Paper title and abstract-based semantic similarity visualization method and system

Abstract

The invention discloses a semantic similarity visualization method and system based on paper titles and abstracts. The semantic similarity visualization method comprises the steps that firstly, an own integrated literature database is constructed, and text embedding vectors and image embedding vectors, matched with the paper titles and abstracts, of all papers are generated; then, a t-SNE nonlinear dimensionality reduction algorithm is used for defining probability distribution of an embedded vector of each paper in a high-dimensional space and mapping the probability distribution to a low-dimensional space to keep a relative distance between the papers; clustering the embedded vector of each thesis after dimension reduction, and extracting a thesis keyword with the highest occurrence frequency in each cluster as a topic tag; then, according to the semantic similarity, a topic label is distributed to each thesis, and the number of times of citation of the thesis is calculated by using PageRank; and finally, performing visual display on the papers based on the clustering result and the citation times. According to the method, the internal relation between the papers is mined and presented in a visual mode by utilizing a visualization technology, so that the academic research efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of paper retrieval and visualization analysis, and specifically relates to a semantic similarity visualization method and system based on paper titles and abstracts. Background Art

[0002] Traditional document management tools are digital academic auxiliary software used to systematically collect, organize, cite and share academic documents. However, due to their lack of deep semantic analysis, they only rely on surface features and cannot deeply capture potential semantic information. When dealing with polysemous words or synonyms, these tools are often unable to accurately distinguish the meanings in different contexts, resulting in limited accuracy of retrieval results. Therefore, errors are prone to occur when dealing with polysemous words or synonyms, resulting in limited accuracy of retrieval results. They also lack contextual information processing and are difficult to adapt to interdisciplinary data. The contextual differences of the same vocabulary in different disciplines can easily lead to information bias. At the same time, the traditional academic document display format is not visually sufficient, mostly in lists or tree structures, lacking dynamic semantic relationship display. It is difficult for users to identify semantic associations of papers in massive data, which greatly limits scientific research efficiency.

[0003] Current academic paper platforms also face numerous challenges. For example, access restrictions prevent widespread access to valuable research results, hindering the rapid dissemination of knowledge and innovation. With the increasing subdivision and intersection of disciplines, existing platforms are fragmented and limited in functionality, making it difficult to meet the demand for comprehensive, personalized academic paper platforms. Platform search algorithms sometimes fail to accurately understand user query intent, resulting in irrelevant results or missing important papers. Furthermore, most platforms fail to provide personalized recommendations based on user research interests and historical reading history, making it difficult for users to find interesting content within the vast literature and reducing the efficiency of discovering valuable papers.

[0004] In the digital age, paper visualization technology is increasingly integrated with fields such as data science, information science, and cognitive science, forming an interdisciplinary research trend. With the advancement of globalization and the increasing frequency of international academic exchange and collaboration, paper visualization, as an effective tool for academic communication, is becoming increasingly important. Therefore, creating a platform based on paper semantics, keyword visualization, and personalized paper recommendations is crucial for improving the efficiency of academic research and promoting research and innovation in various fields. Summary of the Invention

[0005] The main purpose of this invention is to overcome the shortcomings and deficiencies of the existing technology and provide a semantic similarity visualization method and system based on paper titles and abstracts. By introducing advanced text embedding and topic modeling technologies, the shortcomings of the existing technology in deep semantic analysis, interdisciplinary adaptability and dynamic visualization display are solved.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: In one aspect, the present invention provides a method for visualizing semantic similarity based on paper titles and abstracts, comprising the following steps: Collecting original paper data from multiple academic databases, and then cleaning and filling the data to build a self-contained integrated literature database; the original paper data includes title, author, year, cited literature, first image and abstract; The CLIP model is used to generate text embedding vectors and image embedding vectors that match the title and abstract of each paper in the integrated literature database; Use the t-SNE nonlinear dimensionality reduction algorithm to define the probability distribution of the text embedding vector and image embedding vector of each paper in the high-dimensional space, and map them to the low-dimensional space to maintain the relative distance between each paper; Cluster the text embedding vectors and image embedding vectors of each paper after dimensionality reduction, and extract the most frequently appearing paper keywords in each cluster as topic labels; Assign a topic label to each paper based on semantic similarity and use PageRank to calculate the number of citations of the paper; Visualize the papers based on clustering results and citation counts.

[0007] As a preferred technical solution, the data cleaning and filling includes: Use mean imputation, median imputation, mode imputation and interpolation methods to handle missing values ​​in the original paper data; The box plot method and standard deviation method are used to identify and process outliers in the original paper data, and to remove duplicate and erroneous content.

[0008] As a preferred technical solution, the text embedding vector and image embedding vector that match the paper title and abstract are generated as follows: Each paper in the integrated literature database is input into the CLIP model to obtain a semantic description of each paper. Based on the semantic description, the natural language processing tool library NLTK is used to extract semantic keywords, and a text embedding model is used to generate a text embedding vector that matches each paper with the paper title and abstract. Each paper in the integrated literature database is input into the CLIP model to generate an image embedding vector matching each paper with the paper title and abstract.

[0009] As a preferred technical solution, the CLIP model performs contrastive learning training on image-text pairs containing positive and negative samples, and uses the InfoNCE loss function to optimize the model to obtain text embedding vectors and image embedding vectors that match the paper title and abstract; the positive samples are sample pairs in which the image and text match; the negative samples are sample pairs in which the image and text do not match.

[0010] As a preferred technical solution, the natural language processing tool library NLTK divides the semantic description according to specific punctuation marks and specific character patterns to obtain each character; Combine multiple strings according to the adjacent relationship of characters; Use the probability model to calculate the probability of each character in the string forming a word segmentation, and select the word segmentation scheme with the highest probability to determine the most likely word segmentation position in the string; After determining the word segmentation position, stop words are removed and the TF-IDF method is used to select the keywords of each paper; The text embedding model is used to convert keywords into numerical vectors to obtain the text embedding vectors that match each paper with the paper title and abstract.

[0011] As a preferred technical solution, the t-SNE nonlinear dimensionality reduction algorithm is used to define each vector in a high-dimensional space. x i The probability distribution of P ( i | j ), indicating keywords x j In keywords x i The probability in the local neighborhood of is calculated using the Gaussian distribution: , in σ To set the constant; Defining vectors in low-dimensional space x i The corresponding vector y i The conditional probability distribution of Q ( i | j ), which means in the vector y i Select vectors in the local neighborhood of y j Probability of: , Resize vectors in low-dimensional space y i and vector y j, by minimizing the cost function Q ( i | j ) as close as possible P ( i | j ), realizing the vector of high-dimensional solution space x i Vector mapped to low-dimensional space y i ; The vector includes a text embedding vector and an image embedding vector.

[0012] As a preferred technical solution, the most frequently appearing paper keywords in each cluster are extracted as topic tags, specifically: Use the HDBSCAN algorithm to cluster the text embedding vectors and image embedding vectors of each paper after dimensionality reduction to obtain clusters of semantically similar documents; For each cluster, use c-TF-IDF to extract the most frequently appearing paper keywords from the cluster vector; The BERTopic model is used to generate a topic label for each cluster based on the extracted paper keywords.

[0013] As a preferred technical solution, the PageRank is used to calculate the number of citations of the paper, and the calculation formula is: , in, PR ( a ) i+1 For the i +1 paper in circulation a PR value, PR ( T i ) i Indicates the i Other papers in this cycle T i Point to the paper a PR value, T i For the i Other papers in this cycle, L ( T i ) indicates the i Other papers in this cycle T i Point to the paper a The number of outbound links, i.e. the paper a Number of citations; n is the total number of cycles; when i=0, that is, the initial state, the PR value of all papers is set to 1 / N , N The total number of papers.

[0014] As a preferred technical solution, the paper is visualized based on clustering results and citation counts, specifically: Use the visualization library to build an interactive visualization interface, and select the paper with the greatest semantic similarity to the topic tag as the core paper of the topic based on the semantic similarity between the topic tag and each paper; The core papers of each topic are taken as main nodes, and other related papers are taken as branch nodes. The lines between the nodes represent the citation relationships between the papers, thus constructing a paper knowledge graph. The paper knowledge graph is displayed according to the hierarchy of topics and subtopics, and different levels of paper nodes are distinguished by different colors, sizes and shapes; Each paper is assigned a position in the visualization interface, and the position is arranged and distributed based on semantic similarity. That is, the greater the semantic similarity, the more central it is to the topic.

[0015] Another aspect of the present invention provides a semantic similarity visualization system based on paper titles and abstracts, which is applied to the semantic similarity visualization method based on paper titles and abstracts, and includes a data processing module, a vector representation module, a dimensionality reduction mapping module, a topic clustering module, a citation allocation module, and a visualization display module; The data acquisition module is used to collect paper documents from multiple academic databases as original paper data, and perform data cleaning and filling to build its own integrated document database; the original paper data includes title, author, year, cited documents, first image and abstract; The vector representation module is used to generate a text embedding vector and an image embedding vector that matches the title and abstract of each paper in the own integrated literature database using the CLIP model; The dimensionality reduction mapping module is used to define the probability distribution of the text embedding vector and image embedding vector of each paper in the high-dimensional space using the t-SNE nonlinear dimensionality reduction algorithm, and map them to the low-dimensional space to maintain the relative distance between each paper; The topic clustering module is used to cluster the text embedding vector and image embedding vector of each paper after dimensionality reduction, and extract the paper keywords with the highest frequency in each cluster as topic labels; The citation allocation module is used to assign a topic tag to each paper based on semantic similarity and calculate the number of citations of the paper using PageRank; The visualization display module is used to visualize the papers based on the clustering results and the number of citations.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: This paper uses the CLIP model to embed text and images in papers, more accurately capturing the semantic information of the text and reducing errors caused by synonyms and polysemy compared to traditional keyword-matching-based methods. Furthermore, the t-SNE nonlinear dimensionality reduction algorithm is used to map the high-dimensional embedding vectors of the papers to a low-dimensional space while maintaining the relative distances between papers. This effectively preserves the local structure between paper data points, ensuring that similar papers retain similarity even on the reduced-dimensional visualization page, facilitating an intuitive understanding of the relationships between papers. Clustering is then performed to identify richer and more detailed topics, generating more coherent and interpretable topic descriptions compared to a single LDA model. The PageRank algorithm is then introduced to rank papers, more accurately quantifying their academic impact and providing a more comprehensive reflection of their importance than simple citation counts. Finally, a visualization is performed, constructing a paper knowledge graph based on semantic similarity to intuitively display the relationships between papers. Multi-level interactive functionality is also provided, allowing users to easily explore and retrieve papers, from a coarse-grained "paper galaxy" view to a fine-grained individual paper. This improves user experience and analytical efficiency compared to static visualization methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0018] Figure 1 This is an overall flow chart of the semantic similarity visualization method based on paper titles and abstracts in an embodiment of the present invention.

[0019] Figure 2 Schematic diagram of the process of generating an embedding vector in an embodiment of the present invention.

[0020] Figure 3 This is a visual display diagram in an embodiment of the present invention.

[0021] Figure 4 This is a partially enlarged view of the visual display diagram in an embodiment of the present invention.

[0022] Figure 5 Schematic diagram of the structure of the semantic similarity visualization system based on paper titles and abstracts in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0024] References to "embodiments" in this application mean that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described in this application may be combined with other embodiments.

[0025] The purpose of this invention is to provide a paper visualization method to meet the needs of the academic community in processing massive papers. Its core design principle is to integrate a variety of advanced technologies such as clustering and visualization to achieve efficient collection, processing, analysis and visualization of paper information, while providing users with a personalized service experience. This invention uses clustering algorithms to mine the intrinsic connections between papers, uses visualization technology to present complex paper relationships to users in an intuitive way, and uses text embedding vectors to improve semantic understanding and information retrieval capabilities. Figure 1 As shown, this embodiment discloses a semantic similarity visualization method based on paper titles and abstracts, which is characterized by comprising the following steps: S1. Collect massive amounts of paper documents from multiple academic databases as original paper data, and perform data cleaning and filling to build our own integrated document database as the basis for users to search for papers; among them, the original paper data includes title, author, year, cited literature, first image and abstract, etc.

[0026] In this example, to obtain a rich data source, we extensively collected original paper data from multiple authoritative academic databases, covering several databases in different disciplines, and collected 2,000 papers as the basic data. Navicat for MySQL was selected as the database management system for data storage, and the collected academic paper data was imported into it. The MySQL database has excellent stability and scalability, which can meet the needs of large-scale data storage and query. To improve retrieval efficiency, indexes are created for key information fields such as file name, path, and author information. This allows for rapid location and retrieval of the required paper data, reducing query time and improving system response speed.

[0027] Furthermore, the specific method of data cleaning and filling in the present invention is: For missing values ​​in the original paper data, mean filling, median filling, mode filling and interpolation methods are used to handle them; for outliers in the original paper data, box plot method and standard deviation method are used to identify and handle them to ensure the accuracy and completeness of the data; at the same time, low-quality content such as duplication and errors are removed to improve the quality and availability of the data.

[0028] S2. Use the CLIP model to generate text embedding vectors and image embedding vectors that match the title and abstract of each paper in the integrated literature database.

[0029] Furthermore, in order to present the intrinsic meaning and structure of the paper in a more intuitive and quantifiable way and provide a solid foundation for analysis and research, the present invention uses the CLIP model to achieve efficient semantic retrieval, text embedding and image embedding generation, such as Figure 2 As shown, the specific steps are: S2.1. Input each paper in the own integrated literature database into the CLIP model to obtain the semantic description of each paper.

[0030] Furthermore, the CLIP model handles image encoding, text encoding, and image-based text encoding and decoding tasks through its multimodal hybrid architecture. In this embodiment, the CLIP model performs contrastive learning training on image-text pairs containing positive and negative samples, and uses the InfoNCE loss function to optimize the model. This loss function encourages the model to bring the embedding vectors of positive samples closer while pushing the embedding vectors of negative samples further apart. Through contrastive learning, semantic associations between images and texts are achieved, resulting in text embedding vectors and image embedding vectors that match the title and abstract of the paper. In addition, data enhancement is performed on images and texts before training to increase the robustness and generalization ability of the model. Positive samples refer to sample pairs where images and texts match, and negative samples refer to sample pairs where images and texts do not match.

[0031] S2.2. Based on the semantic description, use the natural language processing tool library NLTK to extract semantic keywords, and use the text embedding model to generate the text embedding vector of each paper.

[0032] Furthermore, the Natural Language Processing (NLTK) library is used to process the semantic description of papers, classify paper text, analyze language structure, and perform other operations, including extracting paper titles, author information, abstracts, and normalizing this information. The NLTK natural language processing tool library mainly performs word segmentation based on rule-based and statistical methods, specifically: First, the semantic description is divided according to specific punctuation marks (such as space, comma, period, etc.) and specific character patterns to obtain each character; Combine multiple strings according to the adjacent relationship of characters; Then, the probability model is used to calculate the probability of each character in the string forming a word. The word segmentation scheme with the highest probability is selected to determine the most likely word segmentation position in the string. If a string has a high probability of appearing as an independent word, then it is more likely to be segmented at this position. The probability calculation formula for adjacent characters forming a word segment is: , in, p ( s ) is the probability of adjacent characters forming a word, s = w 1 w 2… w l A word is composed of adjacent characters. w i For the i characters, p ( w i | w 1… w i-1 ) is given before i- Under the condition of 1 character, i characters w i The probability of occurrence, l for s length, p ( w i | w i-1 ) is the character before the given w i-1 Under the conditions of i characters w i Probability of occurrence.

[0033] After determining the word segmentation position, stop words are removed and the TF-IDF method is used to select the keywords of each paper. The specific method of selecting keywords in this embodiment is: First calculate each word t In the semantic description d The frequency of occurrence in TF ( t , d ); then calculate the inverse of the frequency of each word in the entire semantic description, that is, the inverse document frequency IDF ( t ), which is used to reduce the weight of common words and calculate the TF-IDF value of each word by combining word frequency and inverse document frequency: TF-IDF ( t , d )= TF ( t , d )× IDF ( t ), and finally the keywords of each paper are selected according to the highest TF-IDF value.

[0034] The text embedding model is used to convert keywords into numerical vectors to obtain the text embedding vectors that match each paper with the paper title and abstract.

[0035] In this embodiment, the text embedding model uses Text-embedding-ada-002, and the preprocessed keyword text T , through pre-trained model f (text-embedding-ada-002), the text T Converted into a d -dimensional embedding vector v T =f ( T )∈ R d Text-embedding-ada-002 is a text embedding service provided by OpenAI. It converts text into embedding vectors in a high-dimensional space. These embedding vectors capture the semantic information of the text and play an important role in natural language processing tasks. When calculating the similarity between papers, using these embedding vectors for semantic similarity calculations can improve the accuracy of similarity calculations.

[0036] S2.3. Input each paper in the own integrated literature database into the CLIP model to generate the image embedding vector of each paper.

[0037] S3. Use the t-SNE nonlinear dimensionality reduction algorithm to define the probability distribution of the text embedding vector and image embedding vector of each paper in the high-dimensional space, and map them to the low-dimensional space to maintain the relative distance between each paper.

[0038] Furthermore, the t-SNE nonlinear dimensionality reduction algorithm defines the probability distribution between data points in high-dimensional space and maps it to low-dimensional space while maintaining the relative distance between data points in high-dimensional and low-dimensional space. This effectively preserves the local structure between paper data points, allowing similar papers to maintain similarity on the visualized page after dimensionality reduction, making it easier to intuitively understand the relationship between papers. Specifically: S3.1. Define each vector in high-dimensional space x i The probability distribution of P ( i | j ), which represents a vector xj In vector x i The probability in the local neighborhood of is realized by Gaussian distribution, where x i and x j The similarity between them is determined by their Euclidean distance; the probability distribution needs to make the sum of the probabilities in the local neighborhood of each point a constant. The smaller the distance, the closer the semantics: , in, σ To set the constant, the numerator represents the vector in high-dimensional space. x j Pair Vector x i The conditional probability density of x i and x j The denominator represents the number of x i Other vectors x k The probability density sum is used for normalization to ensure that all x j for x i The sum of the conditional probabilities is 1.

[0039] Defining vectors in low-dimensional space x i The corresponding vector y i The conditional probability distribution of Q ( i | j ), indicating that y i Select from the local neighborhood y j Probability of: , Among them, the molecule is represented in a low-dimensional space, and the vector y j Pair Vector y i The conditional probability density of , the denominator is for all except y i Points outside y k The probability density sum of .

[0040] Finally, by adjusting the vector in the low-dimensional space y i andy j , by minimizing the cost function Q ( i | j ) as close as possible P ( i | j ), realizing the vector of high-dimensional solution space x i Vector mapped to low-dimensional space y i The cost function is: , Where N is the total number of vectors. The vectors include text embedding vectors and image embedding vectors.

[0041] S4. Cluster the text embedding vector and image embedding vector of each paper after dimensionality reduction, and extract the paper keywords with the highest frequency in each cluster as topic labels.

[0042] Furthermore, the clustering process of the present invention is specifically as follows: S4.1. Use the HDBSCAN algorithm to cluster the text embedding vector and image embedding vector of each paper after dimensionality reduction to obtain clusters of semantically similar documents.

[0043] S4.2. For each cluster, use c-TF-IDF to extract the most frequently appearing paper keywords from the cluster vector.

[0044] S4.3. Use the BERTopic model to generate a topic label for each cluster based on the extracted paper keywords.

[0045] S5. Assign a topic tag to each paper based on semantic similarity to ensure that papers on the same topic have high semantic consistency, and use PageRank to calculate the number of citations of the paper.

[0046] Furthermore, the present invention also combines the PageRank algorithm to construct a multi-level visual layout. The most important papers will be displayed first. As the web page is gradually enlarged, other relatively less important papers will gradually emerge based on the citation network. The application of this technology helps users quickly grasp key research trends, which not only ensures the highlighting of core content, but also gradually guides users to explore a wider range of academic resources in depth to meet research needs at different levels. Specifically, PageRank is used to calculate the number of citations of a paper as its importance. The calculation formula is: , in, PR ( a ) i+1 For thei +1 paper in circulation a PR value, PR ( T i ) i Indicates the i Other papers in this cycle T i Point to the paper a PR value, T i For the i Other papers in this cycle, L ( T i ) indicates the i Other papers in this cycle T i Point to the paper a The number of outbound links, i.e. the paper a Number of citations; n is the total number of cycles; when i =0, that is, the initial state, the PR value of all papers is set to 1 / N , N The total number of papers.

[0047] S6. Visualize the papers based on clustering results and citation counts.

[0048] Furthermore, the present invention uses visualization technology to intuitively present the citation network and similarity information between papers: S6.1. Use the visualization library to build an interactive visualization interface, and select the paper with the greatest semantic similarity as the core paper of the topic based on the semantic similarity between the topic tag and each paper.

[0049] S6.2. Take the core papers of each topic (the most influential papers in each topic) as the main nodes, and other related papers as branch nodes. The lines between the nodes represent the citation relationships between the papers, and construct a paper knowledge graph.

[0050] S6.3. Display the paper knowledge graph according to the hierarchy of topics and subtopics, and use different colors, sizes, and shapes to distinguish paper nodes at different levels.

[0051] S6.4. Each paper is assigned a position in the visualization interface, and the positions are arranged and distributed based on semantic similarity. That is, the greater the semantic similarity, the more central it is to the topic.

[0052] Figure 3The initial visualization interface is displayed, which contains the core paper nodes of each topic, and these nodes are distinguished by different colors according to the topic they belong to. When the user zooms in on the visualization interface, Figure 4 The interface shown in Figure 1 details all academic papers in a specific research field. In this view, paper nodes with high semantic similarity are spatially close to each other, intuitively reflecting the semantic relevance between papers.

[0053] It should be noted that, for the sake of convenience, the aforementioned method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present invention is not limited to the described order of actions, because according to the present invention, certain steps can be performed in other orders or simultaneously.

[0054] Based on the same concept as the semantic similarity visualization method based on paper titles and abstracts in the above-mentioned embodiment, the present invention also provides a semantic similarity visualization system based on paper titles and abstracts, which can be used to execute the above-mentioned semantic similarity visualization method based on paper titles and abstracts. For ease of explanation, the structural diagram of the embodiment of the semantic similarity visualization system based on paper titles and abstracts only shows the parts related to the embodiment of the present invention. Those skilled in the art will understand that the illustrated structure does not constitute a limitation of the device, and it can include more or fewer components than shown in the diagram, or combine certain components, or arrange the components differently.

[0055] like Figure 5 As shown, another embodiment of the present invention provides a semantic similarity visualization system based on paper titles and abstracts, including a data processing module, a vector representation module, a dimensionality reduction mapping module, a topic clustering module, a citation allocation module, and a visualization display module; The data acquisition module is used to collect original paper data from multiple academic databases, and to clean and fill the data to build its own integrated literature database; the original paper data includes title, author, year, cited literature, first image and abstract; The vector representation module is used to generate text embedding vectors and image embedding vectors that match the title and abstract of each paper in the integrated literature database using the CLIP model; The dimensionality reduction mapping module is used to define the probability distribution of the text embedding vector and image embedding vector of each paper in the high-dimensional space using the t-SNE nonlinear dimensionality reduction algorithm, and map them to the low-dimensional space to maintain the relative distance between each paper; The topic clustering module is used to cluster the text embedding vectors and image embedding vectors of each paper after dimensionality reduction, and extract the most frequently appearing paper keywords in each cluster as topic labels; The citation allocation module is used to assign a topic label to each paper based on semantic similarity and calculate the number of citations of the paper using PageRank; The visualization display module is used to visualize the papers based on clustering results and citation counts.

[0056] It should be noted that the semantic similarity visualization system based on paper titles and abstracts of the present invention corresponds one-to-one to the semantic similarity visualization method based on paper titles and abstracts of the present invention. The technical features and beneficial effects described in the above-mentioned embodiment of the semantic similarity visualization method based on paper titles and abstracts are applicable to the embodiment of the semantic similarity visualization system based on paper titles and abstracts. For specific contents, please refer to the description in the embodiment of the method of the present invention. No further details will be given here. This is hereby declared.

[0057] In addition, in the implementation of the semantic similarity visualization system based on paper titles and abstracts in the above embodiment, the logical division of each program module is only an example. In actual applications, the above functions can be assigned to different program modules as needed, for example, for the configuration requirements of the corresponding hardware or the convenience of software implementation. That is, the internal structure of the semantic similarity visualization system based on paper titles and abstracts is divided into different program modules to complete all or part of the functions described above.

[0058] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0059] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A semantic similarity visualization method based on paper titles and abstracts, characterized by: The steps include: Collecting original paper data from multiple academic databases, and then cleaning and filling the data to build a self-contained integrated literature database; the original paper data includes title, author, year, cited literature, first image and abstract; The CLIP model is used to generate text embedding vectors and image embedding vectors that match the title and abstract of each paper in the integrated literature database; Use the t-SNE nonlinear dimensionality reduction algorithm to define the probability distribution of the text embedding vector and image embedding vector of each paper in the high-dimensional space, and map them to the low-dimensional space to maintain the relative distance between each paper; Cluster the text embedding vectors and image embedding vectors of each paper after dimensionality reduction, and extract the most frequently appearing paper keywords in each cluster as topic labels; Assign a topic label to each paper based on semantic similarity and use PageRank to calculate the number of citations of the paper; Visualize the papers based on clustering results and citation counts.

2. The semantic similarity visualization method based on paper titles and abstracts according to claim 1 is characterized in that: The data cleaning and filling include: Use mean imputation, median imputation, mode imputation and interpolation methods to handle missing values ​​in the original paper data; The box plot method and standard deviation method are used to identify and process outliers in the original paper data, and to remove duplicate and erroneous content.

3. The semantic similarity visualization method based on paper titles and abstracts according to claim 1 is characterized in that: The text embedding vector and image embedding vector that match the paper title and abstract are generated as follows: Each paper in the integrated literature database is input into the CLIP model to obtain a semantic description of each paper. Based on the semantic description, the natural language processing tool library NLTK is used to extract semantic keywords, and a text embedding model is used to generate a text embedding vector that matches each paper with the paper title and abstract. Each paper in the integrated literature database is input into the CLIP model to generate an image embedding vector matching each paper with the paper title and abstract.

4. The semantic similarity visualization method based on paper titles and abstracts according to claim 3 is characterized in that: The CLIP model is trained through contrastive learning of image-text pairs containing positive and negative samples, and the model is optimized using the InfoNCE loss function to obtain text embedding vectors and image embedding vectors that match the paper title and abstract. The positive samples are sample pairs in which the image and text match; the negative samples are sample pairs in which the image and text do not match.

5. The semantic similarity visualization method based on paper titles and abstracts according to claim 3 is characterized in that: The natural language processing tool library NLTK divides the semantic description according to specific punctuation marks and specific character patterns to obtain each character; Combine multiple strings according to the adjacent relationship of characters; Use the probability model to calculate the probability of each character in the string forming a word segmentation, and select the word segmentation scheme with the highest probability to determine the most likely word segmentation position in the string; After determining the word segmentation position, stop words are removed and the TF-IDF method is used to select the keywords of each paper; The text embedding model is used to convert keywords into numerical vectors to obtain the text embedding vectors that match each paper with the paper title and abstract.

6. The semantic similarity visualization method based on paper titles and abstracts according to claim 1 is characterized in that: The t-SNE nonlinear dimensionality reduction algorithm is used to define each vector in a high-dimensional space x i The probability distribution of P ( i | j ), indicating keywords x j In keywords x i The probability in the local neighborhood of is calculated using the Gaussian distribution: ， in σ To set the constant; Defining vectors in low-dimensional space x i The corresponding vector y i The conditional probability distribution of Q ( i | j ), which means in the vector y i Select vectors in the local neighborhood of y j Probability of: ， Resize vectors in low-dimensional space y i and vector y j , by minimizing the cost function Q ( i | j ) as close as possible P ( i | j ), realizing the vector of high-dimensional solution space x i Vector mapped to low-dimensional space y i ; The vector includes a text embedding vector and an image embedding vector.

7. The semantic similarity visualization method based on paper titles and abstracts according to claim 1 is characterized in that: The most frequently appearing keywords in each cluster are extracted as topic tags, specifically: Use the HDBSCAN algorithm to cluster the text embedding vectors and image embedding vectors of each paper after dimensionality reduction to obtain clusters of semantically similar documents; For each cluster, use c-TF-IDF to extract the most frequently appearing paper keywords from the cluster vector; The BERTopic model is used to generate a topic label for each cluster based on the extracted paper keywords.

8. The semantic similarity visualization method based on paper titles and abstracts according to claim 1 is characterized in that: The PageRank method is used to calculate the number of citations of papers. The calculation formula is: ， in, PR ( a ) i+1 For the i +1 paper in circulation a PR value, PR ( T i ) i Indicates the i Other papers in this cycle T i Point to the paper a PR value, T i For the i Other papers in this cycle, L ( T i ) indicates the i Other papers in this cycle T i Point to the paper a The number of outbound links, i.e. the paper a Number of citations; n is the total number of cycles; when i =0, that is, the initial state, the PR value of all papers is set to 1 / N , N The total number of papers.

9. The semantic similarity visualization method based on paper titles and abstracts according to claim 1 is characterized in that: The paper is visualized based on clustering results and citation counts, specifically: Use the visualization library to build an interactive visualization interface, and select the paper with the greatest semantic similarity to the topic tag as the core paper of the topic based on the semantic similarity between the topic tag and each paper; The core papers of each topic are taken as main nodes, and other related papers are taken as branch nodes. The lines between the nodes represent the citation relationships between the papers, thus constructing a paper knowledge graph. The paper knowledge graph is displayed according to the hierarchy of topics and subtopics, and different levels of paper nodes are distinguished by different colors, sizes and shapes; Each paper is assigned a position in the visualization interface, and the position is arranged and distributed based on semantic similarity. That is, the greater the semantic similarity, the more central it is to the topic.

10. A semantic similarity visualization system based on paper titles and abstracts, characterized by: The semantic similarity visualization method based on paper titles and abstracts as described in any one of claims 1 to 9 comprises a data processing module, a vector representation module, a dimensionality reduction mapping module, a topic clustering module, a citation allocation module, and a visualization display module; The data acquisition module is used to collect paper documents from multiple academic databases as original paper data, and perform data cleaning and filling to build its own integrated document database; the original paper data includes title, author, year, cited documents, first image and abstract; The vector representation module is used to generate a text embedding vector and an image embedding vector that matches the title and abstract of each paper in the own integrated literature database using the CLIP model; The dimensionality reduction mapping module is used to define the probability distribution of the text embedding vector and image embedding vector of each paper in the high-dimensional space using the t-SNE nonlinear dimensionality reduction algorithm, and map them to the low-dimensional space to maintain the relative distance between each paper; The topic clustering module is used to cluster the text embedding vector and image embedding vector of each paper after dimensionality reduction, and extract the paper keywords with the highest frequency in each cluster as topic labels; The citation allocation module is used to assign a topic tag to each paper based on semantic similarity and calculate the number of citations of the paper using PageRank; The visualization display module is used to visualize the papers based on the clustering results and the number of citations.

Images