Cross-domain knowledge theme fusion and evolution analysis method

Abstract

The invention relates to a cross-domain knowledge theme fusion and evolution analysis method. The method comprises the following steps: acquiring data and time slices; subject extraction and semantic representation: independently processing the sub-corpus set of each time slice in each technical field, extracting a subject by adopting a first preset model for the corpus set corresponding to each time slice in different technical fields, and carrying out semantic embedding on the subject by utilizing a second preset model, generating a topic semantic vector of each topic; identifying and classifying shared topics: calculating the fusion degree of the semantic vectors of the topics among different technical fields and the importance of each topic, and dividing the cross-field topic pairs into at least two preset categories according to a first preset fusion degree threshold and a first preset importance threshold; and evolution path construction and analysis: constructing an evolution network of the cross-domain shared topic based on the semantic similarity of the shared topic in the adjacent time slices, and identifying the evolution path type of the shared topic.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more particularly to natural language processing, knowledge discovery and data mining technologies. Specifically, it relates to a cross-domain knowledge topic fusion and evolution analysis method based on semantic modeling and dynamic network analysis. Background Technology

[0002] With the acceleration of global technological innovation, the cross-disciplinary integration of different fields and technologies is increasingly becoming a source of major breakthroughs. Accurately identifying cross-disciplinary knowledge intersections from massive amounts of scientific literature, patent reports, and other textual data, and dynamically tracking their evolution paths, is of crucial strategic significance for predicting technological trends and planning cutting-edge research.

[0003] Existing technical topic analysis methods, such as Latent Dirichlet Allocation (LDA) or Dynamic Topic Model (DTM), primarily perform topic mining and evolutionary analysis within a single domain. These methods are typically based on the bag-of-words model assumption, neglecting the deep semantic connections between words, resulting in insufficient accuracy and interpretability when dealing with interdisciplinary topics whose connotations are constantly evolving.

[0004] In recent years, pre-trained language models, represented by BERT, have made significant progress in capturing contextual semantics. Some studies have attempted to apply them to topic modeling to improve the accuracy of topic representation. However, existing research still has the following limitations: Lack of cross-domain fusion metrics: There is a lack of effective quantitative indicators for the degree of semantic association between topics in different domains, making it difficult to distinguish between truly "shared topics" and superficially "similar topics." Insufficient evolutionary path modeling: Most methods focus on static topic identification or linear evolution within a single domain, lacking a systematic network modeling and analysis framework for the complex dynamic evolutionary paths such as inheritance, differentiation, and fusion between cross-domain topics. Difficulty in predicting potential trends: It is difficult to effectively identify potential hot topics moving from the "periphery" to the "core," making it difficult to provide forward-looking warnings about the explosive potential of emerging interdisciplinary technologies.

[0005] Therefore, there is an urgent need to propose a cross-domain knowledge analysis method that can deeply integrate semantic information, quantify topic associations, and systematically construct dynamic evolution networks to solve the problems of existing technologies in identifying and tracking the integration of cross-disciplinary knowledge. Summary of the Invention

[0006] The purpose of this invention is to provide a cross-domain knowledge topic fusion and evolution analysis method that can effectively quantify the degree of cross-domain topic integration and systematically construct and analyze the dynamic evolution path of shared topics. The technical solution is as follows:

[0007] A method for cross-domain knowledge topic fusion and evolution analysis, characterized by the following steps:

[0008] S1: Data Acquisition and Time Slicing: Acquire text corpus data from at least two different technical fields, and slice the text corpus data according to a preset time unit to form a corpus set corresponding to multiple time segments;

[0009] S2: Topic extraction and semantic representation: Each technical field is processed independently for each time segment of the sub-corpus. For the corpus corresponding to each time segment of different technical fields, the first preset model is used to extract topics, and the second preset model is used to semantically embed the topics to generate topic semantic vectors for each topic.

[0010] S3: Shared topic identification and classification: Calculate the fusion degree of topic semantic vectors between different technical fields and the importance of each topic, and classify cross-field topic pairs into at least two preset categories based on the first preset fusion degree threshold and the first preset importance threshold;

[0011] S4: Evolutionary Path Construction and Analysis: Based on the semantic similarity of shared topics within adjacent time segments, construct an evolutionary network for cross-domain shared topics and identify the evolutionary path types of the shared topics.

[0012] Furthermore, in step S2, the first preset model is a Latent Dirichlet Allocation (LDA) model. The LDA model is used to mine topics in the preprocessed text. By calculating the text perplexity under different topics, the inflection point with the lowest perplexity or tending to converge is selected as the optimal number of topics K for a certain time segment. Each topic consists of a set of keywords and their weights arranged in descending order of probability. The second preset model is a pre-trained language model (BERT) based on the Transformer architecture. The topic semantic vector is obtained by weighted summation of the embedding vectors of keywords within the topic.

[0013] Furthermore, in step S3, the degree of fusion is obtained by calculating the cosine similarity between the semantic vectors of cross-domain topic pairs; the importance is obtained based on the document frequency or word frequency distribution of the topic in the corresponding corpus.

[0014] Furthermore, in step S3, the preset categories include:

[0015] Hot topic sharing: Determined when the degree of integration is greater than the first preset degree of integration threshold and the importance is greater than the first preset importance threshold;

[0016] Edge-shared topic: determined when the degree of fusion is greater than the first preset fusion threshold and the importance is not greater than the first preset importance threshold;

[0017] Hot topic independent theme: determined when the degree of integration is not greater than the first preset degree of integration threshold and the importance is greater than the first preset importance threshold;

[0018] Edge Independent Topic: Determined when the degree of integration is not greater than the first preset degree of integration threshold and the importance is not greater than the first preset importance threshold.

[0019] Furthermore, in step S4, when the cosine similarity between the semantic vectors of two shared topics in adjacent time segments is greater than the second preset evolution threshold, an evolutionary connection relationship is established for the two shared topics in the evolutionary network.

[0020] Furthermore, the evolutionary path types include at least one of the following: persistent path, decay path, nascent path, and mutation path.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] ● Improve the accuracy of topic representation: By combining the topic induction ability of the LDA model and the deep semantic understanding ability of the BERT model, the generated topic semantic vectors can more accurately depict the connotation of the topic, laying a solid foundation for subsequent analysis.

[0023] ● Achieve quantitative assessment of cross-domain integration: Introduce dual indicators of "integration degree" and "importance" and construct a four-quadrant classification model, which can systematically identify cross-domain shared topics at different levels, providing a reliable basis for discovering potential knowledge intersections.

[0024] ● Achieve systematic modeling and visualization of evolutionary paths: By constructing a dynamic evolutionary network across time, it is possible to clearly and intuitively reveal various evolutionary modes of shared topics, such as continuity, emergence, decline, and mutation, and to depict a panoramic view of the integration of interdisciplinary knowledge.

[0025] ●Enhanced the ability to predict technological trends: By analyzing the evolutionary trajectory and type migration of shared hot topics (such as the shift from the periphery to the core), it is possible to effectively identify emerging interdisciplinary research directions with explosive potential. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0027] Figure 1 This is a schematic diagram of the cross-domain corpus time slice structure in an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of a shared knowledge topic matrix in an embodiment of the present invention.

[0029] Figure 3This is a schematic diagram of a cross-domain knowledge topic classification model in an embodiment of the present invention.

[0030] Figure 4 This is a schematic diagram illustrating the construction of the CDT evolution path relationship in an embodiment of the present invention.

[0031] Figure 5 This is a schematic diagram of the topic evolution path type in an embodiment of the present invention.

[0032] Figure 6 This is a graph showing the results of the topic quantity and confusion assessment in the fields of "large language model" and "innovative design" in various time segments in the application case.

[0033] Figure 7 This is a distribution chart showing the integration and importance of cross-domain sharing topics in application cases in 2024.

[0034] Figure 8 This is a schematic diagram illustrating the classification results of cross-domain sharing topics in the application case in 2024.

[0035] Figure 9 This is the overall evolution network diagram of the cross-domain sharing theme from 2018 to 2025 in the application case. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0037] This embodiment provides a method for cross-domain knowledge topic fusion and evolution analysis, the specific process of which includes the following steps:

[0038] Step S101: Obtain data and time slices.

[0039] Choose at least two target technology areas, such as Area A (e.g., "Natural Language Processing") and Area B (e.g., "Educational Technology"). Using web crawlers or database APIs, obtain relevant literature or patent data within a specified time span (e.g., 2015 to 2025) from sources such as Web of Science, CNKI, and patent databases (e.g., Derwent). Extract the title, abstract, keywords, and other textual information from each data entry to construct the original corpora for Area A and Area B, respectively.

[0040] Subsequently, using "year" as the unit of time, the entire time span was divided into multiple consecutive time segments (T1, T2, ..., Tn). The corpus for each domain was then segmented according to these time segments, forming a series of timestamped subcorpora, such as... Figure 1 As shown.

[0041] Step S102: Topic extraction and semantic representation.

[0042] Each domain's sub-corpus is processed independently for each time segment. For example... Figure 3 As shown, it specifically includes:

[0043] (1) Topic Extraction: The Latent Dirichlet Allocation (LDA) model was used to mine topics from the preprocessed text (e.g., word segmentation and stop word removal). By calculating the perplexity of the model under different topics, the inflection point with the lowest perplexity or the point that tends to converge was selected as the optimal number of topics K for that time segment. Each topic consists of a set of keywords and their weights arranged in descending order of probability.

[0044]

[0045] Among them, Perplexity z P(topic) represents the text perplexity of a given year z, M represents the total number of documents in that year, Nm is the number of words in document m, and K is the number of predefined topics. i |m) represents the topic i The probability of a word appearing in document m, where wm,n refers to the nth word in document m, and P(wm,n|topic) is the probability of the word appearing in document m. i () refers to the words wm,n in the topic i The probability of it appearing in the middle.

[0046] (2) Semantic embedding: A pre-trained BERT model is used as the encoder. The first N (e.g., N=15) keywords and their weights extracted from each topic by LDA are input into the BERT model to obtain the context embedding vector of each keyword.

[0047] (3) Vector dimensionality reduction and aggregation: To reduce computational complexity and noise impact, dimensionality reduction methods such as principal component analysis (PCA) can be used to reduce the high-dimensional word embedding vectors (e.g., 768-dimensional) to a lower dimension (e.g., 64-dimensional). Finally, based on the original weights of each keyword in the topic, the dimensionality-reduced word vectors are weighted and summed to obtain a unique topic semantic vector representing the core semantics of the entire topic.

[0048] Each topic *topici* contains a set of word weights *ωj* and corresponding words *wordj*. BERT maps each word *wordj* to a semantically rich embedding vector *aj*. The topic vectors for domains A and B are constructed by calculating a weighted sum of these word embedding vectors. For example, the i-th topic vector for domain A in time period 1 is calculated as follows:

[0049]

[0050] Step S103: Shared topic identification and classification.

[0051] A cross-disciplinary research topic (CDT) refers to a novel research topic composed of thematic elements from two different fields. The relationship between these topics can be represented by a cross-disciplinary research topic matrix, such as... Figure 2 As shown in the figure, CDT(1,2) contains the content of Topic 1 in research field A and Topic 2 in research field B, serving as a new topic identifier for cross-disciplinary research.

[0052] To better understand the characteristics of each topic in the CDT matrix, this invention further introduces a calculation method for evaluating the degree of CDT integration and importance.

[0053] (1) Calculate the degree of fusion and importance:

[0054] Similarity (Sim): For any topic i in domain A and any topic j in domain B within the same time segment, their similarity Sim(i,j) is obtained by calculating the cosine similarity between their topic semantic vectors V_i and V_j. The higher this value, the stronger the semantic connection between the two topics.

[0055] Suppose domain A has P topics and domain B has Q topics, and these topics are represented by topic vectors. p and topic q Let p represent a topic in domain A and q represent a topic in domain B. Then the integration degree of CDT(p,q) is calculated as follows:

[0056]

[0057] Popularity (Pop): For each topic, its importance can be measured by the frequency of its distribution in the corresponding corpus, i.e., how many documents contain the main topic. This value reflects the research popularity or attention given to the topic within its own field.

[0058] (2) Four-quadrant classification:

[0059] Set a fusion threshold θ_sim (e.g., 0.5) and an importance threshold θ_pop (e.g., 0.5). For example... Figure 3 As shown, based on the integration degree Sim(i,j) of each cross-domain topic pair (i,j) and their respective importance Pop(i) and Pop(j) (which can be averaged or considered comprehensively), they are divided into four categories:

[0060] ● Hotspot Shared Topics (HST): Sim > θ_sim and Pop > θ_pop. This represents a core hotspot of common interest in both fields.

[0061] ● Edge Shared Topic (MST): Sim > θ_sim and Pop ≤ θ_pop. Represents a potential, emerging, or niche intersection between two fields.

[0062] ●Hot Independent Topics (HIT): Sim ≤ θ_sim and Pop > θ_pop. These represent popular research areas within each discipline, but which have not yet formed close intersections.

[0063] ● Marginal Independent Topics (MIT): Sim≤θ_sim and Pop≤θ_pop. These represent relatively niche or independent topics within a given field.

[0064] Step S104: Evolutionary path construction and analysis.

[0065] (1) Constructing evolutionary networks: such as Figure 4 As shown, for two adjacent time segments T_n and T_{n+1}, the semantic similarity between any shared topic pair C(i,j) in T_n and any shared topic pair C(p,q) in T_{n+1} is calculated. This semantic similarity can be calculated by combining the similarity of the corresponding domain topic vectors in C(i,j) and C(p,q). If the similarity is higher than a preset evolution threshold θ_rel (e.g., 0.5), then C(p,q) is considered to be the semantic inheritor of C(i,j), and a directed edge from C(i,j) to C(p,q) is established in the evolutionary network.

[0066] (2) Identify evolutionary path types:

[0067] ●Continuous path: The topic has a unique, strongly related successor in the later time segment.

[0068] ● Decline path: The theme has no successor in the later time segment, indicating that its popularity has declined or it has been replaced by new technologies.

[0069] ● Emerging Path: The topic appears in the current time segment, but has no clear source in the previous time segment, representing the birth of a new research direction.

[0070] ●Mutation path: Themes diverge (one parent theme evolves into multiple sub-themes) or merge (multiple parent themes merge into one sub-theme).

[0071] like Figure 5 As shown, by analyzing the path types, lifecycles, and type migrations (such as the evolution of MST into HST) of HST and MST in the evolutionary network, key technology convergence trajectories and cutting-edge directions with future explosive potential can be effectively identified.

[0072] Based on the above method, this embodiment takes the cross-integration of the two fields of "Large Language Model (LLM)" and "Innovative Design (ID)" as an example to specifically illustrate the application process of the method of the present invention.

[0073] 1. Data Acquisition and Processing

[0074] We collected 30,929 articles related to "Large Language Models" and 74,314 articles related to "Innovation Design" from the Web of Science database between January 2018 and May 2025. The period from 2018 to 2025 was divided into eight time segments, using "year" as the unit.

[0075] 2. Cross-domain topic mining and modeling

[0076] LDA topic modeling was performed on the corpora from the two domains across eight time segments. Based on... Figure 6 The perplexity assessment results shown indicate the optimal number of topics for each time segment. Based on the perplexity calculation, the optimal number of topics (K) for the Large Language Model (LLM) domain in each time segment is set to 9, 7, 9, 9, 5, 9, 9, and 9, respectively. For the Innovative Design (ID) domain, the optimal K values ​​are determined to be 6, 7, 9, 6, 9, 8, 9, and 9. Taking 2024 as an example, 9 topics were extracted from the LLM domain and 9 topics were extracted from the ID domain. The topic knowledge content is shown in Tables 1 and 2.

[0077] Table 1. Core Research Themes of “LLM” in 2024

[0078]

[0079]

[0080] Table 2. Core Research Themes in the “ID” Field in 2024

[0081]

[0082]

[0083] The semantic vectors of each topic are obtained using the method described in step S102. Then, according to step S103, the fusion degree and importance of cross-domain topic pairs are calculated, and thresholds θ_sim = 0.5 and θ_pop = 0.5 are set for classification. The calculation results of cross-domain topic importance and fusion degree in 2024 are as follows: Figure 7 As shown in the figure. The analysis found that since the release of GPT-3 in 2020, the integration of topics such as "knowledge training" in the LLM field with research in the ID field has significantly increased.

[0084] like Figure 8 As shown, in the 2024 classification results, three hot shared topics (HSTs) were identified, such as "fusion of semantic understanding and material optimization", "intelligent algorithm and control design based on pre-trained model", and "text-driven knowledge modeling platform for digital construction", all of which are closely related to the topic of "text knowledge extraction and model building" in the LLM field.

[0085] 3. Construction and Analysis of Cross-Domain Shared Topic Evolution Path

[0086] Setting the evolution threshold θ_rel = 0.5, a global shared-topic evolution network was constructed for the period 2018-2025, such as... Figure 9 As shown.

[0087] Path Tracing Analysis 1: Information-Driven Product Innovation Design Framework

[0088] We take the 2018 HST topic CDT(5,6) "Industrial Product Design Based on Natural Language Processing" as a starting point. In 2019, influenced by the BERT model, this topic diverged into several directions. From 2020 to 2022, driven by the GPT-2 / 3 model, numerous mutation paths emerged regarding collaborative design tools. By 2024-2025, with the maturation of multimodal LLM and Agent technologies, more complex application topics such as "Digital Construction Knowledge Platform" and "Sustainable Technology Collaborative Development Platform" have evolved. This path clearly demonstrates the transformation of LLM's role in innovative design from "auxiliary tool" to "design engine" and then to "core intelligent agent."

[0089] Path Tracing Analysis II: Integration of Large-Scale Models in the Context of Digital Enterprises

[0090] A retrospective analysis was conducted, starting with the 2025 HST topic CDT(6,2), "Knowledge Construction and Management Methods for Complex Systems." The study found that its main origins can be traced back to a batch of "intelligent service" related topics that emerged in 2022 due to the release of InstructGPT, such as "user behavior recognition" and "dynamic response platform construction." These emerging topics incorporated the concept of "digital enterprise" in 2023 and converged with earlier topics such as "content generation" driven by the Transformer architecture, ultimately forming a stable and continuously evolving mainstream direction.

[0091] This case demonstrates that the method of the present invention can effectively identify key technology integration points and clearly and systematically trace their evolution, which has high practical value for predicting technology development trends.

[0092] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for cross-domain knowledge topic fusion and evolution analysis, characterized in that, Includes the following steps: S1: Data Acquisition and Time Slicing: Acquire text corpus data from at least two different technical fields, and slice the text corpus data according to a preset time unit to form a corpus set corresponding to multiple time segments; S2: Topic extraction and semantic representation: Each technical field is processed independently for each time segment of the sub-corpus. For the corpus corresponding to each time segment of different technical fields, the first preset model is used to extract topics, and the second preset model is used to semantically embed the topics to generate topic semantic vectors for each topic. S3: Shared topic identification and classification: Calculate the fusion degree of topic semantic vectors between different technical fields and the importance of each topic, and classify cross-field topic pairs into at least two preset categories based on the first preset fusion degree threshold and the first preset importance threshold; S4: Evolutionary Path Construction and Analysis: Based on the semantic similarity of shared topics within adjacent time segments, construct an evolutionary network for cross-domain shared topics and identify the evolutionary path types of the shared topics.

2. The method according to claim 1, characterized in that, In step S2, the first preset model is a Latent Dirichlet Allocation (LDA) model. The LDA model is used to mine topics in the preprocessed text. By calculating the text perplexity under different topics, the inflection point with the lowest perplexity or tending to converge is selected as the optimal number of topics K for a certain time segment. Each topic consists of a set of keywords and their weights arranged in descending order of probability. The second preset model is a pre-trained language model (BERT) based on the Transformer architecture. The topic semantic vector is obtained by weighted summation of the embedding vectors of keywords within the topic.

3. The method according to claim 1, characterized in that, In step S3, the degree of fusion is obtained by calculating the cosine similarity between the semantic vectors of cross-domain topic pairs; the importance is obtained based on the document frequency or word frequency distribution of the topic in the corresponding corpus.

4. The method according to claim 3, characterized in that, In step S3, the preset categories include: Hot topic sharing: Determined when the degree of integration is greater than the first preset degree of integration threshold and the importance is greater than the first preset importance threshold; Edge-shared topic: determined when the degree of fusion is greater than the first preset fusion threshold and the importance is not greater than the first preset importance threshold; Hot topic independent theme: determined when the degree of integration is not greater than the first preset degree of integration threshold and the importance is greater than the first preset importance threshold; Edge Independent Topic: Determined when the degree of integration is not greater than the first preset degree of integration threshold and the importance is not greater than the first preset importance threshold.

5. The method according to claim 1, characterized in that, In step S4, when the cosine similarity between the semantic vectors of two shared topics in adjacent time segments is greater than the second preset evolution threshold, an evolutionary connection relationship is established for the two shared topics in the evolutionary network.

6. The method according to claim 1, characterized in that, The evolutionary path types include at least one of the following: persistent path, decay path, nascent path, and mutation path.

Images