Scientific research theme evolution method based on text embedding vector clustering

Abstract

The invention discloses a scientific research theme evolution method based on text embedding vector clustering, and relates to a data processing technology, the method comprises the following steps: collecting scientific research literature data, preprocessing, identifying and extracting field keywords, generating text embedding vectors corresponding to literatures, and constructing a scientific research literature representation vector library; taking a preset time slice as a unit, performing density adaptive clustering on the text embedding vector, generating a scientific research subject cluster in the corresponding time slice, and dynamically extracting Top-N subject keywords of the scientific research subject cluster based on keyword weights and semantic vector center distances of literatures in the cluster; and constructing a dual-verification mechanism of vector space similarity and keyword semantic overlap ratio, dynamically screening a topic evolution relationship between scientific research topic clusters in adjacent time slices, introducing a keyword overlap ratio to filter associated noise, constructing a high-confidence topic evolution chain, and revealing development venation and technical migration rules in the field of scientific research.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a research topic evolution method based on text embedding vector clustering. Background Technology

[0002] The dynamic evolution of research topics not only reflects the trajectory of academic development but also reveals technological migration trends, providing significant guidance for research decision-making and forward-looking innovation analysis. However, with the increasing volume of research literature data and the rapid development of interdisciplinary research, research topics exhibit highly diverse and complex cross-temporal relationships. Traditional topic evolution analysis methods have significant shortcomings when dealing with such large-scale, multi-source literature data: First, traditional methods typically rely on static topic models, which cannot adapt to the differences in the rate of topic change across different fields and struggle to capture the dynamic evolution process of topics. Second, existing topic evolution methods, when dealing with cross-temporal topic associations, usually rely solely on keyword matching or simple vector similarity calculations, making them susceptible to noise interference and leading to the incorrect association of semantically unrelated topics, thus affecting the accuracy of topic evolution analysis. Third, traditional methods often fail to clearly depict the semantic evolution path of topics, especially when facing interdisciplinary research topics, making it difficult to accurately present the semantic evolution path of topics and meet the practical needs of high-precision evolution analysis. Summary of the Invention

[0003] The purpose of this invention is to provide a research topic evolution method based on text embedding vector clustering, which aims to solve the pain points of traditional topic evolution analysis, such as weak static modeling ability, high cross-time topic association noise, and ambiguous semantic evolution path.

[0004] The technical solution to achieve the objective of this invention is as follows: On the one hand, this invention provides a research topic evolution method based on text embedding vector clustering, including the following steps: Collect scientific literature data, preprocess the scientific literature data, identify and extract domain keywords, generate corresponding text embedding vectors, and construct a scientific literature representation vector library. Using a preset time slice as a unit, density adaptive clustering is performed on the text embedding vector to generate scientific research topic clusters within the corresponding time slice. Based on the keyword weights and semantic vector central distances of the documents within the cluster, the Top-N topic keywords of the scientific research topic clusters are dynamically extracted. A dual verification mechanism of vector space similarity and keyword semantic overlap is constructed to dynamically filter the topic evolution relationship between scientific research topic clusters in adjacent time slices, introduce keyword overlap ratio to filter associated noise, and construct a high-confidence topic evolution chain based on the filtered topic evolution relationship.

[0005] Specifically, the scientific research literature data includes various document types such as academic journal articles, conference papers, and patent documents; the scientific research literature data is collected, and preprocessing operations are performed on the collected scientific research literature data, including text cleaning, format standardization, noise filtering, and timestamp standardization, to generate standardized literature data; Furthermore, the standardized document data is subjected to automatic language discrimination. Based on the discrimination results, the corresponding word segmentation tool is selected to segment the text content of the document data, remove punctuation marks and stop words, and extract the metadata information of the document to obtain the segmented document data.

[0006] Furthermore, for the segmented document data, domain keywords are identified and extracted, including the following steps: Construct a research domain dictionary for the target field, which includes core and technical terms in the target field; The dictionary matching method is used to initially screen the text content of the documents, retaining words from the domain dictionary as the initial candidate keyword set, and filtering out general words and non-domain words. Calculate the word frequency (TF-IDF) of each word in the initial candidate keyword set within the text of each document, and its inverse document frequency (IF) across all documents to obtain the TF-IDF weight of each initial candidate keyword. Sort the keywords in descending order according to their TF-IDF weights and then filter them. One set of words as secondary keywords; A word co-occurrence network is constructed based on the secondary keyword set. Words are used as nodes and co-occurrence frequency is used as edge weights. The TextRank algorithm is used to calculate the importance score of each word. The calculation is iterative until convergence. The TF-IDF weights and TextRanK importance scores of each word in the secondary keyword set are weighted and fused according to a preset ratio to obtain a comprehensive score. The keywords with the highest comprehensive scores are then selected. The vocabulary is used as the domain keywords for each document, among which .

[0007] Furthermore, based on the language type of the documents, an appropriate pre-trained text semantic model is selected to semantically encode the text content of the documents, transforming the text content into a low-dimensional semantic vector representation. Specifically, the text is segmented, with long texts being split into multiple segments, each input into a pre-trained semantic model for processing to extract the basic semantic vector of the document text. Simultaneously, the domain keywords extracted from each document are processed by text embedding to obtain keyword enhancement vectors. The basic semantic vector of the document is then concatenated with the keyword enhancement vectors to obtain the text embedding vector corresponding to the document. Based on the text embedding vectors corresponding to all document data, a scientific literature representation vector library is constructed. Preferably, the pre-trained text semantic model is the BERT model, which is pre-trained on a massive general corpus and has strong semantic understanding capabilities. At the same time, a scientific literature database is constructed to fine-tune the BERT model so that it can better understand the professional terms and expressions in scientific literature.

[0008] Furthermore, based on the standard timestamps of the literature data, the literature data is divided into time dimensions, and the text embedding vectors in the scientific literature representation vector library are divided into multiple consecutive preset time slices. Density adaptive clustering is performed on the text embedding vectors in each time slice to generate scientific research topic clusters. Then, the top-N topic keywords of the scientific research topic clusters are dynamically extracted by combining the keyword TF-IDF weights and semantic vector central distances of the literature within the clusters, so as to enhance the semantic interpretability of the scientific research topic clusters. Preferably, the granularity of time slices is dynamically determined based on the rate of change of the subject matter in the target research field, so as to achieve precise splitting of the time dimension; Specifically, the DBSCAN algorithm with adaptive neighborhood radius is used to perform density-adaptive clustering of text embedding vectors within each preset time slice to generate research topic clusters, including the following steps: For each time slice Calculate the semantic distance between all text embedding vectors, specifically: for any two text embedding vectors within the time slice... and Calculate the cosine similarity and then calculate the corresponding semantic distance based on the cosine similarity. Based on the semantic distance, a local semantic distance matrix is ​​constructed. ,in, ; For each text embedding vector According to the local semantic distance matrix Calculate the distance distribution between its k-nearest neighbors, and dynamically calculate based on the distance distribution. adaptive radius ; Based on the aforementioned adaptive neighborhood radius, the DBSCAN algorithm is used to cluster the text embedding vectors within the time slice. For any text embedding vector... According to its domain radius and the local semantic distance matrix Semantic distance less than or equal to the domain radius All text embedding vectors are grouped into the domain set. If the domain set The number of samples is greater than or equal to the preset minimum number of samples. Then Mark each core point as a core point; for each core point, group the text embedding vectors within its domain radius into the same research cluster until the domain cannot be expanded further. If a text embedding vector does not belong to any cluster and is not labeled as a core point, it is marked as a noise point and removed during time slicing. After clustering all text embedding vectors in the dataset, we obtain... Research topic clusters .

[0009] Specifically, semantic enhancement is performed on each research topic cluster. The TF-IDF weights and semantic vector central distances of the keywords in the documents corresponding to the text embedding vectors contained in each research topic cluster are analyzed. The Top-N topic keywords of the research topic cluster are dynamically extracted, including the following steps: For all documents contained within the current research topic cluster, obtain the domain keyword set corresponding to each document, and remove duplicate words to obtain the cluster-level keyword set; Based on the intra-cluster literature data, the word frequency of each keyword in each document in the cluster-level keyword set is counted, and the document frequency of each keyword is counted based on the occurrence of all documents. Its inverse document frequency is calculated, and the TF-IDF weight of each keyword is calculated. The text embedding vectors of all documents within the topic cluster are normalized, and the mean of all text embedding vectors within the cluster is taken to calculate the semantic center vector of the topic cluster, which serves as the core representation of the overall semantics within the cluster. A pre-trained keyword semantic model is used to semantically encode each keyword to obtain the corresponding word vector. The semantic vector center distance between the word vector and the semantic center vector of the cluster is calculated. The higher the similarity, the stronger the correlation between the keyword and the core semantics of the topic cluster. For each keyword, its TF-IDF weight and the semantic central distance of its corresponding word vector are weighted and fused according to a preset ratio to obtain the semantic contribution of each keyword; Sort all keywords in descending order of semantic contribution, based on the preset number of keywords. Select the highest score from the sorted results. These words were selected as the Top-N keywords for this research topic cluster.

[0010] Furthermore, a dual verification mechanism of vector space similarity and keyword semantic overlap is constructed. Combined with dynamic threshold screening and keyword overlap ratio filtering, a high-confidence topic evolution chain is constructed to accurately identify evolutionary events such as the continuation, splitting, extinction and rebirth of scientific research topics, and realize interactive visualization of topic evolution paths. Specifically, the similarity of the vector space of research topic clusters in adjacent time slices and the semantic overlap of keywords are calculated, and then filtered using a dynamic threshold. This includes the following steps: Using continuous time slices as units, all research topic clusters included in the previous time period are paired with all research topic clusters included in the next time period to form a candidate set of related topic clusters to be verified, ensuring that no potential evolutionary relationships between adjacent time periods are missed. For each pair of research topic clusters in the candidate set of related topic clusters, extract the topic clusters from the previous time period. The semantic center vector and the topic cluster of the next period The semantic center vectors are used to calculate the vector space similarity between the two using the cosine similarity algorithm, which directly reflects the semantic association strength between topic cluster pairs. Extracting thematic clusters from the previous time period Top-N topic keywords and subsequent topic clusters The top-N topic keywords are used to construct two sets of topic keywords, and the Jaccard similarity coefficient algorithm is used to calculate the ratio of the number of elements in the intersection to the number of elements in the union of the two topic keyword sets, thus obtaining the semantic overlap of the keywords. ; The vector space similarity and key semantic overlap of each pair of topic clusters in the candidate set of related topic clusters are compared with preset dynamic thresholds. Only when both indicators meet the preset dynamic threshold requirements are the topic clusters included in the evolutionary candidate set, thus initially screening out the evolutionary candidate set with semantic relevance. Specifically, to further improve the accuracy of evolutionary relationships, keyword overlap ratio is introduced to filter cross-time period association noise, including the following steps: For the topic cluster pairs in the evolutionary candidate set, the sum of the TF-IDF weights of the overlapping topic keywords in the two topic clusters and the sum of the TF-IDF weights of the topic keywords in each of the two topic clusters are calculated. The ratio of the sum of the weights of the overlapping topic keywords to the sum of the total weights is defined as the keyword overlap ratio, highlighting the influence of high-weight core keywords on evolutionary association. A preset overlap ratio filtering threshold is set. If the keyword overlap ratio of a topic cluster pair is lower than the preset overlap ratio filtering threshold, its evolutionary relationship is judged as a false association and removed, and finally a set of evolutionary relationships with high confidence is obtained.

[0011] Specifically, evolutionary events are identified based on a set of high-confidence evolutionary relationships. These events are categorized into four types: theme continuation, theme splitting, theme extinction, and theme emergence. The rules for determining these evolutionary events are as follows: If a single topic cluster in the previous period has a unique high-confidence association with a single topic cluster in the next period, and the core semantics are highly consistent, then it is determined to be a topic continuation; If a single topic cluster in the previous period has a high confidence correlation with multiple topic clusters in the next period, and the core semantics of the topic clusters in the next period are all subsets of the topic clusters in the previous period, then it is determined to be a topic split. If a topic cluster in the previous period does not form a high-confidence association with any topic cluster in the next period, and there are no derived associated objects, it is determined that the topic has disappeared. If a topic cluster in a later time period does not form a high-confidence association with any topic cluster in the previous time period, and it represents a newly added semantic direction within the domain, then it is determined to be a newly emerging topic.

[0012] Specifically, a topic evolution network is constructed with time slices as the horizontal axis, topic clusters as nodes, and high-confidence evolutionary relationships as directed edges. The entire process of the birth, continuation, splitting, or extinction of each topic cluster is sorted out in chronological order and connected to form a complete high-confidence topic evolution chain. Each topic cluster node includes the topic keywords of the topic cluster and is associated with metadata such as the collection of documents it contains.

[0013] Optionally, multi-dimensional visual evolution path diagrams or dynamic cloud maps can be generated based on the theme evolution network to intuitively show the evolution process of the theme and dynamically demonstrate the technology transfer patterns in the scientific research field. Furthermore, feedback from domain experts and academic trend data can be combined to further optimize the display effect of the theme evolution path and improve the interpretability and operability of the theme evolution process.

[0014] Compared with the prior art, the significant advantages of this invention are: 1. Using preset time slices as units, density-adaptive clustering is performed on the text embedding vectors of scientific research literature. Combined with semantic enhancement methods, the changes in scientific research topics are dynamically captured, and the semantic interpretability of scientific research topics is enhanced, accurately reflecting the continuous evolution of the scientific research field. 2. The innovative integration of vector space similarity and keyword semantic overlap dual verification mechanism, along with the introduction of keyword overlap ratio, can effectively filter associated noise data, construct a high-confidence evolution chain, accurately identify theme continuation, splitting, extinction and new events, and ultimately achieve interactive visualization of theme evolution path, revealing the development context and technology transfer laws in the scientific research field. Attached Figure Description

[0015] Figure 1 A flowchart of a research topic evolution method based on text embedding vector clustering; Figure 2 This is a flowchart of the process for identifying and extracting domain keywords in this invention; Figure 3 This is a flowchart of density adaptive clustering within a preset time slice in this invention; Figure 4 This is a flowchart for constructing a high-confidence topic evolution chain in this invention. Detailed Implementation

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0017] like Figure 1 As shown, this invention discloses a research topic evolution method based on text embedding vector clustering, including the following steps: Collect scientific literature data, preprocess the scientific literature data, identify and extract domain keywords, generate corresponding text embedding vectors, and construct a scientific literature representation vector library. Using a preset time slice as a unit, density adaptive clustering is performed on the text embedding vector to generate scientific research topic clusters within the corresponding time slice. Based on the keyword weights and semantic vector central distances of the documents within the cluster, the Top-N topic keywords of the scientific research topic clusters are dynamically extracted. A dual verification mechanism of vector space similarity and keyword semantic overlap is constructed to dynamically filter the topic evolution relationship between scientific research topic clusters in adjacent time slices, introduce keyword overlap ratio to filter associated noise, and construct a high-confidence topic evolution chain based on the filtered topic evolution relationship.

[0018] In this embodiment, the scientific literature data includes various document types such as academic journal articles, conference papers, and patent documents; scientific literature data is collected, and preprocessing operations are performed on the collected scientific literature data. The preprocessing includes text cleaning, format standardization, noise filtering, and timestamp standardization to generate standardized literature data. Furthermore, the standardized document data is subjected to automatic language discrimination. Based on the discrimination results, the corresponding word segmentation tool is selected to process the text content of the document data into words, remove punctuation marks and stop words, and extract the metadata information of the document to obtain the word-segmented document data. As an optional implementation method, jieba segmentation is used for Chinese document data.

[0019] like Figure 2 As shown, for the segmented literature data, identifying and extracting domain keywords includes the following steps: Construct a research domain dictionary for the target field, which includes core and technical terms of the target field. The domain dictionary is supported for regular updates as the target research field develops. The dictionary matching method is used to initially screen the text content of the documents, retaining words from the domain dictionary as the initial candidate keyword set, and filtering out general words and non-domain words. Calculate the word frequency (TF-IDF) of each word in the initial candidate keyword set within the text of each document, and its inverse document frequency (IF) across all documents to obtain the TF-IDF weight of each initial candidate keyword. Sort the keywords in descending order according to their TF-IDF weights and then filter them. One set of words as secondary keywords; A word co-occurrence network is constructed based on the secondary keyword set. Words are used as nodes and co-occurrence frequency is used as edge weights. The TextRank algorithm is used to calculate the importance score of each word. The calculation is iterative until convergence. The TF-IDF weights and TextRanK importance scores of each word in the secondary keyword set are weighted and fused according to a preset ratio to obtain a comprehensive score. The keywords with the highest comprehensive scores are then selected. The vocabulary is used as the domain keywords for each document, among which .

[0020] Furthermore, based on the language type of the documents, an appropriate pre-trained text semantic model is selected to semantically encode the text content of the documents, transforming the text content into a low-dimensional semantic vector representation. Specifically, the text content is segmented, with long texts being split into multiple segments, which are then input into the text semantic model for processing to extract the basic semantic vector of the document text. At the same time, the domain keywords extracted from each document are processed by text embedding to obtain keyword enhancement vectors. The basic semantic vector of the document is then concatenated with the keyword enhancement vector to obtain the text embedding vector corresponding to the document. Based on the text embedding vectors corresponding to all document data, a scientific literature representation vector library is constructed. Preferably, the pre-trained text semantic model is the BERT model, which is pre-trained on a massive general corpus and has strong semantic understanding capabilities. At the same time, a scientific literature database is constructed to fine-tune the BERT model so that it can better understand the professional terminology and expressions in scientific literature.

[0021] In the analysis of the evolution of research literature topics, traditional static modeling methods cannot capture the changes in topics across time periods, especially when dealing with a large number of interdisciplinary literature books. Based on this, this embodiment constructs a temporal topic clustering and semantic enhancement method to capture the dynamic evolution of research topics across time periods. First, based on the standard timestamp of the literature data, the literature data is divided into time dimensions, and the text embedding vectors in the research literature representation vector library are divided into multiple consecutive preset time slices. Density adaptive clustering is performed on the text embedding vectors in each time slice to generate research topic clusters. Then, the top-N topic keywords of the research topic clusters are dynamically extracted by combining the keyword TF-IDF weights and semantic vector central distance of the literature within the clusters to enhance the semantic interpretability of the research topic clusters. Preferably, the granularity of time slices is dynamically determined based on the rate of change of the subject matter in the target research field, so as to achieve precise segmentation of the time dimension. As an optional implementation method, the frequency of subject matter change of scientific literature in the target field in the past 5 years is statistically analyzed, and the annual subject matter change rate is calculated. If the annual subject matter change rate exceeds 30%, the quarterly or semi-annual granularity is used; otherwise, the annual granularity is used. like Figure 3 As shown, the DBSCAN algorithm with adaptive neighborhood radius is used to perform density-adaptive clustering of text embedding vectors within each preset time slice to generate research topic clusters, including the following steps: For each time slice Let the set of all text embedding vectors within a time slice be denoted as Calculate the semantic distance between all text embedding vectors, specifically for the set Any two text embedding vectors within and Calculate the cosine similarity and then calculate the corresponding semantic distance based on the cosine similarity. The calculation formula is as follows: , Based on the calculated semantic distance between text embedding vector pairs, a local semantic distance matrix is ​​constructed. , The smaller the semantic distance between vector pairs, the more similar their meanings are. For sets Each text embedding vector in According to the local semantic distance matrix Calculate the distance distribution between its k-nearest neighbors, and dynamically calculate the text embedding vector based on the distance distribution. adaptive radius The calculation formula is: , In the formula It is an adjustment coefficient, with a value range of [0,1]. Density clustering is performed using the DBSCAN algorithm with adaptive radius for each text embedding vector. According to its domain radius and the local semantic distance matrix Semantic distance less than or equal to the domain radius The entire set of text embedding vectors constitutes the domain set. If the domain set The number of samples within is greater than or equal to the preset minimum number of samples. Then Mark each core point; for each core point, group all text embedding vectors within its domain radius into the same research cluster until the domain can no longer be expanded. If a text embedding vector does not belong to any cluster and is not labeled as a core point, it is marked as a noise point and removed. This process is repeated after the set is complete. After clustering all text embedding vectors in the dataset, we obtain... Research topic clusters .

[0022] Specifically, to enhance the semantic interpretability of each topic, semantic enhancement is performed on each research topic cluster. The TF-IDF weights and semantic vector central distances of the keywords in the corresponding documents within each research topic cluster are analyzed. The Top-N topic keywords of the research topic cluster are dynamically extracted, including the following steps: For all documents contained within the current research topic cluster, obtain the domain keyword set corresponding to each document, remove duplicate words, and obtain the cluster-level keyword set. Based on the intra-cluster literature data, the word frequency of each keyword in each document in the cluster-level keyword set is counted, and the document frequency of each keyword is counted based on the occurrence of all documents. The inverse document frequency is calculated to calculate the TF-IDF weight of each keyword. Normalize all text embedding vectors within the current research topic cluster, and take the mean of all text embedding vectors within the cluster to obtain the semantic center vector of the research topic cluster, which serves as the core representation of the overall semantics within the cluster. A pre-trained keyword semantic model is used to process each candidate keyword to obtain the corresponding word vector. The semantic vector center distance between the word vector and the semantic center vector of the cluster is calculated. The higher the similarity, the stronger the correlation between the keyword and the core semantics of the topic cluster. For each keyword, its TF-IDF weight and the semantic central distance of its corresponding word vector are weighted and fused according to a preset ratio to obtain the semantic contribution of each keyword; Sort all keywords in descending order of semantic contribution, based on the preset number of keywords. Select the highest score from the sorted results. These words were selected as the Top-N keywords for this research topic cluster.

[0023] like Figure 4As shown, this embodiment provides an innovative multimodal topic evolution association engine. By constructing a dual verification mechanism of vector space similarity and keyword semantic overlap, combined with dynamic threshold screening and keyword overlap ratio filtering, a high-confidence topic evolution chain is constructed to accurately identify evolutionary events such as the continuation, splitting, extinction and rebirth of scientific research topics, and realize interactive visualization of topic evolution paths, thereby revealing the development context and technology migration laws of the scientific research field. Specifically, the similarity of the vector space of research topic clusters in adjacent time slices and the semantic overlap of keywords are calculated, and then filtered using a dynamic threshold. This includes the following steps: Using continuous time slices as units, all research topic clusters included in the previous time period are paired with all research topic clusters included in the next time period to form a candidate set of related topic clusters to be verified, ensuring that no potential evolutionary relationships between adjacent time periods are missed. For each pair of research topic clusters in the candidate set of related topic clusters, extract the topic clusters from the previous time period. The semantic vector center and the topic cluster of the next period The semantic vector centers are used to calculate the vector space similarity between the two using the cosine similarity algorithm, which directly reflects the semantic association strength between topic cluster pairs. Extracting thematic clusters from the previous time period Top-N topic keywords and subsequent topic clusters The top-N topic keywords are used to construct two sets of topic keywords, and the Jaccard similarity coefficient algorithm is used to calculate the ratio of the number of elements in the intersection to the number of elements in the union of the two topic keyword sets, thus obtaining the semantic overlap of the keywords. The formula for quantifying the degree of overlap in the core semantics of two topic clusters is as follows: , In the formula and These represent the sets of keywords corresponding to each research topic cluster. The vector space similarity and key semantic overlap of each topic cluster in the candidate set of related topic clusters are compared with preset dynamic thresholds. Only when both indicators meet the preset dynamic threshold requirements are the topic clusters included in the evolutionary candidate set, thus initially screening out the evolutionary candidate set with semantic relevance. Preferably, the preset dynamic thresholds include a vector space similarity threshold and a keyword semantic overlap threshold. Based on the topic change characteristics of the target research field, a dynamic threshold adjustment model is constructed. If the topic of the field changes rapidly (e.g., frequent technological iterations, dense emerging directions), the double verification threshold can be appropriately reduced to avoid missing potential evolutionary associations. If the topic of the field changes slowly (stable discipline development, concentrated research directions), the double verification threshold can be appropriately increased to strictly filter weak association noise. As an optional implementation method, a default dynamic threshold is preset, wherein the default threshold for vector space similarity is >0.8 and the default threshold for keyword semantic overlap is >0.5. Specifically, to further improve the accuracy of evolutionary relationships, keyword overlap ratio is introduced to filter cross-time period association noise, including the following steps: For topic cluster pairs in the evolutionary candidate set, the sum of TF-IDF weights of the overlapping topic keywords in the two topic clusters and the sum of the sum of TF-IDF weights of the topic keywords of each topic cluster are calculated. The ratio of the sum of the weights of the overlapping topic keywords to the total sum is defined as the keyword overlap ratio, highlighting the influence of high-weight core keywords on evolutionary association. A preset overlap ratio filtering threshold is set. If the overlap ratio of the topic keywords of a topic cluster pair is lower than the preset overlap ratio filtering threshold, its evolutionary relationship is judged as a false association and removed, and finally a set of evolutionary relationships with high confidence is obtained. Optionally, a semantic consistency check is introduced for secondary verification, which specifically includes: combining the domain dictionary and the pre-trained semantic model to verify the semantic consistency of overlapping keywords of topic cluster pairs in the evolutionary relation set, so as to eliminate abnormal associations that are literally identical but semantically unrelated. Specifically, evolutionary events are identified based on a set of high-confidence evolutionary relationships. These events are categorized into four types: theme continuation, theme splitting, theme extinction, and theme emergence. The rules for determining these evolutionary events are as follows: If a single topic cluster in the previous period has a unique high-confidence association with a single topic cluster in the next period, and the core semantics are highly consistent, then it is determined to be a topic continuation; If a single topic cluster in the previous period has a high confidence correlation with multiple topic clusters in the next period, and the core semantics of the topic clusters in the next period are all subsets of the topic clusters in the previous period, then it is determined to be a topic split. If a topic cluster in the previous period does not form a high-confidence association with any topic cluster in the next period, and there are no derived associated objects, it is determined that the topic has disappeared. If a topic cluster in a later time period does not form a high-confidence association with any topic cluster in the previous time period, and it represents a newly added semantic direction within the domain, then it is determined to be a newly emerging topic.

[0024] Specifically, a topic evolution network is constructed with time slices as the horizontal axis, topic clusters as nodes, and high-confidence evolutionary relationships as directed edges. The entire process of the birth, continuation, splitting, or extinction of each topic cluster is sorted out in chronological order and connected to form a complete high-confidence topic evolution chain. Each topic cluster node includes the topic keywords of the topic cluster and is associated with metadata such as the collection of documents it contains.

[0025] Optionally, multi-dimensional visual evolution path diagrams or dynamic cloud maps can be generated based on the theme evolution network to intuitively show the evolution process of the theme and dynamically demonstrate the technology transfer patterns in the scientific research field. Furthermore, feedback from domain experts and academic trend data can be combined to further optimize the display effect of the theme evolution path and improve the interpretability and operability of the theme evolution process.

[0026] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A research topic evolution method based on text embedding vector clustering, characterized in that, Includes the following steps: Collect scientific literature data, preprocess the scientific literature data, identify and extract domain keywords, generate corresponding text embedding vectors, and construct a scientific literature representation vector library. Using a preset time slice as a unit, density adaptive clustering is performed on the text embedding vector to generate scientific research topic clusters within the corresponding time slice. Based on the keyword weights and semantic vector central distances of the documents within the cluster, the Top-N topic keywords of the scientific research topic clusters are dynamically extracted. A dual verification mechanism of vector space similarity and keyword semantic overlap is constructed to dynamically filter the topic evolution relationship between scientific research topic clusters in adjacent time slices, introduce keyword overlap ratio to filter associated noise, and construct a high-confidence topic evolution chain based on the filtered topic evolution relationship.

2. The research topic evolution method based on text embedding vector clustering as described in claim 1, characterized in that, The scientific literature data is preprocessed, including text cleaning, format standardization, noise filtering, and timestamp standardization, to generate standardized literature data. The standardized literature data is then subjected to automatic language discrimination. Based on the discrimination results, the corresponding word segmentation tool is selected to segment the literature text content, remove punctuation marks and stop words, and extract the metadata information of the literature to obtain the word-segmented literature data.

3. The research topic evolution method based on text embedding vector clustering as described in claim 1, characterized in that, Identify and extract domain keywords, including: A dictionary for scientific research is constructed, and a dictionary matching method is used to initially screen the text content of the documents, retaining the words in the dictionary for scientific research as an initial set of candidate keywords; Calculate the word frequency (TF-IDF) of each word in the initial candidate keyword set within the text of each document, and its inverse document frequency (IF) across all documents to obtain the corresponding TF-IDF weights. Sort the keywords in descending order according to these TF-IDF weights, and then filter them. One set of words as secondary keywords; A word co-occurrence network is constructed based on the secondary keyword set. Words are used as nodes and co-occurrence frequency is used as edge weights. The TextRank algorithm is used to calculate the importance score of each word. The calculation is iterative until convergence. The TF-IDF weights and TextRanK importance scores of each word in the secondary keyword set are weighted and fused according to a preset ratio to obtain a comprehensive score. The keywords with the highest comprehensive scores are then selected. The vocabulary is used as the domain keywords for each document.

4. The research topic evolution method based on text embedding vector clustering as described in claim 1, characterized in that, The process of generating the corresponding text embedding vector includes: selecting an appropriate pre-trained text semantic model to semantically encode the text content of the document according to the language type of the document, and extracting the basic semantic vector of the document. Simultaneously, text embedding processing is performed on the domain keywords of each document to obtain keyword enhancement vectors. The basic semantic vector of the document is concatenated with the keyword enhancement vectors to obtain the text embedding vector corresponding to the document. A scientific literature representation vector library is constructed based on the text embedding vectors corresponding to all document data.

5. The research topic evolution method based on text embedding vector clustering as described in claim 1, characterized in that, The step of performing density-adaptive clustering of the text embedding vectors using preset time slices includes: dividing the document data into time dimensions based on the standard timestamps of the document data, dividing the text embedding vectors in the scientific literature representation vector library into multiple consecutive preset time slices, performing density-adaptive clustering of the text embedding vectors in multiple time slices using the DBSCAN clustering algorithm with adaptive neighborhood radius to generate scientific research topic clusters, and then dynamically extracting the Top-N topic keywords of the scientific research topic clusters by combining the keyword TF-IDF weights and semantic vector central distances of the documents within the clusters.

6. The research topic evolution method based on text embedding vector clustering as described in claim 5, characterized in that, Density-adaptive clustering is used to generate research topic clusters, including: For each time slice, calculate the cosine similarity between any two text embedding vectors within the time slice, and calculate the corresponding semantic distance based on the cosine similarity. Construct a local semantic distance matrix based on the semantic distance. For each text embedding vector in a time slice, the distance distribution of its k-nearest neighbors is calculated based on the local semantic distance matrix, and its adaptive neighborhood radius is dynamically calculated based on the distance distribution. Based on the adaptive neighborhood radius, the DBSCAN algorithm is used to cluster the text embedding vectors within the time slice. If the number of samples in the neighborhood of a certain text embedding vector is greater than or equal to the preset minimum number of samples... If so, the text embedding vector is marked as the core point, and its text embedding vectors in the same research cluster are classified into the same research cluster. If a text embedding vector does not belong to any scientific research cluster and is not marked as a core point, it is marked as a noise point and removed. After completing the clustering process, a set of research topic clusters within the corresponding time slice is obtained.

7. The research topic evolution method based on text embedding vector clustering as described in claim 5, characterized in that, Dynamically extract the Top-N topic keywords of the research topic cluster, including: For each research topic cluster, obtain the set of domain keywords for the literature within the cluster, remove duplicate words, and obtain the cluster-level keyword set; Based on intra-cluster literature data, the term frequency and inverse document frequency of each keyword are statistically analyzed, and the corresponding TF-IDF weights are calculated. Normalize all text embedding vectors within the cluster and calculate the semantic center vector of the research topic cluster; The semantic similarity between each keyword and the semantic center vector is calculated using a pre-trained keyword semantic model, which serves as a measure of the relevance between the keyword and the core semantics of the topic cluster. The semantic contribution of each keyword is obtained by weighted fusion of the TF-IDF weight and semantic central distance. All keywords were sorted in descending order of their semantic contribution, and the keywords with the highest scores were selected. These keywords are selected as the Top-N keywords for this research topic cluster.

8. The research topic evolution method based on text embedding vector clustering as described in claim 1, characterized in that, Dynamically filtering thematic evolution relationships between research theme clusters in adjacent time slices includes the following steps: Using time slices as units, research topic clusters from the previous time period and the next time period are paired up to form a candidate set of related topic clusters; Calculate the cosine similarity of the semantic center vectors of each pair of topic clusters, extract the top-N topic keywords of each pair of topic clusters, and calculate the semantic overlap of the keyword set using the Jaccard similarity coefficient. The vector space similarity and keyword semantic overlap of each pair of topic clusters are compared with a preset dynamic threshold to obtain an evolutionary candidate set.

9. The research topic evolution method based on text embedding vector clustering as described in claim 1, characterized in that, Introducing keyword overlap ratio filtering to filter cross-time period noise associations, including: For topic cluster pairs in the evolutionary candidate set, the sum of TF-IDF weights of the overlapping topic keywords in the two topic clusters and the sum of the sum of TF-IDF weights of the topic keywords in each of the two topic clusters are calculated. The ratio of the sum of the weights of the overlapping topic keywords to the sum of the total weights is defined as the keyword overlap ratio. A preset overlap ratio filtering threshold is set. If the overlap ratio of the topic keywords of a topic cluster pair is lower than the preset overlap ratio filtering threshold, its evolutionary relationship is judged as a false association and removed to obtain a set of evolutionary relationships with high confidence.

10. The research topic evolution method based on text embedding vector clustering as described in claim 1, characterized in that, Construct a high-confidence topic evolution chain and implement interactive visualization, including: Evolutionary events are identified based on a set of high-confidence evolutionary relationships. Thematic evolutionary events are classified into four types of evolutionary events according to preset rules: thematic continuation, thematic splitting, thematic extinction, and thematic rebirth. Using time slices as the horizontal axis, theme clusters as nodes, and high-confidence evolutionary relationships as directed edges, a theme evolution network is constructed. The entire process of the birth, continuation, splitting, or extinction of each theme cluster is sorted out in chronological order, and a complete high-confidence theme evolution chain is formed.

Images