation results are as follows: S_purpose = max(0.7, 0.65, 0.72, …) = 0.72; S_mechanism = max(0.5, 0.48, 0.55, …) = 0.55; S_evaluation = max(0.6, 0.58, 0.62, …) = 0.62; S_application = max(0.8, 0.75, 0.78, …) = 0.80.

[0188] Next, the system calculates the novelty score for each dimension. The novelty score is inversely related to the similarity score, and the calculation formula is N_dim = 1 - S_dim: N_purpose = 1 - 0.72 = 0.28; N_mechanism = 1 - 0.55 = 0.45; N_evaluation = 1 - 0.62 = 0.38; N_application = 1 - 0.80 = 0.20.

[0189] Then, the system calculates the overall novelty score N using the maximum value method, with the formula N = 1 - max(S_purpose,S_mechanism, S_evaluation, S_application): N = 1 - max(0.72, 0.55, 0.62, 0.80) = 1 - 0.80 = 0.20.

[0190] Finally, the system determines the novelty level based on the comprehensive novelty score N. The determination rules are as follows: N > 0.8 is "novel", 0.6 < N ≤ 0.8 is "significant improvement", 0.3 < N ≤ 0.6 is "minor improvement", and N ≤ 0.3 is "applied to known technology". Since N = 0.20 ≤ 0.3, the system determines the novelty level to be "applied to known technology".

[0191] Analysis of the judgment criteria: This innovation has a certain degree of novelty in the technical solution dimension (N_mechanism = 0.45), but due to its high similarity to existing external patents in the application scenario dimension (S_application = 0.80), the overall novelty score is lowered to 0.20, which is a conventional application of known technology. This result indicates that simply adjusting the technical solution while using the same application scenario is unlikely to generate substantial innovative value.

[0192] Example 3: Four-Quadrant Decision Process Based on the calculation results of Examples 1 and 2 above, the overall repetition rate of the subject under investigation is R = 0.663, and the overall novelty score is N = 0.20.

[0193] The system maps (R, N) to a four-quadrant decision matrix. For example... Figure 5 As shown, the four-quadrant matrix has the overall repetition R as the horizontal axis and the overall novelty score N as the vertical axis, with the midpoint thresholds of R_mid = 0.5 and N_mid = 0.5, respectively.

[0194] Since R = 0.663 > 0.5 and N = 0.20 < 0.5, the topic falls in the fourth quadrant (not recommended).

[0195] According to the four-quadrant decision-making rule, the decision recommendation for the fourth quadrant is "not recommended to initiate the project." The system outputs a complete decision recommendation: "This project has core duplication with existing internal achievements (R = 0.663) and lacks novelty compared to external patents (N = 0.20). It is considered repetitive research with no innovative value, and initiation of the project is not recommended." Simultaneously, the system generates personalized adjustment suggestions based on the similarity contribution of each dimension: "The application scenario dimension has a similarity of 0.80. It is recommended to expand the application scope from 'super high-rise building construction progress management' to include 'construction collaboration management in complex environments' to enhance differentiation from existing patents." The three embodiments described above fully demonstrate the entire process of this invention, from repetition determination and novelty assessment to four-quadrant decision-making. Through this process, the system can automatically and quantitatively complete the novelty search for research projects and output clear decision-making suggestions, effectively assisting research managers in making project initiation decisions.

[0196] The research topic deduplication auxiliary decision-making method and system presented in this invention deeply integrates large language model and vector retrieval technology and applies it to the research project establishment and novelty search scenario. It constructs a complete technical system including multi-algorithm fusion similarity calculation, Facet-based novelty quantitative evaluation, and multi-level duplication analysis, so as to realize a quantitative answer to the decision question of "whether the topic is worth establishing".

[0197] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A research topic deduplication auxiliary decision-making method integrating large language model and vector retrieval, characterized in that, The method includes: Obtain the text input of the topic to be investigated, preprocess the text of the topic to be investigated and extract structured information to obtain structured topic features; Perform multi-path parallel retrieval, which includes keyword-based retrieval, semantic vector-based retrieval, and knowledge graph-based retrieval. After fusing the retrieval results, obtain an internal result candidate set and an external patent candidate set, and use a large language model to reorder the candidate results. The comprehensive duplication score is calculated based on the internal candidate set of results. The calculation of the comprehensive duplication score includes: decomposing the topic to be investigated and the internal results into multiple preset dimensions, calculating the comprehensive similarity of each dimension, and obtaining the comprehensive similarity score by weighted fusion of the results of multiple text similarity algorithms. Then, the similarity scores of each dimension are weighted and summed to obtain the comprehensive duplication score. The comprehensive novelty score is calculated based on an external patent candidate set. The calculation of the comprehensive novelty score includes: using a large language model to decompose the innovation points of the subject to be investigated into multiple core dimensions, calculating the similarity between each dimension and the external patent candidate set, and then calculating the comprehensive novelty score based on the similarity of each dimension. Based on the overall repeatability score and the overall novelty score, a project initiation decision recommendation is output through a four-quadrant decision matrix.

2. The method according to claim 1, characterized in that, The multiple text similarity algorithms include at least two of the following: word frequency-inverse document frequency cosine similarity, BM25 score, and Jaccard similarity coefficient; the multiple preset dimensions include at least two of the following: research content dimension, technical route dimension, expected results dimension, and innovation point dimension; the multiple core dimensions include at least two of the following: research purpose dimension, technical solution dimension, verification method dimension, and application scenario dimension.

3. The method according to claim 1, characterized in that, The weighted fusion of the comprehensive similarity uses configurable weight parameters, and the formula for calculating the comprehensive repetition is the weighted sum of the similarity of each dimension and the corresponding dimension weight. The dimension weights are configurable according to the focus of the novelty search task.

4. The method according to claim 1, characterized in that, The overall novelty score is calculated using the maximum value method.

5. The method according to claim 1, characterized in that, The method further includes: determining the novelty level based on the comparison result of the comprehensive novelty score and the preset threshold; and / or determining the repetition type based on the comparison result of the comprehensive repetition degree and the preset threshold.

6. The method according to claim 1, characterized in that, The four-quadrant decision matrix uses the overall repetition score as the horizontal axis and the overall novelty score as the vertical axis. It maps the topics to four quadrants according to the comparison results of the two dimensions with their respective intermediate thresholds. Each quadrant corresponds to different project decision recommendations.

7. A research topic deduplication auxiliary decision-making system integrating large language model and vector retrieval, characterized in that, The system includes: The user interaction layer is used to obtain input on the research topic and display the decision results; The application service layer is used to schedule the novelty search process, generate reports, and provide decision-making suggestions. The intelligent analysis layer is used to execute the research topic deduplication auxiliary decision-making method as described in any one of claims 1-6; The search engine layer efficiently retrieves candidate documents related to the research topic from external patent databases and internal results databases, and performs fine sorting on the candidate set. The data storage layer persistently stores all the data required for the system to operate.

8. The system according to claim 7, characterized in that, The user interaction layer includes a topic upload entry, a text paste input box, a novelty search report viewing panel, and a project proposal display area. Furthermore, the user interaction layer presents a four-quadrant decision matrix in a visual manner.

9. The system according to claim 7, characterized in that, The intelligent analysis layer deploys a multi-dimensional similarity calculation module, which encapsulates the weighted fusion calculation logic of TF-IDF cosine similarity, BM25 score and Jaccard similarity coefficient.

10. The system according to claim 9, characterized in that, The intelligent analysis layer is equipped with a novelty quantification evaluation module and a repeatability analysis module. The novelty quantification assessment module is configured to work in conjunction with the multi-dimensional similarity calculation module to calculate the overall novelty score using the maximum value method; The repeatability analysis module is configured to work in conjunction with the multi-dimensional similarity calculation module to calculate the overall repeatability using a weighted sum formula.

Images