A method and system for mining subject evolution paths of battery field literature

Abstract

This invention discloses a method and system for mining thematic evolution paths in battery-related literature, relating to data processing. The method includes: acquiring a set of battery-related literature; performing structured representation on each document; dividing the document into subsets according to a preset time window; independently performing semantic vectorization on the documents in each window; conducting deep semantic thematic modeling and representation enhancement; generating thematic tags through a large language model; quantitatively evaluating the multi-dimensional thematic influence based on the document size and citation count; generating a comprehensive thematic evaluation archive; constructing an evolutionary directed graph based on the thematic evolution correlation strength of adjacent time windows; identifying thematic evolution patterns; and packaging the data into a standard data package and compiling it into a structured analysis report. This application addresses the technical problems of insufficient accuracy in thematic identification and weak ability to depict evolutionary processes in existing technologies, enabling in-depth mining of cutting-edge technological trends.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a method and system for mining topic evolution paths in battery-related literature. Background Technology

[0002] Accurately grasping the technological development trends and effectively identifying core research directions in the battery field has significant industrial and academic value for R&D layout, scientific research planning, and industrial policy formulation in the energy materials field. Currently, the mainstream technical methods for thematic analysis and evolution research of scientific literature can be divided into three categories: traditional text mining and bibliometric methods based on bag-of-words models, word co-occurrence relationships, and probability statistics; deep semantic learning thematic model methods based on deep neural network semantic representation combined with unsupervised clustering; and text analysis and generation methods based on large-scale pre-trained semantic models for text summarization and induction. Existing technical methods lack a deep semantic understanding of professional knowledge in the battery field, and their evaluation dimensions are singular, lacking a quantitative and interpretative framework for dynamic evolution analysis. This makes it difficult to accurately distinguish semantically similar technical topics within the field, easily misjudging the true value of research topics, and failing to accurately characterize and explain the dynamic evolution process of topics.

[0003] Currently, in the relevant technologies, the identification and evolution analysis of topics in scientific and technological literature in the battery field suffers from technical problems such as insufficient accuracy of analysis results, lack of objectivity in evaluation, and weak ability to characterize and explain the evolution process. Summary of the Invention

[0004] This application provides a method and system for mining thematic evolution paths in battery-related literature. It employs techniques such as acquiring battery-related literature from scientific and technological literature databases and performing structured processing; dividing the literature into subset sequences according to preset time windows; independently semantically vectorizing the literature in each time window; enhancing thematic mining through deep semantic thematic modeling and representation; generating thematic tags for each window using a large language model; quantitatively evaluating thematic influence from multiple dimensions based on the scale of literature within the theme and citation counts; generating a comprehensive thematic evaluation archive; constructing a directed thematic evolution graph based on the thematic evolution correlation strength of adjacent time windows and identifying evolutionary patterns such as continuation and splitting; and encapsulating the evolutionary patterns, thematic tags, and the comprehensive evaluation archive into a standard data package, integrating and compiling them into a structured analysis report. These techniques address the technical problems of insufficient accuracy in analysis results, lack of objectivity in evaluation, and weak ability to characterize and explain the evolutionary process in existing thematic identification and evolutionary analysis of battery-related scientific and technological literature. The application achieves the technical effects of improving the accuracy of thematic identification in battery-related scientific and technological literature, enhancing the objectivity of thematic evaluation, and strengthening the characterization and explanation of the thematic evolutionary process.

[0005] This application provides a method for mining thematic evolution paths of literature in the battery field, comprising: obtaining a set of literature related to the battery field from a scientific and technological literature database; performing structured representation on each document; dividing the document set according to a preset time window to obtain a document subset sequence; based on the document subset sequence, independently performing semantic vectorization processing on the documents within each time window, performing deep semantic theme modeling and representation enhancement, and generating theme tags within each time window through a large language model; for the theme tags within each time window, performing multi-dimensional quantitative evaluation of theme influence based on the scale of documents within the theme and the number of citations, and generating a comprehensive theme evaluation file; constructing a directed evolution graph between themes based on the theme evolution correlation strength of adjacent time windows, and identifying theme evolution patterns; and encapsulating the theme evolution patterns, theme tags, and comprehensive theme evaluation files into a standard data package, and compiling them into a structured analysis report.

[0006] In a possible implementation, the following processing is performed: the structured representation includes a title, abstract, publication year, and citation count, and the title and abstract are preprocessed, including cleaning, word segmentation, stop word removal, part-of-speech filtering, and word form restoration.

[0007] In a possible implementation, based on the document subset sequence, the documents within each time window are independently semantically vectorized, and deep semantic topic modeling and representation enhancement are performed. A large language model is used to generate topic tags for each time window, and the following processing is performed: a pre-trained sentence embedding model is used to convert the concatenated text of the document title and abstract into a high-dimensional semantic vector; the high-dimensional semantic vector is dimensionality reduced, and unsupervised clustering is performed to automatically identify semantically dense document clusters, with each document cluster defined as a candidate topic; the mean of the original high-dimensional semantic embedding vectors of all documents under each candidate topic is calculated as the semantic center vector, and the topic representation is optimized to obtain an optimized set of topic keywords; the set of topic keywords is integrated with representative documents to construct prompt words containing domain constraints, core topic information, and generation requirements; the large model is then invoked to generate semantic tags for each topic, resulting in topic tags for each time window.

[0008] In a possible implementation, the high-dimensional semantic vector is dimensionality reduced and unsupervised clustering is performed to automatically identify semantically dense document clusters. Each document cluster is defined as a candidate topic. The following processing is performed: the high-dimensional semantic vector is reduced to a low-dimensional space using the UMAP algorithm to obtain a low-dimensional vector; the HDBSCAN density clustering algorithm is used to automatically identify semantically dense document clusters based on the low-dimensional vector, and each document cluster is defined as a candidate topic.

[0009] In a possible implementation, the mean of the original high-dimensional semantic embedding vectors of all documents under each candidate topic is calculated as the semantic center vector. The topic representation is then optimized to obtain an optimized set of topic keywords. The following processing is then performed: For each candidate topic, candidate keywords are extracted based on document word frequency using a TF-IDF-like algorithm, and initial statistical weights are calculated. For each candidate keyword, the semantic similarity between it and the corresponding topic semantic center vector is calculated. The initial weights of the candidate keywords are reweighted based on the semantic similarity using a nonlinear enhancement function to generate reweighted weights. The candidate keywords are sorted according to the reweighted weights, and the top M words with the highest weights are selected to form the set of topic keywords.

[0010] In a possible implementation, based on the evolutionary association strength of topics in adjacent time windows, an evolutionary directed graph is constructed between topics to identify topic evolution patterns and perform the following processing: calculate the evolutionary association strength between topics in adjacent time windows and construct a similarity matrix; set a dynamic threshold based on the distribution characteristics of the similarity matrix, filter out topic pairs that are higher than the dynamic threshold, and generate directed evolutionary edges; construct the evolutionary directed graph with all topics as nodes and evolutionary edges as directed edges; analyze the connection patterns of each node based on the evolutionary directed graph to identify the topic evolution patterns.

[0011] In possible implementations, the following processing is performed: the theme evolution patterns include continuation, splitting, merging, emerging, and decay.

[0012] In a possible implementation, the following process is performed: the upper quartile or mean of all non-zero similarities in the similarity matrix is ​​added to a standard deviation as a dynamic threshold.

[0013] In a possible implementation, the following processing is performed: the indicators for multi-dimensional quantitative evaluation of topic influence include size indicators, influence indicators, and vitality indicators; wherein, the size indicator is the number of documents within the topic; the influence indicator includes average citation influence per article and total citation influence; and the vitality indicator is a weighted average of size vitality and influence vitality.

[0014] This application also provides a system for mining the theme evolution path of literature in the battery field, comprising: a literature acquisition module, used to acquire a set of literature related to the battery field from a scientific and technological literature database, perform structured representation on each document, and divide the document set according to a preset time window to obtain a document subset sequence; a theme tag generation module, used to independently perform semantic vectorization processing on the documents within each time window based on the document subset sequence, perform deep semantic theme modeling and representation enhancement, and generate theme tags within each time window through a large language model; a theme influence quantification and evaluation module, used to perform multi-dimensional theme influence quantification and evaluation based on the document size and citation count within the theme tags in each time window, and generate a comprehensive theme evaluation file; a theme evolution pattern recognition module, used to construct a directed evolution graph between themes based on the theme evolution correlation strength between adjacent time windows, and identify theme evolution patterns; and a structured analysis report generation module, used to encapsulate the theme evolution patterns, theme tags, and comprehensive theme evaluation files into a standard data package, and compile them into a structured analysis report.

[0015] This application proposes a method and system for mining thematic evolution paths in battery-related literature. First, it retrieves a collection of battery-related literature from a scientific literature database. Each document is then structurally represented, and the collection is divided into subsets according to a preset time window. Next, based on these subsets, the literature within each time window undergoes independent semantic vectorization, deep semantic thematic modeling, and representation enhancement. A large language model generates thematic tags for each time window. Then, for each thematic tag within a time window, a multi-dimensional thematic influence quantification assessment is performed based on the size of the literature within the theme and its citation count, generating a comprehensive thematic assessment archive. Furthermore, based on the thematic evolution correlation strength between adjacent time windows, a directed evolutionary graph between themes is constructed to identify thematic evolution patterns. Finally, the thematic evolution patterns, thematic tags, and the comprehensive thematic assessment archive are packaged into a standard data package and compiled into a structured analysis report. Through this process, the method and system proposed in this application achieve the technical effects of improving the accuracy of thematic identification in battery-related scientific literature, enhancing the objectivity of thematic assessment, and strengthening the characterization and explanation of thematic evolution processes. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a flowchart illustrating a method for mining topic evolution paths in battery-related literature, as provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the structure of a topic evolution path mining system for battery-related literature, provided as an embodiment of this application.

[0019] Figure labeling: Document acquisition module 10, topic tag generation module 20, topic influence quantification assessment module 30, topic evolution pattern recognition module 40, structured analysis report generation module 50. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] This application provides a method for mining topic evolution paths in battery-related literature, such as... Figure 1 As shown, the method includes:

[0022] Step S100: Obtain a set of literature related to the battery field from the scientific and technological literature database, perform structured representation on each literature, and divide the literature set according to a preset time window to obtain a literature subset sequence. The structured representation includes title, abstract, publication year and citation count. The title and abstract are preprocessed, including text cleaning, word segmentation, stop word removal, part-of-speech filtering and word form restoration.

[0023] Specifically, battery-related literature within the target scope is obtained from professional scientific and technological literature databases. For example, databases such as Web of Science, Scopus, and CNKI are selected as data sources. Advanced searches are conducted using core keywords in the battery field, such as "lithium-ion battery," "solid-state battery," and "lithium metal anode," limiting the document type to academic papers, patent documents, and other target types to obtain the raw data. The core metadata of each document is then organized into a standardized structured format. For example, the raw data is structured using Python's Pandas library, extracting the title, abstract, publication year, and citation count fields for each document and constructing a structured data table. Missing citation counts are padded with zeros, and the publication year format is standardized. The extracted titles and abstracts undergo text preprocessing, including text cleaning, word segmentation, stop word removal, part-of-speech filtering, and lemmatization. For Chinese text, the focus is on text cleaning, word segmentation, stop word removal, and part-of-speech filtering, while for English text, lemmatization is added. Then, all documents are processed by windowing according to a pre-defined time window, forming a sequence of document subsets arranged chronologically. For example, the time window length can be set to 2 or 3 years. Structured document data is divided by publication year using sliding windows or non-overlapping windows. For instance, documents from 2016 to 2025 are divided into 5 non-overlapping subsets with 2-year windows. Each subset serves as a document set for a given time window, and the window numbers are named sequentially according to time, forming an ordered sequence of document subsets.

[0024] Step S200: Based on the document subset sequence, semantic vectorization is performed independently on the documents within each time window, deep semantic topic modeling and representation enhancement are carried out, and topic tags within each time window are generated through a large language model.

[0025] Specifically, the literature data is split into time windows to ensure that the literature data in each window is processed independently, avoiding semantic interference across time periods. For the literature in each window, the title and abstract text are first concatenated, and then semantic vectorization, dimensionality reduction clustering, semantic center calculation, and keyword optimization are performed sequentially to obtain the candidate topics and optimized keyword set for each window. Representative documents are selected, such as the top 5 most cited documents in the topic, and their core content is extracted and integrated with the topic keyword set to construct prompt words containing domain constraints and topic content. Large language models, such as DeepSeek, Wenxin Yiyan, and Tongyi Qianwen, are called to guide the model to generate semantic tags that conform to the professional standards of the battery field through prompt engineering. For example, the semantic tag "Lithium metal anode dendrite suppression and solid electrolyte interface stability" is generated for topics containing keywords such as "lithium dendrites," "solid electrolyte interface," and "suppression." Finally, standardized topic tags corresponding to all topics in each time window are obtained, realizing accurate identification and enhanced representation of battery field literature topics.

[0026] In one possible implementation, based on the document subset sequence, the documents within each time window are independently semantically vectorized, and deep semantic topic modeling and representation enhancement are performed. Topic tags for each time window are generated using a large language model. Step S200 further includes: Step S210 involves using a pre-trained sentence embedding model to convert the concatenated text of document titles and abstracts into high-dimensional semantic vectors. Specifically, sentence embedding models pre-trained on massive amounts of scientific and technological texts, such as Sentence-BERT, SciBERT, and all-MiniLM-L6-v2, are selected as the core models. These models take the concatenated text string as input and output a fixed-dimensional high-dimensional semantic vector. The embedding dimension can be selected according to actual needs, such as 768 dimensions or 384 dimensions. The pre-processed concatenated text of the documents is batch-input into the pre-trained sentence embedding model. The model's encoding layer performs semantic encoding on the text, outputting a high-dimensional semantic vector corresponding to each document. The semantic vectors of all documents constitute the high-dimensional semantic vector set for that time window.

[0027] Step S220 involves dimensionality reduction of the high-dimensional semantic vectors and unsupervised clustering to automatically identify semantically dense document clusters. Each document cluster is defined as a candidate topic. Specifically, the high-dimensional semantic vector set is simplified by using the UMAP algorithm to reduce its dimensionality, preserving the core semantic relationships between vectors and avoiding semantic information loss. Then, an unsupervised clustering algorithm such as HDBSCAN density clustering is used to automatically identify semantically dense clusters based on the low-dimensional vectors. Each semantically dense document cluster is defined as a candidate topic, and the document IDs contained in each candidate topic are recorded, achieving a preliminary division from document vectors to topics. Simultaneously, documents not clustered are marked as noise and removed. For example, after dimensionality reduction and clustering of 10,000 documents in a certain time window, 25 semantically dense clusters (25 candidate topics) are obtained, and the remaining 50 documents are considered noise.

[0028] Step S230: Calculate the mean of the original high-dimensional semantic embedding vectors of all documents under each candidate topic as the semantic center vector, and optimize the topic representation to obtain an optimized set of topic keywords. Specifically, candidate keywords are extracted using a TF-IDF-like algorithm and initial weights are calculated; the cosine similarity between each candidate keyword and the topic semantic center vector is calculated; based on the semantic similarity, the initial weights of the candidate keywords are reweighted using the Sigmoid function; the keywords are sorted according to the reweighted weights, and the top M words with the highest weights are selected to form the optimized set of topic keywords. The value of M is generally 5-10, taking into account the characteristics of the battery domain topic. For example, if M is 8, then 8 core keywords are retained for each topic.

[0029] Step S240: Integrate the aforementioned keyword set with representative documents to construct prompt words. Call the large model to generate semantic tags for each topic, obtaining topic tags within each time window. Specifically, select representative documents from each candidate topic, using the selection rule of the top 3-5 most cited documents within the topic, and extract the titles and core abstract information of the representative documents. Construct prompt words for the large language model, which include domain constraints, core topic information, and generation requirements. Call the pre-trained large language model, inputting the constructed prompt words into the model's text input interface. Set the model's output format to a single topic tag, obtain the semantic tags generated by the model, and regenerate tags that do not conform to the specifications. Finally, generate concise, professional semantic tags that conform to the academic norms of the battery field for each candidate topic, achieving standardized topic naming. For example, generate the semantic tag "Lithium metal anode dendrite suppression and SEI film regulation" for topics containing keywords such as "lithium metal anode," "dendrite suppression," "solid electrolyte interface," and "SEI film."

[0030] In one possible implementation, the high-dimensional semantic vectors are dimensionality reduced and unsupervised clustering is performed to automatically identify semantically dense document clusters. Each document cluster is defined as a candidate topic. Step S220 further includes step S221, using the UMAP algorithm to reduce the high-dimensional semantic vectors to a low-dimensional space, obtaining low-dimensional vectors. Specifically, the high-dimensional semantic vector set obtained in step S210 is used as the input data for the UMAP algorithm. The core parameters of the algorithm are set, where the low-dimensional space dimension is set to 2 or 3, the number of nearest neighbors is set to 15-50, the minimum distance is set to 0.1-0.5, and the distance metric is cosine distance. Then, the core process of the UMAP algorithm is executed: first, a nearest neighbor graph of the high-dimensional vectors is constructed; then, the nearest neighbor graph is mapped to the low-dimensional space by optimizing the objective function, ensuring that the distance relationship between the low-dimensional vectors is consistent with that of the high-dimensional vectors, preserving the local and global semantic structure between the high-dimensional vectors to the greatest extent, and finally outputting a low-dimensional vector set suitable for clustering.

[0031] Step S222: The HDBSCAN density clustering algorithm is used to automatically identify semantically dense document clusters based on the low-dimensional vectors, defining each document cluster as a candidate topic. Specifically, the low-dimensional vector set obtained in step S221 is used as the input data for the HDBSCAN algorithm. The core parameters of the algorithm are set, with the minimum cluster size set to 5-20, the minimum number of samples set to be the same as the minimum cluster size, and cosine distance selected as the distance metric. Then, the core process of the HDBSCAN algorithm is executed: first, the density reachability of the low-dimensional vectors is calculated, a clustering tree is constructed, and then clusters of different densities are obtained through tree cutting. Noise points with excessively low density are automatically removed. Finally, the clustering results are output, including the document number corresponding to each cluster and the number of clusters, defining each cluster as a candidate topic. By automatically identifying semantically dense document clusters based on vector density features, without the need to pre-set the number of clusters, this method can effectively handle the non-uniform distribution characteristics of battery-related literature data and improve the accuracy of topic segmentation.

[0032] In one possible implementation, the mean of the original high-dimensional semantic embedding vectors of all documents under each candidate topic is calculated as the semantic center vector. The topic representation is then optimized to obtain an optimized set of topic keywords. Step S230 further includes step S231, where, for each candidate topic, candidate keywords are extracted based on document word frequency using a TF-IDF-like algorithm, and initial statistical weights are calculated. Specifically, all document texts within each candidate topic are merged, and Chinese word segmentation / English lemma restoration, removal of stop words and domain-general meaningless words are performed to construct a topic word frequency statistics library. Then, a TF-IDF-like algorithm is executed to calculate the topic-specific word frequency (TF) of each word, which is the ratio of the number of times the word appears in the merged text of that topic to the total number of words in that topic. The inverse topic frequency (IDF) of the word is then calculated, which is the logarithm of the total number of topics within that time window to the number of topics containing that word. The initial statistical weight is the product of TF and IDF. Finally, the words are sorted in descending order according to the initial statistical weights, and the top 20-30 words are selected as the initial candidate keywords for that candidate topic.

[0033] Step S232: For each candidate keyword, calculate the semantic similarity between it and the corresponding topic semantic center vector. Specifically, use the same pre-trained sentence embedding model as in step S210, and input each candidate keyword as an independent text model to generate a keyword semantic vector with the same dimension as the document semantic vector. Then, use the topic semantic center vector and the semantic vector of each candidate keyword as the calculation objects, and use cosine similarity as the similarity calculation method to output the semantic similarity value corresponding to each candidate keyword.

[0034] Step S233: The initial weights of candidate keywords are reweighted based on the semantic similarity using a non-linear enhancement function to generate reweighted weights. Specifically, the reweighting calculation method is determined: the reweighted weight is the product of the initial statistical weight and the non-linear enhancement function based on semantic similarity. The non-linear enhancement function may include the Sigmoid function, exponential function, etc. The initial statistical weight and corresponding semantic similarity of each candidate keyword are substituted into the reweighting formula to calculate the reweighted weight. For general terms with semantic similarity below a set threshold, an additional weight decay coefficient can be set to further reduce their weight. Finally, the reweighted weight corresponding to each candidate keyword is output.

[0035] Step S234: Sort the candidate keywords according to their reweighted weights, and select the top M keywords with the highest weights to form the topic keyword set. Specifically, all initial candidate keywords under each candidate topic are sorted in descending order according to their reweighted weights to form an ordered keyword list. A fixed number M is set based on the research characteristics and topic representation requirements of the battery field. The top M keywords with the highest weights are selected from the ordered keyword list to form the optimized keyword set for that candidate topic, and keywords with lower weights are removed.

[0036] Step S300: For each topic tag within a time window, a multi-dimensional quantitative assessment of topic influence is conducted based on the scale of literature within the topic and the number of citations, generating a comprehensive topic assessment file. The indicators for the multi-dimensional quantitative assessment of topic influence include scale indicators, influence indicators, and vitality indicators. The scale indicator is the number of literatures within the topic; the influence indicator includes average citation influence per article and total citation influence; and the vitality indicator is a weighted average of scale vitality and influence vitality.

[0037] Specifically, based on the topic segmentation obtained in step S200, for each topic within each time window, the number of all documents and citation counts for each topic are extracted, and the size index, influence index, and vitality index are calculated sequentially. The following processing is then performed: 1. Scale Indicator Calculation. The scale indicator is calculated directly from the number of documents included in the topic. This indicator quantifies the research activity and output volume of this research direction, reflecting the size of the research community or the breadth of attention received by the technical direction. The larger the value, the higher the activity level of the research direction within this time window, the larger the research community, or the wider the attention received by the technical direction.

[0038] 2. Calculation of Impact Indicators. Impact indicators aim to assess the average academic quality and recognition of research findings within a given topic. They are obtained through statistical analysis of citation counts for all literature within the topic, including average citation impact per paper and total citation impact. Average citation impact per paper is the arithmetic mean of citation counts for all literature within the topic, eliminating the influence of topic size and directly measuring the average impact of individual studies, reflecting the average quality of the topic. Total citation impact is the sum of citation counts for all literature within the topic, reflecting the overall academic influence of the topic.

[0039] 3. Vitality Index Calculation. The vitality index is used to identify emerging topics that are in a rapid growth phase or have development potential. Its calculation comprehensively considers the topic's growth trend in scale and the rate of influence accumulation within a recent time window. Specifically, it calculates the compound annual growth rate of the topic within the current time window compared to the set of preceding topics with which it has an evolutionary relationship in the previous time window. Specifically, the vitality index first calculates scale vitality and influence vitality separately. Scale vitality is calculated based on the growth in the number of documents, representing the difference between the current topic size and the total size of preceding topics, divided by the larger of the total size of preceding topics and a minimum constant. For emerging topics, the total size of preceding topics is recorded as 0, and the minimum constant is set to 1 to prevent division by zero. Influence vitality is calculated based on the growth in average citations per article, representing the difference between the current topic's average citation influence per article and the weighted average average citation influence per article of preceding topics, divided by the weighted average average citation influence per article of preceding topics and a minimum constant. The larger value among the constants is the weighted average citation influence of the preceding topics, which is the sum of the total citation influence of all preceding topics divided by the total size of the preceding topics. If it is an emerging topic, the weighted average citation influence of the preceding topics is recorded as 0. Then, the size vitality and influence vitality are input into the Sigmoid function for normalization, and the results are mapped to the interval of 0 to 1 to smooth out extreme values. Then, the normalized size vitality and influence vitality are weighted and summed according to the weight coefficients α and β to obtain the comprehensive vitality index. α and β are set according to the analysis requirements.

[0040] 4. Generation of Comprehensive Evaluation Files for Each Topic. A structured comprehensive evaluation file is generated for each topic, integrating its semantic representation and all quantitative indicators. The file contains core information such as topic tags, optimized keyword set, topic semantic center vector, scale indicators, article citation impact, total citation impact, and comprehensive indicators, and is stored in structured formats such as JSON and Excel.

[0041] Step S400: Based on the correlation strength of topic evolution in adjacent time windows, construct a directed graph of topic evolution to identify topic evolution patterns.

[0042] Specifically, the comprehensive evaluation files of topics from two adjacent time windows are extracted chronologically, and the fusion similarity between all topic pairs is calculated. The fusion similarity is obtained by weighting vector cosine similarity and keyword Jaccard similarity to construct a similarity matrix. A dynamic threshold is set based on the distribution characteristics of the similarity matrix, and topic pairs with similarity higher than the threshold are selected, generating directed evolutionary edges with the direction from the previous window to the next window, and the edge weight being the fusion similarity. Topics from all time windows are used as nodes, and the selected directed evolutionary edges are used as edges to construct a global directed graph of topic evolution, where the node attribute is the corresponding comprehensive evaluation file. Finally, the in-degree and out-degree characteristics of each node in the directed graph are analyzed. Based on the values ​​of in-degree and out-degree, five typical evolutionary modes—continuation, splitting, fusion, emerging, and decay—are automatically identified, achieving a classification and characterization of the topic evolution process.

[0043] In one possible implementation, based on the correlation strength of topic evolution between adjacent time windows, a directed evolution graph between topics is constructed to identify topic evolution patterns. Step S400 further includes: Step S410: Calculate the evolutionary association strength between topics in adjacent time windows and construct a similarity matrix. Specifically, for two adjacent time windows, obtain all topics within the preceding and following time windows respectively, and calculate the evolutionary association strength between topic pairs using a fusion metric. The fusion metric includes: calculating the cosine similarity between the semantic center vectors of topic pairs to characterize the similarity of deep research connotations and semantic orientations; calculating the Jaccard similarity (i.e., the ratio of intersection size to union size) between the optimized keyword sets of topic pairs to characterize the overlap at the explicit research terminology level. The cosine similarity and the Jaccard similarity are weighted and summed according to preset weights to obtain the fusion similarity; wherein the preset weight of the cosine similarity is configured to be greater than the preset weight of the Jaccard similarity. Traverse all topic pairs in adjacent time windows and perform the above calculations, outputting the calculation results in matrix form to construct a two-dimensional similarity matrix. The rows and columns of the similarity matrix correspond to the topics within the preceding and following time windows, respectively, and the values ​​of the matrix elements are the fusion similarity of the corresponding topic pairs.

[0044] Step S420: Based on the distribution characteristics of the similarity matrix, a dynamic threshold is set to filter out topic pairs with similarity values ​​higher than the dynamic threshold and generate directed evolutionary edges. In one specific implementation, the dynamic threshold can be dynamically determined according to the numerical distribution characteristics of the similarity matrix (for example, preferably set to 0.45, or statistically determined based on data distribution characteristics). All matrix elements in the similarity matrix are traversed. If the fusion similarity corresponding to a certain matrix element is greater than or equal to the dynamic threshold, it is determined that there is a significant evolutionary correlation between the corresponding topic pairs, and a directed evolutionary edge is generated accordingly. The direction of the directed evolutionary edge is strictly configured to point from the source topic of the previous time window to the target topic of the next time window in chronological order; the weight of the directed evolutionary edge is assigned as the corresponding fusion similarity to measure and characterize the strength of the evolutionary relationship. For topic pairs with fusion similarity lower than the dynamic threshold, no evolutionary edge is generated.

[0045] Step S430: Construct the evolutionary directed graph using all topics as nodes and evolutionary edges as directed edges. Specifically, assign a unique identifier containing a time window number and a topic number to each topic node, and store the attribute information (including topic tags, keyword sets, semantic center vectors, scale indicators, and vitality indicators, etc.) from the pre-generated topic comprehensive evaluation file as node attributes. Map the directed evolutionary edges and their corresponding weights and directions to the corresponding topic nodes. Due to the natural hierarchical nature of time windows, the topic evolutionary directed graph is configured as a structured directed acyclic graph (DAG) data model to fully depict the technological evolution and inheritance of domain topics over time. A global directed graph is constructed using graph data structure libraries, such as Python's NetworkX and PyTorch Geometric. The node set and edge set are imported into the library to generate standardized directed graph data, enabling visualization and structured representation of the evolutionary process of battery-related topics.

[0046] Step S440: Based on the analysis of the connection patterns of each node in the evolutionary directed graph, the theme evolution patterns are identified. These patterns include continuation, splitting, merging, emerging, and decay. Specifically, the in-degree and out-degree of each node in the evolutionary directed graph are calculated. The in-degree is the number of directed evolutionary edges pointing to that node, representing the number of predecessor topics for that theme. The out-degree is the number of directed evolutionary edges pointing from that node to other nodes, representing the number of successor topics for that theme. Based on the in-degree and out-degree values, five evolutionary modes are defined using the following rules: Continuation mode: Node in-degree approximately 1, out-degree 1, indicating stable and continuous evolution of the research direction; Splitting mode: Node in-degree approximately 1, out-degree ≥ 2, representing the differentiation of technical routes or the refinement of research focus; Fusion mode: Node in-degree ≥ 2, out-degree 1, reflecting the cross-fertilization or integrated innovation of different technical routes, a common mode for generating new breakthroughs; Emerging mode: Node in-degree 0, indicating that the topic may be a newly emerging research direction or a cutting-edge hotspot; Decline mode: Node out-degree 0, suggesting that the research direction may be temporarily stagnant, replaced, or integrated into other directions. All nodes in the evolutionary directed graph are traversed, and the evolutionary mode is labeled for each node's corresponding topic according to the rules.

[0047] Step S500: The topic evolution pattern, topic tags, and topic comprehensive evaluation files are packaged into a standard data package and compiled into a structured analysis report.

[0048] Specifically, standardized data packages are constructed for each topic, each evolutionary relationship, and each evolutionary chain. These packages are in JSON format and contain core content such as semantic information (topic tags, keywords), quantitative indicators (scale, influence, vitality), evolutionary information (evolutionary patterns, predecessor / successor topics, similarity weights), and domain constraints (specific sub-directions within the battery field). Based on prompting engineering, prompt templates of different granularities are constructed. A large language model is used to generate in-depth analysis text for each data package, including in-depth topic descriptions, evolutionary relationship narratives, and interpretations of evolutionary chain trends. The analysis text generated by the large language model, the visualized evolutionary directed graph, and the statistical tables of quantitative indicators are integrated and compiled into a structured analysis report according to a modular structure. The report includes sections such as an executive summary, a brief methodology overview, a panoramic view of the domain topics, in-depth analysis of key evolutionary processes, identification of emerging frontiers and declining topics, and conclusions and prospects. It is output in Word, PDF, and other formats.

[0049] This application employs a series of technical methods, including acquiring battery-related literature from a scientific and technological literature database and performing structured processing, dividing the literature into subset sequences according to preset time windows, independently semantically vectorizing the literature in each time window, enhancing topic mining through deep semantic topic modeling and representation, generating topic tags for each window using a large language model, quantitatively evaluating topic influence from multiple dimensions based on the number of documents and citations within a topic, generating a comprehensive topic evaluation archive, constructing a directed topic evolution graph based on the evolutionary correlation strength of topics in adjacent time windows and identifying evolutionary patterns such as continuation and splitting, encapsulating evolutionary patterns, topic tags, and comprehensive evaluation archives into a standard data package, and integrating and compiling them into a structured analysis report. These methods address the technical problems of insufficient accuracy in analysis results, lack of objectivity in evaluation, and weak ability to characterize and explain the evolutionary process in existing battery-related scientific and technological literature topic identification and evolutionary analysis. This achieves the technical effect of improving the accuracy of topic identification in battery-related scientific and technological literature, enhancing the objectivity of topic evaluation, and strengthening the characterization and explanation of the topic evolutionary process.

[0050] In the above text, refer to Figure 1 This paper describes in detail a method for mining topic evolution paths in battery-related literature according to an embodiment of the present invention. Next, we will refer to... Figure 2 This invention describes a topic evolution path mining system for battery-related literature according to an embodiment of the present invention.

[0051] This invention provides a topic evolution path mining system for battery-related literature, addressing the technical problems of insufficient accuracy in analysis results, lack of objectivity in evaluation, and weak ability to characterize and explain the evolution process in existing battery-related scientific and technological literature topic identification and evolution analysis. The system aims to improve the accuracy of topic identification in battery-related scientific and technological literature, enhance the objectivity of topic evaluation, and strengthen the characterization and explanation of the topic evolution process. The system includes: a literature acquisition module 10, a topic tag generation module 20, a topic influence quantification evaluation module 30, a topic evolution pattern recognition module 40, and a structured analysis report generation module 50.

[0052] The document acquisition module 10 is used to acquire a collection of documents related to the battery field from a scientific and technological literature database, perform structured representation on each document, and divide the document collection according to a preset time window to obtain a document subset sequence. The topic tag generation module 20 is used to independently perform semantic vectorization processing on the documents within each time window based on the document subset sequence, perform deep semantic topic modeling and representation enhancement, and generate topic tags within each time window through a large language model. The topic influence quantification assessment module 30 is used to perform multi-dimensional topic influence quantification assessment based on the document size and citation count within each topic tag, and generate a comprehensive topic assessment file. The topic evolution pattern recognition module 40 is used to construct an evolutionary directed graph between topics based on the topic evolution correlation strength between adjacent time windows and identify topic evolution patterns. The structured analysis report generation module 50 is used to encapsulate the topic evolution patterns, topic tags, and comprehensive topic assessment files into a standard data package and compile them into a structured analysis report.

[0053] The detailed description of the specific configuration of the document acquisition module 10 is explained as follows: As mentioned above, the document acquisition module 10 may further include: a structured representation including title, abstract, publication year and citation count, and text preprocessing of the title and abstract, the preprocessing including cleaning, word segmentation, stop word removal, part-of-speech filtering and word form restoration.

[0054] The specific configuration of the topic tag generation module 20 is described in detail below: As mentioned above, based on the document subset sequence, the documents within each time window are independently semantically vectorized, and deep semantic topic modeling and representation enhancement are performed. Topic tags for each time window are generated through a large language model. The topic tag generation module 20 may further include: a text conversion unit that uses a pre-trained sentence embedding model to convert the concatenated text of the document title and abstract into a high-dimensional semantic vector; an unsupervised clustering unit that performs dimensionality reduction on the high-dimensional semantic vector and performs unsupervised clustering to automatically identify semantically dense document clusters, with each document cluster defined as a candidate topic; a topic optimization unit that calculates the mean of the original high-dimensional semantic embedding vectors of all documents under each candidate topic as the semantic center vector, optimizes the topic representation, and obtains an optimized set of topic keywords; and a topic tag generation unit that integrates the set of topic keywords with representative documents, constructs prompt words, calls the large model, and generates semantic tags for each topic, thus obtaining topic tags for each time window.

[0055] Specifically, the high-dimensional semantic vector is subjected to dimensionality reduction processing and unsupervised clustering to automatically identify semantically dense document clusters. Each document cluster is defined as a candidate topic. The unsupervised clustering unit may further include: a dimensionality reduction subunit used to reduce the high-dimensional semantic vector to a low-dimensional space using the UMAP algorithm to obtain a low-dimensional vector; and a clustering subunit used to automatically identify semantically dense document clusters based on the low-dimensional vector using the HDBSCAN density clustering algorithm, defining each document cluster as a candidate topic.

[0056] The optimization unit further includes: a candidate keyword extraction subunit, which calculates the mean of the original high-dimensional semantic embedding vectors of all documents under each candidate topic as the semantic center vector, and optimizes the topic representation to obtain an optimized topic keyword set; a candidate keyword extraction subunit, which extracts candidate keywords based on document word frequency using a TF-IDF-like algorithm for each candidate topic and calculates initial statistical weights; a semantic similarity calculation subunit, which calculates the semantic similarity between each candidate keyword and the corresponding topic semantic center vector; a reweighting subunit, which reweights the initial weights of the candidate keywords using a non-linear enhancement function based on the semantic similarity to generate reweighted weights; and a topic keyword set generation subunit, which sorts the candidate keywords according to the reweighted weights and selects the top M words with the highest weights to form the topic keyword set.

[0057] The specific configuration of the topic evolution pattern recognition module 40 is described in detail below: As mentioned above, based on the topic evolution correlation strength of adjacent time windows, an evolutionary directed graph between topics is constructed to identify topic evolution patterns. The topic evolution pattern recognition module 40 may further include: a similarity matrix construction unit for calculating the evolutionary correlation strength between topics in adjacent time windows and constructing a similarity matrix; a directed evolution edge generation unit for setting a dynamic threshold based on the distribution characteristics of the similarity matrix, filtering out topic pairs higher than the dynamic threshold, and generating directed evolution edges; an evolutionary directed graph construction unit for constructing the evolutionary directed graph with all topics as nodes and evolution edges as directed edges; and a topic evolution pattern recognition unit for analyzing the connection pattern of each node based on the evolutionary directed graph to identify the topic evolution pattern.

[0058] The theme evolution pattern recognition unit may further include: the theme evolution pattern includes continuation, splitting, merging, emerging and decaying.

[0059] The directed evolutionary edge generation unit may further include: taking the upper quartile or mean of all non-zero similarities in the similarity matrix plus a standard deviation as a dynamic threshold.

[0060] The detailed description of the specific configuration of the topic influence quantitative assessment module 30 is explained as follows: As mentioned above, the topic influence quantitative assessment module 30 may further include: indicators for multi-dimensional topic influence quantitative assessment, including scale indicators, influence indicators and vitality indicators, wherein the scale indicator is the number of documents within the topic; the influence indicator includes average citation influence and total citation influence; and the vitality indicator is a weighted average of scale vitality and influence vitality.

[0061] The topic evolution path mining system for battery-related literature provided in this embodiment of the invention can execute the topic evolution path mining method for battery-related literature provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0062] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for mining topic evolution paths in battery-related literature, characterized in that, include: A collection of literature related to the field of batteries is obtained from a scientific and technological literature database. Each article is represented in a structured manner, and the collection is divided according to a preset time window to obtain a sequence of literature subsets. Based on the document subset sequence, the documents in each time window are independently semantically vectorized, and deep semantic topic modeling and representation enhancement are performed. The topic tags in each time window are generated through a large language model. For each time window, based on the number of documents and citations within the topic, a multi-dimensional quantitative assessment of the topic's influence is conducted to generate a comprehensive topic assessment file. Based on the correlation strength of topic evolution in adjacent time windows, a directed graph of topic evolution is constructed to identify topic evolution patterns; The aforementioned topic evolution patterns, along with topic tags and topic comprehensive evaluation files, are packaged into a standard data package and compiled into a structured analysis report.

2. The method for mining topic evolution paths in battery-related literature as described in claim 1, characterized in that, The structured representation includes the title, abstract, publication year, and citation count. The title and abstract are preprocessed, including text cleaning, word segmentation, stop word removal, part-of-speech filtering, and word form restoration.

3. The method for mining topic evolution paths in battery-related literature as described in claim 1, characterized in that, Based on the aforementioned document subset sequence, semantic vectorization is performed independently on the documents within each time window. Deep semantic topic modeling and representation enhancement are then conducted, and topic tags for each time window are generated using a large language model, including: A pre-trained sentence embedding model is used to convert the concatenated text of the document's title and abstract into a high-dimensional semantic vector; The high-dimensional semantic vector is reduced in dimensionality and unsupervised clustering is performed to automatically identify semantically dense document clusters, and each document cluster is defined as a candidate topic. The mean of the original high-dimensional semantic embedding vectors of all documents under each candidate topic is calculated as the semantic center vector. The topic representation is then optimized to obtain the optimized set of topic keywords. By integrating the aforementioned set of topic keywords with representative documents, prompt words are constructed, a large model is invoked, and semantic tags are generated for each topic, thus obtaining topic tags within each time window.

4. The method for mining topic evolution paths in battery-related literature as described in claim 3, characterized in that, The high-dimensional semantic vector is reduced in dimensionality and then subjected to unsupervised clustering to automatically identify semantically dense document clusters. Each document cluster is defined as a candidate topic, including: The UMAP algorithm is used to reduce the high-dimensional semantic vector to a low-dimensional space to obtain a low-dimensional vector. The HDBSCAN density clustering algorithm is used to automatically identify semantically dense document clusters based on the low-dimensional vectors, and each document cluster is defined as a candidate topic.

5. The method for mining topic evolution paths in battery-related literature as described in claim 3, characterized in that, The mean of the original high-dimensional semantic embedding vectors of all documents under each candidate topic is calculated as the semantic center vector. This is then used to optimize the topic representation, resulting in an optimized set of topic keywords, including: For each candidate topic, candidate keywords are extracted based on document term frequency using a TF-IDF-like algorithm and initial statistical weights are calculated. For each candidate keyword, calculate the semantic similarity between it and the corresponding topic semantic center vector; The initial weights of candidate keywords are reweighted based on the semantic similarity using a nonlinear enhancement function to generate reweighted weights. Candidate keywords are sorted according to their reweighted weights, and the top M keywords with the highest weights are selected to form the topic keyword set.

6. The method for mining topic evolution paths in battery-related literature as described in claim 1, characterized in that, Based on the correlation strength of topic evolution between adjacent time windows, a directed graph of topic evolution is constructed to identify topic evolution patterns, including: Calculate the evolutionary association strength between topics in adjacent time windows and construct a similarity matrix; Based on the distribution characteristics of the similarity matrix, a dynamic threshold is set to filter out topic pairs that are higher than the dynamic threshold and generate directed evolution edges. Construct the aforementioned directed evolutionary graph using all themes as nodes and evolutionary edges as directed edges; Based on the analysis of the connection patterns of each node in the evolutionary directed graph, the theme evolution pattern is identified.

7. The method for mining topic evolution paths in battery-related literature as described in claim 6, characterized in that, The thematic evolution patterns include continuation, splitting, merging, emerging, and decay.

8. The method for mining topic evolution paths in battery-related literature as described in claim 6, characterized in that, The upper quartile or mean of all non-zero similarities in the similarity matrix is ​​added to a standard deviation as the dynamic threshold.

9. The method for mining topic evolution paths in battery-related literature as described in claim 1, characterized in that, Indicators for multi-dimensional quantitative assessment of thematic influence include scale indicators, influence indicators, and vitality indicators; The scale indicator is the number of documents within the topic; the influence indicator includes average citation influence and total citation influence; and the vitality indicator is a weighted average of scale vitality and influence vitality.

10. A topic evolution path mining system for battery-related literature, characterized in that, The system is used to implement the topic evolution path mining method for battery-related literature as described in any one of claims 1-9, the system comprising: The document acquisition module is used to acquire a collection of documents related to the field of batteries from a scientific and technological literature database, perform structured representation on each document, and divide the document collection according to a preset time window to obtain a document subset sequence. The topic tag generation module is used to independently perform semantic vectorization processing on the documents within each time window based on the document subset sequence, perform deep semantic topic modeling and representation enhancement, and generate topic tags for each time window through a large language model. The topic influence quantitative assessment module is used to conduct multi-dimensional quantitative assessment of topic influence for topic tags within each time window, based on the scale of literature and citation count within the topic, and generate a comprehensive topic assessment file; The topic evolution pattern recognition module is used to construct a directed graph of topic evolution based on the correlation strength of topic evolution in adjacent time windows, and to identify topic evolution patterns. The structured analysis report generation module is used to encapsulate the topic evolution pattern, topic tags, and topic comprehensive evaluation files into a standard data package and compile them into a structured analysis report.

Images