Scientific research subject similarity examination method, device, equipment and medium

Abstract

The invention discloses a similarity examination method, device and equipment for scientific research subjects and a medium. The method belongs to the technical field of natural language processing and comprises the steps of extracting at least two key indexes of a to-be-examined scientific research subject; determining at least two text vectors for each key index by using at least two text vectorization tools; for each key index, calculating each text similarity value with the historical scientific research subject; the text similarity values are fused to determine index similarity; performing weighted fusion on the similarity values of the indexes of the to-be-examined scientific research subject, and determining the subject initial similarity of the to-be-examined scientific research subject and the historical scientific research subject; and when the project initial similarity exceeds a preset threshold value, outputting a similarity result of the to-be-examined scientific research project and the historical scientific research project by using the large language model. And the similarity between the subjects is determined based on the multi-granularity text vector, so that the accuracy of similarity review of the scientific research subjects is improved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method, apparatus, equipment, and medium for similarity review of scientific research projects. Background Technology

[0002] Text similarity review of research projects is a crucial step in the research management of enterprises and research institutions. It can prevent duplicate projects and is of great significance for promoting the high-quality development of science and technology, optimizing the allocation of scientific and technological resources, and improving the quality and efficiency of research projects.

[0003] Current mainstream text similarity assessment methods mostly rely on string edit distance or term frequency statistics based on TF-IDF (Term Frequency-Inverse Document Frequency), lacking effective modeling of deep semantic features. Furthermore, natural language exhibits significant polysemy, synonymy, and semantic drift, meaning the same research intent may be expressed through different terms, and seemingly similar wording may point to entirely different research questions. Therefore, relying solely on literal matching based on term frequency statistics is insufficient to accurately characterize the substantive connections between research topics, resulting in inaccurate text similarity assessment results. Summary of the Invention

[0004] The purpose of this invention is to provide a method, apparatus, device, and medium for reviewing the similarity of scientific research topics. By using various text vectorization tools, semantic features of text are obtained from multiple granularities, and the similarity between scientific research topics is determined based on the generated multiple text vectors, thus solving the problem of inaccurate text similarity review results for existing scientific research topics.

[0005] This invention is achieved through the following technical solution:

[0006] The first aspect of this application provides a method for reviewing the similarity of research topics, including:

[0007] Extract key textual information from the research project materials to be reviewed, and extract at least two key indicators from the key textual information; the key textual information is textual descriptive information that can characterize the key information of the project.

[0008] At least two text vectorization tools are used to vectorize the textual description information of key indicators, and at least two text vectors are determined for each key indicator; the at least two text vectorization tools generate corresponding text vectors by extracting text features at different granularities.

[0009] For each of the key indicators, calculate the text similarity value between each text vector of the key indicator and each text vector corresponding to the historical research topic.

[0010] The text similarity values ​​are fused together to determine the similarity between the key indicators of the research project to be reviewed and the corresponding key indicators of historical research projects.

[0011] The similarity values ​​of each indicator of the research topic to be reviewed are weighted and fused to determine the initial similarity between the research topic to be reviewed and the historical research topic.

[0012] When the initial similarity of the research topic exceeds a preset threshold, the key text information of the research topic to be reviewed and the key text information of the historical research topics are input into a fine-tuned large language model to obtain the similarity result between the research topic to be reviewed and the historical research topics output by the large language model.

[0013] In one feasible implementation, the text similarity values ​​are fused to determine the similarity between the key indicators of the research topic under review and the corresponding key indicators of historical research topics, including:

[0014] The text similarity values ​​are input into a pre-trained attention network to obtain the weights of each text similarity value output by the attention network.

[0015] Based on the weights of the text similarity values, the text similarity values ​​are weighted and fused to determine the similarity between the key indicators of the research topic to be reviewed and the corresponding key indicators of historical research topics.

[0016] In one feasible implementation, training the attention network includes:

[0017] Construct a training sample set; the training sample set includes research topic samples and binary similarity values ​​of different research topic pairs labeled;

[0018] During each iteration of training, the text similarity values ​​corresponding to the two research topics are input into the attention network to be trained, and the weight values ​​output by the attention network are obtained.

[0019] Based on the output weight values, the similarity between the two research topics is determined.

[0020] Based on the similarity between the stated similarity and the labeled similarity value, the value of the loss function is determined. The parameters of the attention network are adjusted according to the principle of minimizing the value of the loss function until the attention network converges.

[0021] In one feasible implementation, the method further includes:

[0022] Obtain expert review opinions on the similarity results output by the large language model;

[0023] The review opinions are used for fine-tuning the large language model and the attention network.

[0024] In one feasible implementation, the text vectorization tool uses three types: a text vectorization tool that represents the statistical features of the text, a text vectorization tool that represents the semantic features at the word level, and a text vectorization tool that represents the contextual features at the sentence level.

[0025] In one feasible implementation, the method further includes:

[0026] The extracted key text information is segmented into words, and keywords are extracted from the segmented key text information. The text information after keyword extraction is used as the text description information of key indicators.

[0027] In one feasible implementation, the input to the large language model also includes a prompt template, which includes the titles, research objectives, research content, technical routes, and innovative points of the two research projects.

[0028] The prompt template also includes the output information of the large language model, which includes the similarity results between the research topic to be reviewed and the historical research topics, as well as the structured review report of the similarity review of the research topic to be reviewed, and determines the necessary elements that must be included in the review report.

[0029] The second aspect of this application provides a similarity review device for scientific research topics, comprising:

[0030] The key indicator extraction unit is used to extract key text information from the research project materials to be reviewed, and to extract at least two key indicators from the key text information; the key text information is textual description information that can characterize the key information of the project.

[0031] The multi-granularity text vector generation unit uses at least two text vectorization tools to vectorize the textual description information of key indicators, and determines at least two text vectors for each key indicator; the at least two text vectorization tools generate corresponding text vectors by extracting text features at different granularities.

[0032] The text similarity determination unit calculates the text similarity value between each text vector of the key indicator and each text vector corresponding to the historical research topic for each key indicator.

[0033] The indicator similarity determination unit is used to fuse the text similarity values ​​to determine the indicator similarity between the key indicators of the research topic to be reviewed and the corresponding key indicators of historical research topics.

[0034] The initial similarity determination unit is used to perform weighted fusion of the similarity values ​​of various indicators of the research topic to be reviewed, and to determine the initial similarity between the research topic to be reviewed and the historical research topic.

[0035] The similarity output unit is used to input the key text information of the research topic to be reviewed and the key text information of the historical research topics into a fine-tuned large language model when the initial similarity of the topic exceeds a preset threshold, so as to obtain the similarity result between the research topic to be reviewed and the historical research topics output by the large language model.

[0036] A third aspect of this application provides an electronic device, including: a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the above-described method.

[0037] A fourth aspect of this application provides a storage medium, comprising: storing a program or instructions on the storage medium, wherein the program or instructions, when executed by a processor, implement the steps of the above-described method.

[0038] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0039] This application uses multiple text vectorization tools to generate multiple text vectors for each key indicator. Since multiple text vectorization tools extract text features based on different granularities, multi-granular semantic features can be obtained, solving the misjudgment caused by single-granularity feature extraction when there are differences in terminology or substantive meaning, and improving the accuracy of the review results. In addition, this embodiment also performs similarity comparison for each indicator separately, further refining the comparison granularity. For the research topics to be reviewed whose initial similarity exceeds a preset threshold, semantic analysis is also performed through a large language model to determine the similarity results between the research topic to be reviewed and historical research topics, further improving the accuracy of the final similarity results. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0041] Figure 1 A flowchart illustrating a method for reviewing the similarity of scientific research topics provided in this application embodiment;

[0042] Figure 2A schematic diagram of the structure of a similarity review device for scientific research projects provided in this application embodiment;

[0043] Figure 3 This is a schematic diagram of the structure of a computing device provided in an embodiment of this application. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments and accompanying drawings. The illustrative embodiments and descriptions of this invention are for explanation only and are not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments in this application without creative effort are within the scope of protection of this application.

[0045] As will be known to those skilled in the art, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0046] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not expressly listed or inherent to such process, method, product, or apparatus.

[0047] Example 1:

[0048] Embodiment 1 of this application provides a method for reviewing the similarity of scientific research topics, in order to solve the problem of inaccurate text similarity review results in existing scientific research topics.

[0049] The subject executing this method can be any computing device capable of implementing the method, such as a server, mobile phone, personal computer, smart wearable device, smart robot, etc.

[0050] Furthermore, the embodiments of this application do not limit the execution order of different steps. When using the method provided in the embodiments of this application, the execution order of different steps can be adjusted according to actual needs.

[0051] For ease of description, the following uses a similarity review device for scientific research projects as the subject of this method to provide a detailed description of the method provided in this application embodiment.

[0052] like Figure 1 The diagram shown is a flowchart illustrating the specific implementation of a similarity review method for scientific research topics provided in this application, including the following steps 11-16:

[0053] Step 11: Extract key text information from the research project materials to be reviewed, and extract at least two key indicators from the key text information.

[0054] Research materials come in various formats, including Word documents, PDF (Portable Document Format) files, and scanned copies of paper materials. These heterogeneous materials need to be converted into computer-readable text information. This can be achieved using technologies such as OCR (Optical Character Recognition) and large-scale visual models for scanned documents and PDF text. Furthermore, key text information needs to be extracted from the converted text, such as the main body of the research materials, and key indicators like titles, research content, research objectives, and innovative points, as well as the corresponding textual descriptions.

[0055] The key text information refers to textual descriptions that can characterize key information of the research topic; such as the textual body of research materials.

[0056] In one feasible implementation, this embodiment will also perform structured processing on the extracted key text information, key indicators and other information, and store them in a preset database. The key information of a scientific research project will occupy a storage ID (Identification, unique identifier) ​​in the database.

[0057] To further improve the efficiency and accuracy of subsequent processing, this embodiment also includes: performing word segmentation on the extracted key text information, and extracting keywords from the segmented key text information, using the keyword-extracted text information as text description information for key indicators. This stage mainly includes basic processing steps such as word segmentation, sentence segmentation, and paragraph division, aiming to transform the original text into a standardized input form with clear structure and coherent semantics.

[0058] This process involves using text segmentation tools (such as jieba) to segment the original text, and then filtering out irrelevant words using a stop word list to reduce noise interference. Building on this, keyword extraction techniques (such as TF-IDF, YAKE! (YetAnother Keyword Extractor), or TextRank) are used to select more domain-representative terms for constructing a concise vocabulary. This strategy not only significantly reduces the dimensionality of subsequent TF-IDF vectors, effectively alleviating the high-dimensional sparsity problem, but also improves the semantic focus of the features.

[0059] It should be noted that the aforementioned database already stores key information about historical research topics, as well as text information after text segmentation and keyword extraction. Furthermore, this embodiment also includes: constructing and persisting a TF-IDF vocabulary and corresponding vectorized encoders (such as TfidfVectorizer) based on the historical research topics in the database. Simultaneously, a word embedding model (such as Word2Vec or FastText) needs to be trained using this corpus, and the trained word vector model needs to be saved to support subsequent distributed semantic similarity calculation tasks.

[0060] Sentence segmentation can be based on regular expressions, using punctuation marks such as periods, question marks, exclamation marks, and ellipses in both Chinese and English to divide the text into semantically complete sentence units.

[0061] Paragraph segmentation requires comprehensive consideration of text structural features (such as heading levels and natural paragraph separators), sentence semantic integrity, and the maximum input length limit of the selected embedding model. Taking bge-large-zh-v1.5 as an example, its maximum context length is 512 tokens. Therefore, during segmentation, it should be ensured that each paragraph ends with a complete sentence and the total length does not exceed this limit. This avoids semantic breaks or encoding anomalies caused by truncation, maintaining the semantic coherence of paragraphs while ensuring model compatibility.

[0062] Step 12: Use at least two text vectorization tools to vectorize the text description information of the key indicators, and determine at least two text vectors for each key indicator.

[0063] The at least two text vectorization tools generate corresponding text vectors by extracting text features at different granularities.

[0064] In one feasible implementation, this embodiment selects three text vectorization tools, namely:

[0065] Text vectorization tools that represent statistical features of text, such as TF-IDF; text vectorization tools that represent word-level semantic features, such as Word2Vec; and text vectorization tools that represent sentence-level contextual features, such as SentenceTransformer.

[0066] Among them, TF-IDF: Based on term frequency-inverse document frequency statistics, it maps text into sparse vectors in the vocabulary space. Its dimension is equal to the size of the global vocabulary and is usually high. It mainly captures surface co-occurrence features at the vocabulary level.

[0067] Word2Vec: First, it obtains the word vector of each keyword (usually 100-dimensional or 200-dimensional), and then aggregates them into a dense vector of the entire text through mean pooling, reflecting shallow distributed semantic information.

[0068] Sentence-BERT: It takes the entire text (such as research content or innovation points) as input and generates context-aware sentence-level dense embeddings. Its vector dimension is determined by the pre-trained model (usually 768 or 1024 dimensions), which can effectively model deep semantic structures.

[0069] This embodiment employs a strategy of using multiple text vector representations, taking into account statistical features, word-level semantics, and sentence-level context, providing a heterogeneous but complementary vector foundation for subsequent multi-source similarity fusion. Simultaneously, the vector representations corresponding to the key textual information of the research topic are persistently stored in the topic information database to support efficient subsequent similarity calculation and retrieval.

[0070] Step 13: For each of the key indicators, calculate the text similarity value between each text vector of the key indicator and each text vector corresponding to the historical research topic.

[0071] For example, for the key indicator of innovation, the corresponding text vectors are the statistical feature vector of innovation, the word-level semantic feature vector of innovation, and the sentence-level contextual feature vector of innovation. Similarly, the key indicator of innovation in historical research projects also has these three text vectors. Similarity calculations are performed on the statistical feature vectors of the innovation points of the two research projects to obtain the lexical similarity value; similarity calculations are performed on the word-level semantic feature vectors to obtain the word vector-level similarity value; and similarity calculations are performed on the sentence-level contextual feature vectors to obtain the context-aware sentence-level similarity value. Finally, the text similarity values ​​(lexical similarity value, word vector-level similarity value, and context-aware sentence-level similarity value) are obtained.

[0072] This embodiment uses the cosine similarity algorithm, and the formula for calculating text similarity can be expressed as:

[0073] ;

[0074] In the formula, A and B represent the text vectors of the research project to be reviewed and the historical research project under the same key indicator, respectively. , represents the component values ​​of dimension i of the text vector. Vectors A and B can be derived from statistical feature vectors of TF-IDF, word-level semantic feature vectors of Word2Vec, or sentence-level contextual feature vectors of Sentence-BERT, and must be vectors from the same text vectorization tool.

[0075] It should be noted that the range of cosine similarity is [−1,1], where 1 represents completely in the same direction (highly similar), 0 represents orthogonal (no correlation), and -1 represents opposite direction (extremely dissimilar).

[0076] To facilitate subsequent analysis and fusion, the results are usually normalized and mapped to the [0,1] interval, for example, through linear transformation:

[0077] .

[0078] Ultimately, standardized similarity scores for the two research projects under each key indicator can be obtained, which can be used to support comprehensive evaluation and similarity ranking.

[0079] Step 14: Merge the text similarity values ​​to determine the similarity between the key indicators of the research topic to be reviewed and the corresponding key indicators of historical research topics.

[0080] To integrate the text similarity values, an attention-based dynamic weighted fusion mechanism is adopted. The three similarity values ​​calculated in step 13, namely the lexical similarity value based on TF-IDF, the word vector similarity value based on Word2Vec, and the context-aware sentence similarity value based on Sentence-BERT, are used to generate the final fused similarity score of the key indicator (such as "innovation point").

[0081] The weights for fusing the text similarity values ​​are determined using an attention network. Specifically, this involves: inputting each text similarity value into a pre-trained attention network to obtain the weights output by the attention network for each text similarity value; and then, based on these weights, performing a weighted fusion of the text similarity values ​​to determine the similarity between the key indicators of the research topic under review and the corresponding key indicators of historical research topics.

[0082] The training of this attention network includes the following steps 1-4:

[0083] Step 1: Construct a training sample set; the training sample set includes research topic samples and binary similarity values ​​of different research topic pairs labeled with these samples. The binary similarity values ​​can be binary annotations performed by domain experts, forming a high-quality dataset. This dataset accurately reflects the criteria for judging duplicate proposals in research management and serves as the basis for training and evaluating the attention network.

[0084] The binary labeling is as follows: the similarity of several pairs of research topics in the research topic sample is labeled as 1 or 0; where 1 indicates substantial similarity and 0 indicates dissimilarity.

[0085] Step 2: During each iteration of training, input the text similarity values ​​corresponding to the two research topics into the attention network to be trained, and obtain the weight values ​​output by the attention network.

[0086] Constructing the input feature vector for the attention network: For a key metric, construct the input feature vector based on its three text similarity values. ,in, Represents word-level similarity values. Represents word vector-level similarity values. This represents the context-aware sentence-level similarity value.

[0087] The attention network to be trained can be a lightweight attention network (such as a single fully connected layer + softmax), with output normalized weights: [α,β,γ]=Softmax(Wx+b);

[0088] In the formula, , The parameters are learnable and satisfy α+β+γ=1, where the values ​​of α, β, and γ are all non-negative.

[0089] Step 3: Determine the similarity between the two research topics based on the output weight values.

[0090] The similarity score of the merged indicators is: Furthermore, the similarity between the two research topics is mapped through the sigmoid activation function: p = σ(s).

[0091] Step 4: Based on the similarity values ​​between the stated similarity and the labeled similarity values, determine the value of the loss function. Adjust the parameters of the attention network according to the principle of minimizing the value of the loss function until the attention network converges.

[0092] The attention network parameters are optimized using manually labeled y∈{0,1} and the cross-entropy loss function:

[0093] ;

[0094] By using backpropagation to jointly update W and b, the fusion strategy directly serves the final discrimination objective. This mechanism tightly integrates expert knowledge (manual annotation) with data-driven learning, achieving an upgrade from fixed fusion (where the weights corresponding to the similarity values ​​of each text are fixed) to task-adaptive fusion, significantly improving the accuracy and generalization ability of repetitive review of scientific research projects.

[0095] Step 15: Perform weighted fusion of the similarity values ​​of each indicator of the research topic to be reviewed to determine the initial similarity between the research topic to be reviewed and the historical research topic.

[0096] After determining the similarity of each key indicator, the initial similarity of the research projects is calculated by weighted fusion of the similarity of each indicator. The weighting in this step is reasonably set based on practical experience in project review and opinions from field experts. For example, project titles usually have limited information and contribute relatively little to the overall similarity, so they can be given a lower weight; while core elements such as research content or technical routes better reflect the essence of the project and should be given a higher weight.

[0097] Assuming there are n key metrics, the similarity of the i-th key metric is... The corresponding weight is ,satisfy ≥0 and The initial similarity between the research topic to be reviewed and the historical research topics can be expressed as:

[0098] ;

[0099] Initial similarity of topics The values ​​remain within the [0,1] range, facilitating subsequent threshold judgment or ranking comparison. The weight values ​​can be dynamically optimized through expert scoring, Analytic Hierarchy Process (AHP), or learning methods based on historical review data to improve the accuracy and adaptability of similarity assessment.

[0100] Step 16: When the initial similarity of the research topic exceeds a preset threshold, input the key text information of the research topic to be reviewed and the key text information of the historical research topics into the fine-tuned large language model to obtain the similarity result between the research topic to be reviewed and the historical research topics output by the large language model.

[0101] The preset threshold can be determined based on the performance analysis results of the previously manually labeled validation set, such as setting the preset threshold to 0.5.

[0102] Alternatively, the Top-K strategy can be used to select research topic pairs for input into the large language model. That is, the initial similarity of each research topic to be reviewed and each historical research topic is arranged in descending order of similarity value, and the top K pairs with the highest similarity value are selected.

[0103] The research topic pairs whose initial similarity exceeds a preset threshold, as well as the research topic pairs corresponding to the top K positions with high initial similarity values, are further judged using a large language model.

[0104] For research topic pairs that are not screened for inclusion in the large language model, they are considered to have no similarity risk. This stage achieves efficient filtering, reducing the number of research topic pairs requiring in-depth analysis to 1%-5% of their original size, significantly reducing the subsequent computational load.

[0105] The large language model used in this embodiment can be a large language model finely tuned for the scientific research field, such as DeepSeek, Qwen, GLM (General Language Model), or a privately deployed scientific research-specific large language model.

[0106] When performing similarity judgments in a large language model, the following are included:

[0107] Input a pre-constructed standardized prompt template into the large language model. The prompt content includes key indicators such as the titles, research objectives, research content, technical routes, and innovative points of the two research projects.

[0108] Clearly define the output of the large language model:

[0109] 1. Output similarity results Fine-grained similarity scores for ∈[0,1] reflect whether there is substantial duplication;

[0110] 2. Generate a structured review report, which must include the following essential elements:

[0111] The report should include the following: similarity assessment conclusions (e.g., "highly overlapping," "partially related," "no substantial duplication"), analysis of key overlap points (e.g., "both focus on Transformer-based medical image segmentation"), explanations of differences (e.g., "Project A focuses on CT (Computed Tomography) images, while Project B focuses on MRI (Magnetic Resonance Imaging), with different application scenarios"), duplication risk level (high / medium / low), and recommended measures (e.g., "manual review recommended," "can be combined for submission," "no conflict"). A structured review report should be output in JSON format for easy system display and archiving.

[0112] In one feasible implementation, this embodiment also includes storing research project pairs, project A-IDs and project B-IDs within the research project pairs, initial project similarity, and similarity results in a structured format to ensure the review process is traceable and reusable. After data is stored in the database, it supports functions such as real-time access by the expert review platform, audit log generation, model performance retrospective analysis, and querying of similar project historical records.

[0113] In one feasible implementation, this embodiment further includes: obtaining expert review opinions on the similarity results output by the large language model; using the review opinions for fine-tuning the large language model and the attention network.

[0114] Fine-tuning of a large language model can involve: analyzing deviations in semantic understanding, logical reasoning, or domain judgment (such as misjudging "similar methods are repetitive"), specifically enhancing the definition constraints, example guidance, or output format requirements in the prompts, and verifying the effectiveness of the new templates through testing.

[0115] Fine-tuning the attention network can be achieved by constructing a supervision signal using verification labels (similar / dissimilar), and then fine-tuning the attention network (i.e., the lightweight module that generates [α, β, γ]). For example, when experts frequently correct "high..." Value + Low When a case is identified as "value = dissimilar", the model automatically learns a weight allocation strategy that reduces α and increases γ.

[0116] This joint optimization mechanism not only enhances the deep reasoning capabilities of large models but also strengthens the discrimination sensitivity of multi-granularity fusion in the early stages, forming an end-to-end adaptive evolution capability from front-end features to back-end decisions, thereby continuously improving the overall review accuracy.

[0117] This application uses multiple text vectorization tools to generate multiple text vectors for each key indicator. Since multiple text vectorization tools extract text features based on different granularities, multi-granular semantic features can be obtained, solving the misjudgment caused by single-granularity feature extraction when there are differences in terminology or substantive meaning, and improving the accuracy of the review results. In addition, this embodiment also performs similarity comparison for each indicator separately, further refining the comparison granularity. For the research topics to be reviewed whose initial similarity exceeds a preset threshold, semantic analysis is also performed through a large language model to determine the similarity results between the research topic to be reviewed and historical research topics, further improving the accuracy of the final similarity results.

[0118] Example 2:

[0119] To address the problem of inaccurate text similarity review results for existing scientific research projects, and based on the same inventive concept as Embodiment 1, this application also provides a similarity review fault device for scientific research projects.

[0120] A schematic diagram of the specific structure of the device is shown below. Figure 2 As shown, it includes the following functional units 21-26:

[0121] The key indicator extraction unit 21 is used to extract key text information from the research project materials to be reviewed, and to extract at least two key indicators from the key text information; the key text information is textual description information that can characterize the key information of the project.

[0122] Key indicators include title, research content, research objectives, and innovative points.

[0123] The multi-granularity text vector generation unit 22 uses at least two text vectorization tools to vectorize the text description information of key indicators, and determines at least two text vectors for each key indicator; the at least two text vectorization tools generate corresponding text vectors by extracting text features at different granularities.

[0124] There are three types of text vectorization tools: those that represent the statistical features of text, those that represent the semantic features at the word level, and those that represent the contextual features at the sentence level.

[0125] The text similarity determination unit 23 calculates the text similarity value between each text vector of the key indicator and each text vector corresponding to the historical research topic for each key indicator.

[0126] The indicator similarity determination unit 24 is used to fuse the text similarity values ​​to determine the indicator similarity between the key indicators of the research topic to be reviewed and the corresponding key indicators of historical research topics.

[0127] The indicator similarity determination unit is specifically used to: input the text similarity values ​​into a pre-trained attention network to obtain the weights of the text similarity values ​​output by the attention network; and perform weighted fusion of the text similarity values ​​based on the weights of the text similarity values ​​to determine the indicator similarity between the key indicators of the research topic to be reviewed and the corresponding key indicators of historical research topics.

[0128] Training an attention network includes:

[0129] A training sample set is constructed, which includes research topic samples and binary similarity values ​​of different research topic pairs labeled with annotations. During each training iteration, the text similarity values ​​corresponding to the two research topics are input into the attention network to be trained, and the weight values ​​output by the attention network are obtained. Based on the weight values ​​output, the similarity between the two research topics is determined. Based on the similarity between the two research topics and the labeled similarity values, the value of the loss function is determined. The parameters of the attention network are adjusted according to the principle of minimizing the value of the loss function until the attention network converges.

[0130] The initial similarity determination unit 25 is used to perform weighted fusion of the similarity values ​​of various indicators of the research topic to be reviewed, and to determine the initial similarity between the research topic to be reviewed and the historical research topic.

[0131] The similarity output unit 26 is used to input the key text information of the research topic to be reviewed and the key text information of the historical research topics into a fine-tuned large language model when the initial similarity of the topic exceeds a preset threshold, so as to obtain the similarity result between the research topic to be reviewed and the historical research topics output by the large language model.

[0132] The examination device in this embodiment is also used for:

[0133] Obtain expert review opinions on the similarity results output by the large language model; use the review opinions for fine-tuning the large language model and the attention network.

[0134] The review device in this embodiment further includes a word segmentation unit, which is used to: perform word segmentation processing on the extracted key text information, and extract keywords from the key text information after word segmentation processing, and use the text information after keyword extraction as text description information of key indicators.

[0135] This application uses multiple text vectorization tools to generate multiple text vectors for each key indicator. Since multiple text vectorization tools extract text features based on different granularities, multi-granular semantic features can be obtained, solving the misjudgment caused by single-granularity feature extraction when there are differences in terminology or substantive meaning, and improving the accuracy of the review results. In addition, this embodiment also performs similarity comparison for each indicator separately, further refining the comparison granularity. For the research topics to be reviewed whose initial similarity exceeds a preset threshold, semantic analysis is also performed through a large language model to determine the similarity results between the research topic to be reviewed and historical research topics, further improving the accuracy of the final similarity results.

[0136] Based on the same inventive concept as the foregoing embodiments of this application, this application also provides a computing device.

[0137] like Figure 3 As shown, the computing device includes a memory 31 and a processor 32. The memory 31 can be configured to store various other data to support operation on the electronic device. Examples of such data include instructions for any application or method used to operate on the electronic device. The memory 31 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0138] The processor 32, coupled to the memory 31, is used to execute the computer program stored in the memory 31 to perform a similarity review method for a scientific research project as described in the foregoing embodiments.

[0139] When the processor 32 executes the computer program to perform a method for reviewing the similarity of scientific research topics, multiple text vectors are generated for each key indicator by using various text vectorization tools. Since the multiple text vectorization tools extract text features based on different granularities, multi-granular semantic features can be obtained, solving the misjudgment caused by single-granularity feature extraction when there are differences in terminology or substantive meaning, and improving the accuracy of the review results. In addition, this embodiment also performs similarity comparison for each indicator separately, further refining the comparison granularity. For topics whose initial similarity exceeds a preset threshold, semantic analysis is also performed using a large language model to determine the similarity results between the research topic to be reviewed and historical research topics, further improving the accuracy of the final similarity results.

[0140] When the processor 32 executes the computer program in the memory 31, in addition to the functions described above, it can also perform other functions, as detailed in the descriptions of the preceding embodiments.

[0141] Furthermore, such as Figure 3 As shown, the computing device also includes other components such as a display 34, a communication component 33, a power supply component 35, and an audio component 36. Figure 3 The diagram only shows some components and does not mean that the computing device includes only these components. Figure 3 The components shown.

[0142] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a computer, can implement the methods provided in the above embodiments.

[0143] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0144] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments.

[0145] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for reviewing the similarity of scientific research projects, characterized in that, include: Extract key textual information from the research project materials to be reviewed, and extract at least two key indicators from the key textual information; the key textual information is textual descriptive information that can characterize the key information of the project. At least two text vectorization tools are used to vectorize the textual description information of key indicators, and at least two text vectors are determined for each key indicator; the at least two text vectorization tools generate corresponding text vectors by extracting text features at different granularities. For each of the key indicators, calculate the text similarity value between each text vector of the key indicator and each text vector corresponding to the historical research topic. The text similarity values ​​are fused together to determine the similarity between the key indicators of the research project to be reviewed and the corresponding key indicators of historical research projects. The similarity values ​​of each indicator of the research topic to be reviewed are weighted and fused to determine the initial similarity between the research topic to be reviewed and the historical research topic. When the initial similarity of the research topic exceeds a preset threshold, the key text information of the research topic to be reviewed and the key text information of the historical research topics are input into a fine-tuned large language model to obtain the similarity result between the research topic to be reviewed and the historical research topics output by the large language model.

2. The method according to claim 1, characterized in that, The text similarity values ​​are fused to determine the similarity between the key indicators of the research project under review and the corresponding key indicators of historical research projects, including: The text similarity values ​​are input into a pre-trained attention network to obtain the weights of each text similarity value output by the attention network. Based on the weights of the text similarity values, the text similarity values ​​are weighted and fused to determine the similarity between the key indicators of the research topic to be reviewed and the corresponding key indicators of historical research topics.

3. The method according to claim 2, characterized in that, The training of the attention network includes: Construct a training sample set; the training sample set includes research topic samples and binary similarity values ​​of different research topic pairs labeled; During each iteration of training, the text similarity values ​​corresponding to the two research topics are input into the attention network to be trained, and the weight values ​​output by the attention network are obtained. Based on the output weight values, the similarity between the two research topics is determined. Based on the similarity between the stated similarity and the labeled similarity value, the value of the loss function is determined. The parameters of the attention network are adjusted according to the principle of minimizing the value of the loss function until the attention network converges.

4. The method according to claim 2, characterized in that, The method further includes: Obtain expert review opinions on the similarity results output by the large language model; The review opinions are used for fine-tuning the large language model and the attention network.

5. The method according to claim 1, characterized in that, The text vectorization tool uses three types: a text vectorization tool that represents the statistical features of the text, a text vectorization tool that represents the semantic features at the word level, and a text vectorization tool that represents the contextual features at the sentence level.

6. The method according to claim 1, characterized in that, The method further includes: The extracted key text information is segmented into words, and keywords are extracted from the segmented key text information. The text information after keyword extraction is used as the text description information of key indicators.

7. The method according to claim 1, characterized in that, The input to the large language model also includes a prompt template, which includes the titles, research objectives, research content, technical routes, and innovative points of the two research projects; The prompt template also includes the output information of the large language model, which includes the similarity results between the research topic to be reviewed and the historical research topics, as well as the structured review report of the similarity review of the research topic to be reviewed, and determines the necessary elements that must be included in the review report.

8. A similarity review device for scientific research projects, characterized in that, include: The key indicator extraction unit is used to extract key text information from the research project materials to be reviewed, and to extract at least two key indicators from the key text information; the key text information is textual description information that can characterize the key information of the project. The multi-granularity text vector generation unit uses at least two text vectorization tools to vectorize the textual description information of key indicators, and determines at least two text vectors for each key indicator; the at least two text vectorization tools generate corresponding text vectors by extracting text features at different granularities. The text similarity determination unit calculates the text similarity value between each text vector of the key indicator and each text vector corresponding to the historical research topic for each key indicator. The indicator similarity determination unit is used to fuse the text similarity values ​​to determine the indicator similarity between the key indicators of the research topic to be reviewed and the corresponding key indicators of historical research topics. The initial similarity determination unit is used to perform weighted fusion of the similarity values ​​of various indicators of the research topic to be reviewed, and to determine the initial similarity between the research topic to be reviewed and the historical research topic. The similarity output unit is used to input the key text information of the research topic to be reviewed and the key text information of the historical research topics into a fine-tuned large language model when the initial similarity of the topic exceeds a preset threshold, so as to obtain the similarity result between the research topic to be reviewed and the historical research topics output by the large language model.

9. An electronic device, characterized in that, include: A processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method as claimed in any one of claims 1-7.

10. A storage medium, characterized in that, include: The storage medium stores a program or instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1-7.

Images