Interdisciplinary knowledge unit citation network construction and analysis system for academic literature

Abstract

The application discloses a kind of academic literature-oriented interdisciplinary knowledge unit citation network construction and analysis system, belong to knowledge graph construction technical field, its system includes: data acquisition and pretreatment module, obtains target field academic literature dataset and pretreatment, knowledge unit and semantic function identification module, the data after pretreatment are carried out semantic function identification, interdisciplinary knowledge unit citation network construction module, semantic identification result and subject classification information are formed into composite node, with knowledge unit citation relationship as edge, constructs interdisciplinary knowledge unit citation network, and extracts its subnet;Interdisciplinary knowledge unit citation network analysis module, the interdisciplinary knowledge unit citation network and its subnet obtained by construction are analyzed to obtain target field interdisciplinary knowledge citation structure, core knowledge unit, typical citation mode, knowledge colony structure and its evolution trend;The application improves the accuracy of interdisciplinary knowledge fusion path identification and evolution law analysis.

Description

Technical Field

[0001] This invention belongs to the fields of knowledge graph construction, scientific and technological intelligence analysis, semantic recognition of academic documents and interdisciplinary knowledge network analysis, specifically involving a system for constructing and analyzing interdisciplinary knowledge unit citation networks for academic documents. Background Technology

[0002] As scientific research questions become increasingly complex, interdisciplinary knowledge integration has become an important means of scientific innovation. Citations in academic literature are a crucial vehicle for knowledge flow; they not only indicate connections between documents but also implicitly reveal researchers' use, expansion, and comparison of external theories, methods, data, or concepts. Therefore, identifying specific knowledge content from academic literature, characterizing the citation relationships between knowledge content, and further analyzing interdisciplinary knowledge citation paths and patterns have become important technical issues in scientific intelligence analysis, knowledge graph construction, and interdisciplinary research evaluation.

[0003] Existing interdisciplinary knowledge analysis methods mostly use documents, authors, institutions, journals, keywords, topics, or subject categories as basic analytical objects. For example, citation analysis methods can count the number and direction of citations between documents from different disciplines, topic modeling methods can extract research topics, keyword co-occurrence methods can analyze the relationships between research hotspots, and knowledge graph methods can extract research entities and construct entity relationship networks. These methods can reflect the connections between disciplines at a macro level, but the analytical granularity usually remains at the document level, topic level, or keyword level, making it difficult to explain the specific knowledge content cited in interdisciplinary citations and its semantic role in the research.

[0004] Among existing related patents, patent application CN111241293A discloses a knowledge graph algorithm based on academic literature construction; patent application CN120030172A discloses a method, apparatus, device, and medium for constructing knowledge graphs of scientific and technological literature, mainly targeting the construction of knowledge graphs for academic or scientific and technological literature; patent application CN119830157A discloses a citation intent classification method based on dynamic heterogeneous graph fusion network, mainly utilizing the context of the cited sentence and chapter information for citation intent classification. Although the above solutions involve knowledge graph construction and citation intent recognition respectively, they still have shortcomings such as coarse analysis granularity, high dependence on large-scale training sets for specific domain entity recognition, and lack of interdisciplinary dimensions in knowledge unit network analysis. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention aims to provide a system for constructing and analyzing interdisciplinary knowledge unit citation networks for academic literature. Addressing the problems of coarse-grained analysis, heavy reliance on large training sets for entity recognition of knowledge units in specialized fields, and a lack of interdisciplinary dimensions in knowledge unit network analysis, this invention employs a recognition framework that integrates retrieval-enhanced generation technology with a large language model. Under limited sample conditions, it improves the accuracy and stability of entity recognition of knowledge units, lexical function classification, and citation function category classification in academic literature through multi-round validation optimization. Based on this, it constructs an interdisciplinary multi-semantic knowledge unit citation network, refining the interdisciplinary knowledge analysis object from documents, topics, or keywords to the knowledge unit level. Furthermore, through network structure, interdisciplinary influence of knowledge units, interdisciplinary citation patterns of knowledge units, interdisciplinary knowledge communities, and dynamic evolution analysis, it reveals the interdisciplinary knowledge fusion path and its evolutionary laws. This invention can achieve finer-grained, more semantically interpretable, and more interdisciplinary targeted knowledge citation analysis with lower manual annotation costs.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A system for constructing and analyzing interdisciplinary knowledge unit citation networks for academic literature includes a data acquisition and preprocessing module, a knowledge unit and semantic function recognition module, an interdisciplinary knowledge unit citation network construction module, and an interdisciplinary knowledge unit citation network analysis module. The data acquisition and preprocessing module is used to acquire and preprocess the full-text dataset of academic literature in the target field, and generate a dataset of cited bibliographic records, a dataset of citation context, and a dataset of literature citation relationships. The knowledge unit and semantic function recognition module is used to perform semantic function recognition on the preprocessed data, and to identify the citing knowledge unit and its lexical function based on the citing bibliographic dataset, and to identify the cited knowledge unit and its lexical function and citation function based on the citation context dataset. The interdisciplinary knowledge unit citation network construction module is used to combine semantic recognition results and subject classification information into composite nodes, extract knowledge unit citation relationships from the literature citation relationship dataset as edges, construct an interdisciplinary knowledge unit citation network, and then extract interdisciplinary knowledge unit citation subnets based on the subject attributes of the citing knowledge unit and the cited knowledge unit. The interdisciplinary knowledge unit reference network analysis module performs network topology analysis, knowledge unit interdisciplinary influence analysis, knowledge unit interdisciplinary reference pattern analysis, interdisciplinary knowledge community identification, and interdisciplinary knowledge community dynamic evolution analysis on the constructed interdisciplinary knowledge unit reference network and interdisciplinary knowledge unit reference subnets, thereby obtaining the target domain's interdisciplinary knowledge reference structure, core knowledge units, typical reference patterns, knowledge community structure, and its evolution trend.

[0007] The full-text dataset of academic literature in the target field includes bibliographic information, text, and references cited in the literature.

[0008] The preprocessing includes structured parsing of the full-text dataset of academic literature in the target field and two data cleaning processes. The structured parsing includes extracting title, abstract, and keyword fields from the cited references to form a cited reference bibliographic dataset, and extracting citation sentences with reference identifiers from the text of the cited references, and recording the reference number, chapter, publication year, and context information corresponding to the citation sentences to form a citation context dataset. The first data cleaning process removes records that are missing, have abnormal formatting, or whose reference identifiers cannot be parsed from the full-text data of academic literature in the target field; The second data cleaning process removes records with missing fields, invalid reference identifiers, duplicate citation records, and data from the citation bibliographic dataset, citation context dataset, and bibliographic citation relationship dataset that cannot establish a correspondence between citations and cited knowledge units.

[0009] The specific steps for recognizing the knowledge unit and semantic function are as follows: The bibliographical examples of the citing knowledge units and their lexical functions, as well as the citation sentence examples of the citation function, the cited knowledge units and their lexical functions, are manually annotated, and the manually annotated examples are divided into training set, validation set and test set. A knowledge base of manually annotated examples is built based on the training set and combined with an existing domain terminology knowledge base for vectorized storage; The domain terminology knowledge base is used to store the professional terms and definitions of the target domain. When new citation data or citation context data is input, the system retrieves the top-ranked references with the highest cosine similarity from both the manually annotated sample knowledge base and the domain terminology knowledge base. Each set of examples and terms is given, and then reordered, retaining the top-ranked terms with the highest reorder scores. Examples and terms, ; The newly input citation bibliographic data, the retained examples and terms and their definitions, the judgment rules and the output format are all filled into the structured prompt word template, and then input into the large language model for recognition to obtain structured fields of citing knowledge units and their lexical functions, cited knowledge units and their lexical functions and citation functions.

[0010] The knowledge unit and semantic function recognition module also employs accuracy. Recall rate and The evaluation criteria include the citing knowledge unit and its lexical function, the cited knowledge unit and its lexical function, and the citation function, among which precision is considered. Recall represents the proportion of correctly identified results from a large language model among all identified results. This represents the proportion of correct recognition results from the large language model compared to the recognition results on the validation or test set of manually labeled data. The value represents the accuracy. and recall rate The harmonic mean is calculated using the formulas shown in (1), (2), and (3): (1) (2) (3) in, This represents the number of citing knowledge units and their lexical functions, as well as the number of cited knowledge units and their lexical functions, correctly identified by the large language model. This represents the number of citing knowledge units and their lexical functions, as well as the number of cited knowledge units and their lexical functions, that were incorrectly identified or redundantly generated by the large language model. This represents the number of citing knowledge units and their lexical functions, and cited knowledge units and their lexical functions that the large language model failed to identify. ROUGE-1, ROUGE-2, ROUGE-L, and BERTScore are used to evaluate the lexical character similarity between cited and citing knowledge units (CKIs) identified by the large language model and those manually annotated. Specifically, ROUGE-1 evaluates the unarylexical overlap between CKIs identified by the large language model and those manually annotated, calculated based on term overlap or longest common subsequence at different granularities; ROUGE-2 evaluates the bigram overlap between CKIs identified by the large language model and those manually annotated; and ROUGE-L evaluates the lexical character similarity between CKIs identified by the large language model and those manually annotated. The longest common subsequence matching between the cited knowledge units and citing knowledge units identified by the large language model and the manually annotated cited knowledge units and citing knowledge units is used to measure the similarity between the cited knowledge units and citing knowledge units identified by the large language model and the manually annotated cited knowledge units and citing knowledge units at the lexical character level. BERTScore, based on the context semantic vector generated by the BERT model, evaluates the semantic similarity between the cited knowledge units and citing knowledge units identified by the large language model and the manually annotated cited knowledge units and citing knowledge units by calculating the semantic similarity between them. The calculation formulas are shown in (4), (5) and (6): (4) (5) (6) in, This represents the set of lexical units obtained after word segmentation of the citing and cited knowledge units identified by the large language model. This represents the set of lexical units obtained after word segmentation of manually annotated citing and cited knowledge units. Represents a set The first in The context semantic vector corresponding to each word element Represents a set The first in The context semantic vector corresponding to each word element; Represents a set The number of lexical units in the text Represents a set The number of lexical units in the text; Representing vectors with vector Cosine similarity between them; The semantic vectors in the citing and cited knowledge units recognized by the large language model. The maximum semantic similarity obtained by matching the semantic vectors in the manually annotated citing and cited knowledge units; This represents the semantic vectors in manually annotated citing and cited knowledge units. The maximum semantic similarity obtained by matching the semantic vectors of the citing and cited knowledge units identified by the large language model; This represents the semantic recall rate calculated based on the citing and cited knowledge units output by the large language model. This represents the semantic precision calculated based on manually annotated citing and cited knowledge units. express and The harmonic mean; Based on accuracy Recall rate The evaluation metrics include F1 score, ROUGE-1, ROUGE-2, ROUGE-L, and BERTScore. The large language model's recognition results are compared with manually annotated results to identify error types, including entity boundary errors, category misclassification, missed recognitions, multiple recognitions, and output format errors. The structured prompt word template is corrected based on the identified error types. The manually annotated bibliographic examples and citation examples used for verification in this round are added to the manually annotated example knowledge base for the next iteration. This process continues until the accuracy of the large language model's recognition results compared with the manually annotated results is achieved. Recall rate If the F1 score, ROUGE-1, ROUGE-2, ROUGE-L and BERTScore evaluation metrics no longer improve, then the optimization round that reaches the highest evaluation metric is determined as the optimal round. The structured prompt word version of the optimal round, the supplemented manually annotated sample knowledge base, the original domain terminology knowledge base and the large language model are combined to form the final knowledge unit and semantic function recognition model.

[0011] The network model of the interdisciplinary knowledge unit referencing network is defined as a directed weighted graph. ,in, This represents the set of citing knowledge unit nodes and the set of cited knowledge unit nodes. This represents the set of directed edges formed between knowledge units extracted from the literature citation relation dataset based on citation relationships. Represents the set of edge weights. Represents the set of node attributes and edge attributes; The citing or cited knowledge unit node includes the entity name of the citing or cited knowledge unit and its lexical function classification code and subject classification code; the citing knowledge unit node serves as the source node, and the cited knowledge unit node serves as the target node; the lexical function classification code is pre-set according to the lexical function of the citing or cited knowledge unit in scientific research, and includes the research question code, research method code, research method code, data code, and theory or concept code in the citing knowledge unit; the subject classification code is set according to the second-level subject field in the OpenAlex subject classification system, and is used to identify the subject field to which the knowledge unit belongs, including the target subject code and the external subject classification code; the entity name of the citing or cited knowledge unit and its lexical function classification code and subject classification code are concatenated to form a node label.

[0012] The extraction process of the interdisciplinary knowledge unit reference subnet is as follows: Traverse all directed edges in the interdisciplinary knowledge unit reference network, and determine whether they belong to interdisciplinary reference relationships based on the subject classification codes of the source node and the target node. If the source node belongs to the target subject and the target node belongs to a non-target subject, then retain the directed edge and its two endpoints. If the edge represents a reference within the target subject or a reference between external subjects, then remove it. After filtering, delete isolated nodes to form an interdisciplinary knowledge unit reference subnet.

[0013] The network topology analysis analyzes the connectivity, sparsity, clustering, and structural distribution characteristics of the interdisciplinary knowledge unit referencing network and its subnets by analyzing the number of nodes, edges, average degree, average weighted degree, network density, network diameter, average path length, connected components, and modularity. This analysis is used to characterize the referencing structure of the interdisciplinary knowledge unit referencing network. Simultaneously, a network projection method is employed to construct the referencing node projection network and the cited knowledge unit node projection network of the interdisciplinary knowledge unit referencing subnets. The referencing node projection network indicates whether different referencing knowledge units jointly refer to the same cited knowledge unit, while the cited knowledge unit node projection network indicates whether different cited knowledge units are jointly referenced by the same referencing knowledge unit. By calculating the projection network density or average clustering coefficient, the degree of overlap in knowledge needs among referencing knowledge units and the co-reference relationships among cited knowledge units are analyzed, further characterizing the local clustering features of the interdisciplinary knowledge unit referencing subnets. The interdisciplinary influence analysis of knowledge units is used to identify core cited knowledge units and active citing knowledge units in the interdisciplinary knowledge unit citation network. The activity and coreness of nodes are measured by the in-degree, out-degree, weighted in-degree, and weighted out-degree of the nodes and directed edges in the interdisciplinary knowledge unit citation network. Specifically, in-degree represents the number of directed edges pointing to a node, used to measure the number of times a cited knowledge unit node is cited by different citing knowledge unit nodes; out-degree represents the number of directed edges from a node to other nodes, used to measure the activity level of citing cited knowledge unit nodes from external disciplines by citing knowledge unit nodes; weighted in-degree represents the sum of the weights of directed edges pointing to a node, used to measure the cumulative intensity of citations of a cited knowledge unit node; and weighted out-degree represents the cumulative intensity of citations of a node by a given node. The sum of the weights of directed edges pointing from a node to other nodes is used to measure the cumulative citation strength of the citing knowledge unit node to external subject knowledge. The weights of the directed edges are obtained by aggregating the citation frequencies of the same citing knowledge unit, the same cited knowledge unit, and the same citation function category. A webpage ranking algorithm is used to calculate the importance of the citing or cited knowledge unit node in the network, and the core knowledge units and basic theories, key methods, or important data on the core path of knowledge flow are judged based on the importance of the nodes. At the same time, a hyperlink-induced topic search algorithm is used to calculate the authority value and hub value of the citing knowledge unit node. The authority value is used to reflect the degree of recognition of the cited knowledge unit node as an external knowledge source, and the hub value is used to reflect the ability of the citing knowledge unit node to absorb and organize external knowledge. The cross-disciplinary citation pattern analysis of knowledge units is used to reveal the citation combination rules between knowledge units of different lexical function categories. Based on the lexical function categories of the citing and cited knowledge units, citation relationships are divided into six typical citation patterns: research problem citing research method, research problem citing data, research problem citing theory or concept, research method citing research method, research method citing data, and research method citing theory or concept. Each directed edge in the cross-disciplinary knowledge unit citation subnet is used as a statistical object. The lexical function categories of the source node and the target node of the directed edge are read and combined into a citation pattern. Based on each citation pattern, the number of unique patterns, the total number of citations, and the average citation intensity are calculated. The number of unique patterns is the number of combinations of non-repeating citing or cited knowledge unit entity names under this pattern; the total number of citations is the sum of the edge weights of all directed edges under this pattern; and the average citation intensity is the ratio of the total number of citations to the number of unique patterns. The interdisciplinary knowledge community identification method is used to discover sets of citing or cited knowledge units that are closely connected and have strong semantic relationships in the interdisciplinary knowledge unit citation network. A community discovery algorithm is used to divide the interdisciplinary knowledge unit citation network into multiple knowledge communities. The citing knowledge units with the highest out-degree, weighted out-degree, PageRank value, or hub value within each knowledge community are identified as core citing knowledge unit nodes. The cited knowledge units with the highest in-degree, weighted in-degree, PageRank value, or authority value within each knowledge community are identified as core cited knowledge unit nodes. The number of directed edges or the sum of edge weights corresponding to various citation function categories within a knowledge community are counted, and the citation function category with the highest proportion is identified as the main citation function of that community. The number of directed edges or the sum of edge weights corresponding to various chapter positions within a knowledge community are counted, and the chapter position with the highest proportion is identified as the main chapter position of that community. Subsequently, the theme and interdisciplinary integration type of the knowledge community are interpreted by combining the core citing knowledge unit nodes, core cited knowledge unit nodes, main citation functions, and main chapter positions. The interdisciplinary knowledge community dynamic evolution analysis is used to characterize the changing trends of knowledge integration patterns from a time dimension. This analysis constructs a dynamic monitoring framework based on a fixed community structure: using the core knowledge communities identified in network analysis as a benchmark, and according to the citation time attributes in the literature citation relationship dataset, the interdisciplinary knowledge unit citation network is divided into different time slices; by statistically analyzing the changes in citation intensity within each knowledge community under different time slices, activity evolution curves of different interdisciplinary integration patterns are plotted, thereby capturing the macro-migration path of the research focus of the field; then, the annual replacement of core cited knowledge units within each knowledge community is tracked, and the specific content and direction of knowledge unit flow are analyzed to reveal how an external knowledge unit gradually evolves from peripheral exploratory citations into a core infrastructure, or how a theoretical concept is reactivated and explosively grows under new technological conditions.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Improve the fine-grainedness and semantic interpretability of interdisciplinary knowledge analysis Existing interdisciplinary analysis methods often use documents, keywords, themes, or subject categories as analysis objects, making it difficult to delve into the specific knowledge content level. This invention, through knowledge unit entity recognition, models the research questions and methods in the cited documents, as well as the research methods, data, theories, or concepts in the cited sentences and their context, as network nodes. This refines the interdisciplinary knowledge analysis object from the document level and theme level to the knowledge unit level. Therefore, this invention can more accurately describe what specific knowledge content is cited in interdisciplinary citations and the lexical function of that knowledge content in scientific research.

[0015] (2) Reduce the dependence of professional domain knowledge entity recognition on large-scale labeled data Existing professional entity recognition methods typically require large-scale manually annotated corpora, resulting in high labor costs. Furthermore, they are prone to instability in entity boundary recognition within academic texts with complex terminology, numerous abbreviations, and diverse citation contexts. This invention employs a recognition framework that integrates retrieval-enhanced generation technology with a large language model. It utilizes a small-scale manually annotated sample knowledge base and a domain terminology knowledge base as external knowledge sources. Through similar sample retrieval, related term retrieval, structured prompts, and multi-round validation optimization, it improves the recognition capability of the large language model under limited sample conditions. Therefore, this invention can improve the accuracy and stability of knowledge unit entity recognition, lexical function category discrimination, and citation function category discrimination while reducing reliance on large-scale manual annotation.

[0016] (3) Enhance the ability of citation networks to express interdisciplinary knowledge citation paths. Traditional citation networks typically only record whether citation relationships exist between documents or the number of citations, making it difficult to express the semantic connections between specific knowledge units. This invention uses the entity name of the knowledge unit, the lexical function category code, and the subject classification code to form composite nodes. The citation behavior of a citing knowledge unit pointing to a cited knowledge unit is represented as a directed weighted edge with attributes of citation function, citation time, document position, and citation frequency. Therefore, this invention can not only indicate whether a citation has occurred, but also which cited knowledge unit the citing knowledge unit cited, at what research stage the citation occurred, and whether the citation function is usage, extension, or comparison, thereby improving the interpretability of interdisciplinary knowledge citation paths.

[0017] (4) Able to distinguish between the knowledge structure within a domain and the knowledge reference path across disciplines. Existing knowledge graph or topic evolution methods often struggle to accurately distinguish between internal domain knowledge references, external discipline knowledge inputs, and irrelevant references between external disciplines. This invention first constructs a cross-disciplinary overall knowledge unit reference network encompassing both internal and external knowledge references within the target discipline. Then, based on the node discipline classification coding, it filters the reference relationships between target discipline-invoking knowledge units and non-target discipline-invoking knowledge units, forming a cross-disciplinary knowledge unit reference subnetwork. Therefore, this invention can both preserve the overall view of the target domain's knowledge structure and focus on analyzing the specific paths through which external discipline knowledge flows into the target discipline.

[0018] (5) Form a complete technical process from semantic recognition to knowledge reference pattern analysis. Existing technologies typically perform entity extraction, topic analysis, citation statistics, or network visualization separately, resulting in a relatively fragmented analysis process. This invention integrates data collection and preprocessing, knowledge unit and semantic function identification, interdisciplinary knowledge unit citation network construction, and interdisciplinary knowledge unit citation network analysis into a continuous process. It outputs results through network topology analysis, interdisciplinary influence analysis of knowledge units, interdisciplinary citation pattern analysis of knowledge units, interdisciplinary knowledge community identification, and dynamic evolution analysis of interdisciplinary knowledge communities. Therefore, this invention can systematically reveal the interdisciplinary knowledge fusion structure, core cited knowledge units and citing knowledge units, typical citation patterns, knowledge communities, and their evolutionary trends, providing more granular data support for scientific and technological intelligence analysis, interdisciplinary evaluation, and scientific research decision-making.

[0019] In summary, this invention can identify knowledge units and semantic functions in academic literature under limited sample conditions, and transform the identification results into a computable, analyzable, and interpretable interdisciplinary knowledge unit citation network, thereby improving the accuracy of interdisciplinary knowledge fusion path identification and evolutionary law analysis. Attached Figure Description

[0020] Figure 1 This is an overall architecture diagram of the interdisciplinary knowledge unit reference network construction and analysis system of the present invention.

[0021] Figure 2 This is an architecture diagram of the knowledge unit and semantic function recognition module of the present invention.

[0022] Figure 3 This is a structured prompt word template for the cited knowledge unit and its lexical function recognition in this invention.

[0023] Figure 4 This is an example of constructing an interdisciplinary knowledge unit reference network according to the present invention.

[0024] Figure 5 This invention uses a cross-disciplinary knowledge unit reference network analysis framework.

[0025] Figure 6 This is a sub-network diagram for the interdisciplinary knowledge unit reference of this invention.

[0026] Figure 7 This is a statistical analysis of the number of cross-disciplinary knowledge unit citation patterns in this invention. Detailed Implementation

[0027] The present invention will now be described in detail with reference to the accompanying drawings.

[0028] The overall architecture of the interdisciplinary knowledge unit citation network construction and analysis system for academic literature proposed in this invention is as follows: Figure 1As shown, it includes a data acquisition and preprocessing module, a knowledge unit and semantic function recognition module, an interdisciplinary knowledge unit reference network construction module, and an interdisciplinary knowledge unit reference network analysis module. Figure 1 The document sequentially illustrates the preprocessing of the full-text dataset of the target domain literature, the generation of the cited bibliographic dataset, the citation context dataset, and the literature citation relationship dataset, the acquisition of knowledge unit and semantic function recognition results, and the relationship between the interdisciplinary knowledge unit citation network, the interdisciplinary knowledge unit citation subnetwork, and the subsequent network analysis results. The four modules are executed in the order of data preprocessing, semantic recognition, network construction, and network analysis.

[0029] The data acquisition and preprocessing module obtains a full-text dataset of academic literature in the target field, performs structured parsing and data cleaning on the full-text data, and generates a citation bibliographic dataset, a citation context dataset, and a citation relationship dataset. The citation bibliographic dataset is used to identify citing knowledge units and their lexical functions; the citation context dataset is used to identify cited knowledge units and their lexical and citation functions; and the citation relationship dataset is used for subsequent matching to establish citation relationships between citing documents, cited knowledge units, and cited sentences.

[0030] The knowledge unit and semantic function recognition module performs semantic recognition on the preprocessed data. Based on a recognition framework that integrates retrieval enhancement generation technology and a large language model, this module completes the recognition of citing knowledge units and their lexical functions, cited knowledge units and their lexical functions, and citation functions, respectively, and outputs citing knowledge units and their lexical functions, cited knowledge units and their lexical functions, and citation functions for network construction.

[0031] The interdisciplinary knowledge unit citation network construction module combines the recognition results of the knowledge unit and semantic function recognition module with subject classification information to form composite nodes. It extracts knowledge unit citation relationships from the literature citation relationship dataset as edges to construct the interdisciplinary knowledge unit citation network. Based on this, it filters and extracts interdisciplinary knowledge unit citation subnets according to the subject attributes of the citing and cited knowledge units.

[0032] The interdisciplinary knowledge unit reference network analysis module performs network topology analysis, knowledge unit interdisciplinary influence analysis, knowledge unit interdisciplinary reference pattern analysis, interdisciplinary knowledge community identification, and interdisciplinary knowledge community dynamic evolution analysis on the constructed interdisciplinary knowledge unit reference network and interdisciplinary knowledge unit reference subnets, thereby obtaining the target domain's interdisciplinary knowledge reference structure, core knowledge units, typical reference patterns, knowledge community structure, and its evolution trend.

[0033] The data acquisition and preprocessing module The data acquisition and preprocessing module is used to acquire full-text datasets of academic literature in the target domain and organize the data into a structured dataset required for subsequent knowledge unit and semantic function recognition. In one implementation, the target domain is used as the retrieval object, and data such as bibliographic titles, text, and references cited in the literature are obtained from the academic literature database.

[0034] This module performs structured parsing of full-text datasets of literature in the target domain, constructing different datasets for different recognition tasks. For the task of identifying citing knowledge units and their lexical functions, the title, abstract, and keyword fields are extracted from the bibliographic records of the cited documents to form a citing document bibliographic dataset. This dataset is used to identify citing knowledge units such as research questions and research methods in the cited documents. For the tasks of identifying cited knowledge units and their lexical functions and citation functions, citation sentences with reference identifiers are extracted from the text of the cited documents, and the reference number, chapter, publication year, and context information of the corresponding citation sentences are recorded to form a citation context dataset. To provide sufficient background semantic information, adjacent sentences before and after the citation sentence are saved as context windows in the citation context dataset.

[0035] Two data cleaning processes are performed during preprocessing. The first data cleaning targets the full-text dataset of academic literature in the target field, removing records with missing full text, abnormal formatting, and unresolved reference identifiers. The second data cleaning targets the generated datasets of citing references, citation contexts, and citation relationships, removing records with missing fields, invalid reference identifiers, duplicate citations, and data where a correspondence between citing references and cited knowledge units cannot be established. Considering that citations in the introduction and review sections are mostly used for general background information and have a relatively weak connection to the substantive knowledge of the citing references, in the preferred implementation, the citation content in the introduction and review sections is removed, while citation sentences in the methods, results, and discussion sections are retained to improve the effectiveness of subsequent identification of cited knowledge units and citation function recognition.

[0036] The data acquisition and preprocessing module ultimately outputs three types of data: the first is a dataset of citing bibliographic records, used to identify citing knowledge units and their lexical functions; the second is a dataset of citation context, used to identify cited knowledge units and their lexical and citation functions; and the third is a dataset of bibliographic citation relationships, used to establish the correspondence between citing documents, cited knowledge units, and their citation sentences. All three types of data retain fields such as document number, reference number, publication year, chapter position, and subject classification, providing a data foundation for the subsequent construction of interdisciplinary knowledge unit citation networks.

[0037] The knowledge unit and semantic function recognition module The knowledge unit and semantic function recognition module transforms the cited bibliographic dataset and citation context dataset into structured semantic function recognition results. This module comprises three sub-tasks: citation knowledge unit and its lexical function recognition, cited knowledge unit and its lexical function recognition, and citation function recognition. Citation knowledge unit and its lexical function recognition primarily targets the cited bibliographic data, identifying the research questions and methods within the cited literature. Cited knowledge unit and its lexical function recognition primarily targets the citation context data, first identifying the cited object in the citation sentence and then using this citation object to represent the cited knowledge unit in the knowledge citation relationship between documents. The cited knowledge unit includes the research methods, data, theories, or concepts cited in the citation sentence. Citation function recognition determines whether the citation behavior belongs to a functional category such as background, use, extension, or comparison. The module architecture is as follows: Figure 2 As shown. Figure 2 This invention demonstrates the process of knowledge unit and semantic function recognition based on the fusion of retrieval enhancement generation technology and a large language model. A small amount of annotated academic literature semantic function recognition data and domain terminology data are segmented and vectorized to form a knowledge base; the academic literature citation sentences or bibliographic data to be recognized are used as input, and after similarity matching, reordering and prompt word engineering processing, they are input into the large language model, which outputs knowledge unit and semantic function recognition results. Figure 2 The paper also demonstrates a multi-round verification and optimization method, which uses a manually annotated verification set to check the recognition results of knowledge units and semantic functions, and supplements the knowledge base, optimizes prompt words, and updates the recognition effect of the large language model based on the error situation.

[0038] This module employs a recognition framework that integrates retrieval-enhanced generation technology with a large language model. First, two types of external knowledge sources are constructed: one is a manually annotated sample (training set) knowledge base, used to store manually annotated bibliographical examples of citing knowledge units and their lexical functions, as well as manually annotated citation functions, cited knowledge units, and their lexical functions of citation sentences; the other is a domain terminology knowledge base, used to store professional terms and definitions in the target domain. The construction method of the aforementioned manually annotated sample knowledge base is as follows: Citing bibliographical entries and citation contexts are extracted from the full-text dataset of academic literature in the target domain. Based on pre-defined annotation specifications (categorizing citation functions into four types: background, use, extension, and comparison; categorizing cited objects into three types: research methods, data, theory, and concepts; and categorizing lexical functions into two types: research questions and research methods), the bibliographical entries and citation sentences are manually annotated with semantic functions. The acquired manually annotated data is divided into three categories according to its purpose: the first category is the training set, used to construct the aforementioned manually annotated example knowledge base, serving as the initial external knowledge source for retrieval enhancement generation; the second category is the validation set, derived from bibliographic and citation examples in the manually annotated examples that did not participate in the current round of structured prompt word construction, used for performance evaluation and error analysis after each round of optimization, and added to the manually annotated example knowledge base after successful validation; the third category is the test set, independent of the training and validation sets, used only after multiple rounds of optimization have ended for objective evaluation of the final recognition model. Subsequently, the manually annotated example knowledge base and the domain terminology knowledge base are vectorized and stored. When new citation bibliographic data or citation context data is input, similar examples and related terms are retrieved from the manually annotated example knowledge base and the domain terminology knowledge base according to the text to be recognized. For the retrieval of similar examples and terms, instead of setting a fixed similarity threshold, the top-5 documents with the highest cosine similarity are selected as candidate documents. Then, a cross-encoder model is used to re-rank the top-5 candidate documents. Unlike the initial retrieval which only calculates the vector inner product, the cross-encoder concatenates the text to be identified with each candidate document and inputs it into the model for deep semantic interaction calculation, thereby more accurately capturing the semantic relevance between the two. After re-ranking according to the relevance score, the top-3 documents with the highest re-ranking score are retained as external knowledge sources in the structured prompt words.

[0039] After the search is completed, the newly input citation bibliographic data, similar examples, relevant terms and their definitions, judgment rules, and output format are jointly filled into the structured prompt word template. The judgment rules include the priority order between categories, key trigger word matching rules, and disambiguation instructions for common misjudgment patterns. These rules are used to guide the large language model to make classification decisions that conform to academic citation norms when category boundaries are ambiguous. For details, please refer to the structured prompt word template for citation knowledge units and their lexical function recognition. Figure 3The data is then input into a large language model for recognition. Prompt word templates are used to limit the recognition targets, classification scope, and output format of the large language model, preventing it from outputting unparseable natural language descriptions. After the large language model outputs the data, format and rule validations are performed on the output results to obtain structured fields for the citing knowledge units and their lexical functions, the cited knowledge units and their lexical functions, and the citation functions.

[0040] Figure 3 This demonstrates the main components of a structured prompt template, including task background (character roles, task description, category definitions, classification priorities), decision rules, inputs and outputs (related documents and target statements, output specifications), etc. Figure 3 As shown, in the structured prompt word template of this embodiment, the character role is "You are an expert in neuroscience and citation object recognition, with in-depth knowledge of academic writing standards and classification criteria." The task description is "Identify all citation objects associated with citation tags in the target sentence and assign the correct category code (0-3) according to the following definitions, classification priorities and decision rules." The category is defined as "(0) Theory: concepts, definitions, hypotheses, theories, or research results, phenomena and opinions of others; (1) Method: ...; (2) Data: ...; (3) Not specified: ...". The classification priorities from high to low are "1. Method (1); 2. Data (2); 3. Theory (0); 4. Not specified (3)". The judgment rule part is "1. If the cited content involves the usage specifications, application scenarios or operation guidelines of the technology, it is classified as Method (1). 2. If the cited content mainly states the conclusions, opinions or phenomena, it is classified as Theory (0)..." 13. The phrase used for citations should be concise but must retain key contextual meaning (e.g., "PCR primers" instead of "primers"). The related documents and target statements are formatted as "Related documents: {retrieval text}; Related Medical Subject Headings (MeSH): {retrieval keywords}; Target citation: '{query statement}'", where {query statement} is the text to be identified, and {retrieval text} and {retrieval keywords} are similar examples and related terms retrieved from the manually labeled sample knowledge base and the domain terminology knowledge base, respectively, and then reordered based on the text to be identified. The final output specification is: "Output format: Each result must follow the following format: {{[citation label] citation object (category)}}. Multiple results are separated by commas. Do not add extra outer curly braces around all results. Constraints: Maintain academic rigor and conciseness. Do not include explanations, reasoning, or comments. Only output citation objects in the specified format." This structured prompt template is used to constrain the large language model to identify cited knowledge units and their lexical functions according to preset categories and output the identification results in the specified format, thereby improving the standardization and parsingability of the identification results.

[0041] To evaluate the quality of knowledge unit and semantic function recognition results, this module establishes an evaluation system that integrates classification evaluation and entity name generation quality evaluation. For citing knowledge units and their lexical functions, cited knowledge units and their lexical functions, and citation functions, accuracy is used. Recall rate and Values ​​are used as evaluation metrics, among which accuracy is... Recall represents the proportion of correctly identified results from a large language model among all identified results. This represents the proportion of correctly identified results from the large language model among all manually annotated recognition results in the validation or test set. The value represents the accuracy. and recall rate The harmonic mean is calculated using the formulas shown in (1), (2), and (3): (1) (2) (3) in, This represents the number of citing knowledge units and their lexical functions, as well as the number of cited knowledge units and their lexical functions, correctly identified by the large language model. This represents the number of citing knowledge units and their lexical functions, as well as the number of cited knowledge units and their lexical functions, that were incorrectly identified or redundantly generated by the large language model. This indicates the number of citing knowledge units and their lexical functions, as well as cited knowledge units and their lexical functions, that the large language model failed to identify.

[0042] For the task of identifying cited knowledge units and citing knowledge units, since the citing knowledge units and cited knowledge units output by the large language model may not be completely consistent with the literal form of all manually annotated identification results in the validation set or test set, but have the same semantics, ROUGE-1, ROUGE-2, ROUGE-L and BERTScore are further used as lexical character similarity evaluation indicators for the identification results of cited knowledge units and citing knowledge units. ROUGE-1 is used to evaluate the unary word overlap between the cited and citing knowledge units output by the large language model and the manually annotated cited and citing knowledge units. It is calculated based on word overlap or longest common subsequence at different granularities. ROUGE-2 is used to evaluate the bigram overlap between the cited and citing knowledge units output by the large language model and the manually annotated cited and citing knowledge units. ROUGE-L is used to evaluate the matching of the longest common subsequence between the cited and citing knowledge units output by the large language model and the manually annotated cited and citing knowledge units. Thus, it measures the similarity between the cited and citing knowledge units output by the large language model and the manually annotated cited and citing knowledge units at the lexical character level. BERTScore, based on the contextual semantic vectors generated by the BERT model, evaluates the semantic similarity between the cited and citing knowledge units output by the large language model and those manually annotated. The calculation formulas are shown in (4), (5), and (6). Here, we mainly focus on BERTScore. Value, that is The above evaluation system allows for the simultaneous assessment of the accuracy of identifying cited and citing knowledge units based on semantic function categories (i.e., lexical function and citation function categories). Recall rate and The recognition performance is evaluated based on factors such as the similarity of words and characters and semantic similarity between the cited and citing knowledge units, and the evaluation results are used for subsequent optimization of structured prompt words and supplementation of manually annotated sample knowledge base.

[0043] (4) (5) (6) in, This represents the set of lexical units obtained after word segmentation of the citing and cited knowledge units output by the large language model. This represents the set of lexical units obtained after word segmentation of manually annotated citing and cited knowledge units. Represents a set The first in The context semantic vector corresponding to each word element Represents a set The first in The context semantic vector corresponding to each word element; Represents a set The number of lexical units in the text Represents a set The number of lexical units in the text; Representing vectors with vector Cosine similarity between them; The semantic vectors in the citing and cited knowledge units output by the large language model represent the citing knowledge units. The maximum semantic similarity obtained by matching the semantic vectors in the manually annotated citing and cited knowledge units; This represents the semantic vectors in manually annotated citing and cited knowledge units. The maximum semantic similarity obtained by matching the semantic vectors in the citing and cited knowledge units output by the large language model; This represents the semantic recall rate calculated based on the citing and cited knowledge units output by the large language model. This represents the semantic precision calculated based on manually annotated citing and cited knowledge units. express and The harmonic mean. This implementation mainly adopts... Evaluate the overall semantic consistency between the citing and cited knowledge units output by the large language model and the manually annotated citing and cited knowledge units.

[0044] To improve the accuracy and stability of recognition in complex contexts within specialized fields, this module further employs a multi-round validation optimization method. The validation set comprises bibliographical examples and citation examples from manually annotated samples that did not participate in the current round of structured prompt word construction. These examples include manually annotated results of citing knowledge units and their lexical functions, cited knowledge units and their lexical functions, and citation functions. The system uses the validation set to examine the output of the large language model and calculates the accuracy based on the aforementioned evaluation system. Recall rate The F1 score, ROUGE metric, and BERTS score were used to compare the output of the large language model with the manually annotated results, identifying entity boundary errors, category misclassifications, missed identifications, multiple identifications, and output format errors. Based on these error types, the category definitions, decision rules, classification priorities, and output specifications in the structured prompts were corrected. After completing the correction of the structured prompts for the current round, the manually annotated and validated bibliographic examples and citation examples were added to the manually annotated example knowledge base for use in subsequent rounds of retrieval enhancement generation.

[0045] The multi-round validation optimization follows an iterative process of "identification, evaluation, error analysis, prompt word correction, and supplementation of the sample knowledge base." If, after several rounds of optimization, the accuracy of the large language model's output on the validation set... Recall rate If the F1 score, ROUGE metric, and BERTS score no longer improve, then the optimization round in which the evaluation metrics reach their highest values ​​is determined as the optimal round. Assume the... The round with the highest evaluation index value will be the [number]th [round]. The structured prompt word version is denoted as and the structured prompt word version The supplemented manually annotated sample knowledge base, the original domain terminology knowledge base, and the large language model are combined to form the final recognition model RAG_LLM_Pn. Finally, the system uses RAG_LLM_Pn to perform tasks such as identifying citing knowledge units and their lexical functions, identifying cited knowledge units and their lexical functions, and identifying citation functions, and outputs knowledge unit and semantic function recognition results that can be used for subsequent network construction.

[0046] The interdisciplinary knowledge unit reference network construction module The interdisciplinary knowledge unit citation network building module is used to transform the recognition results, subject classification information, and literature citation relationship datasets output by the knowledge unit and semantic function recognition module into a computable network structure. An example of the network structure is shown below. Figure 4 As shown. Figure 4 This invention demonstrates how it associates citing knowledge units in target documents with cited knowledge units in the cited text. The left side of the figure shows citing knowledge units such as research questions and methods in the citing document, while the right side shows cited knowledge units such as research methods, data, theories, or concepts in the cited sentence. The citation border in the middle indicates citation functions such as use, comparison, and expansion. Figure 4This module can intuitively illustrate the basic approach of this invention: constructing a network using knowledge units as nodes and semantically functional reference relationships as edges. It first builds an interdisciplinary knowledge unit reference network, and then, based on this network, selects interdisciplinary knowledge unit reference subnets according to subject attributes, thereby simultaneously supporting both overall analysis of domain knowledge structure and interdisciplinary knowledge flow analysis.

[0047] In terms of network model definition, this invention defines the interdisciplinary knowledge unit referencing network as a directed weighted graph. .in, This represents the set of nodes representing the citing and cited knowledge units. This represents the set of directed edges formed between knowledge units extracted from the literature citation relation dataset based on citation relationships. Represents the set of edge weights. This represents a set of node attributes and edge attributes. Nodes are no longer based on documents, authors, or keywords as the basic unit, but rather on knowledge units (citing and cited knowledge units). Each node consists of the "entity name of the citing or cited knowledge unit, its lexical function classification code, and its subject classification code," used to simultaneously express the knowledge content, the function of the knowledge unit in scientific research, and the subject to which the knowledge unit belongs. Specifically, the entity name of the knowledge unit is output by the knowledge unit and semantic function recognition module; the lexical function classification code is pre-set according to the lexical function of the knowledge unit in scientific research. In this embodiment, the research question in the citing knowledge unit is coded as 0, and the research method is coded as 1; the research method in the cited knowledge unit is coded as 1, the data as 2, and the theory or concept as 3. The subject classification code is set according to the second-level subject field in the OpenAlex subject classification system, used to identify the subject area to which the knowledge unit belongs; in this embodiment, the target subject neuroscience is coded as 0, and other external subjects are coded sequentially from 1 to 25 according to the preset subject classification table. Therefore, the "knowledge unit entity name, vocabulary function category code, and subject classification code" are concatenated to form a node label. For example, "neuroprotective effects_0_0" represents a knowledge unit node whose knowledge unit entity name is neuroprotective effects, vocabulary function category is research question, and subject classification is neuroscience.

[0048] During node construction, source nodes are generated based on the citing knowledge unit, and target nodes are generated based on the cited knowledge unit. The citing knowledge unit includes the research question and research method, while the cited knowledge unit includes research methods, data, theories, or concepts. The entity name of the knowledge unit is combined with its corresponding lexical function category and subject classification category to form a composite node that distinguishes different lexical functions and subject classifications. Through this composite node structure, the same citing or cited knowledge unit entity name is distinguished under different lexical functions or subject classification attributes, avoiding the mixing of knowledge units with different lexical functions or subject classifications.

[0049] During edge construction, based on the literature citation relationship dataset, the citing knowledge units in the citing literature are matched with the cited knowledge units corresponding to their citation sentences, and directed edges are generated in the direction of "the citing knowledge unit pointing to the cited knowledge unit". Edge attributes include citation function, citation time, and chapter position. The citation function indicates whether the citation relationship belongs to categories such as background, use, extension, or comparison; the citation time is represented by the publication year of the citing literature; and the chapter position is represented by the chapter in which the citation sentence is located. Multiple citation relationships with the same source node, target node, and citation function are aggregated, and the citation frequency is used as the edge weight to represent the strength of the citation relationship between the citing and cited knowledge units.

[0050] After constructing the interdisciplinary knowledge unit referencing network, the next step is to filter the interdisciplinary knowledge unit subnets. Specifically, all directed edges in the interdisciplinary knowledge unit referencing network are traversed, and their interdisciplinary referencing relationships are determined based on the subject classification codes of the source and target nodes. If the source node belongs to the target subject and the target node belongs to a non-target subject, the edge and its two endpoints are retained; if the edge represents a referencing within the target subject or between external subjects, it is discarded. After filtering, isolated nodes are deleted, forming the interdisciplinary knowledge unit referencing subnet. This subnet possesses binary, directed, weighted, and multi-semantic attributes, capable of depicting the specific paths by which target subject knowledge units reference external subject knowledge units.

[0051] The interdisciplinary knowledge unit reference network analysis module The interdisciplinary knowledge unit citation network analysis module is used to perform multi-dimensional analysis of interdisciplinary knowledge unit citation networks and subnetworks, outputting the target domain's interdisciplinary knowledge citation structure, core knowledge units, typical citation patterns, knowledge community structure, and evolution trends. This module mainly includes network topology analysis, interdisciplinary influence analysis of knowledge units, interdisciplinary citation pattern analysis of knowledge units, interdisciplinary knowledge community identification, and dynamic evolution analysis of interdisciplinary knowledge communities, such as... Figure 5 As shown. Figure 5This diagram illustrates the process by which, after constructing the interdisciplinary knowledge unit referencing network and its subnets, the invention further conducts network topology analysis, interdisciplinary influence analysis of knowledge units, interdisciplinary referencing pattern analysis of knowledge units, interdisciplinary knowledge community identification, and dynamic evolution analysis of interdisciplinary knowledge communities. The diagram demonstrates that the invention not only constructs networks but also enables multi-dimensional analysis of interdisciplinary knowledge referencing patterns based on these networks.

[0052] Network topology analysis is used to characterize the overall organizational form of interdisciplinary knowledge unit referencing networks. Taking the interdisciplinary knowledge unit referencing network and its subnets as computational objects, the network's connectivity, sparsity, clustering, and structural distribution characteristics are analyzed by calculating indicators such as the number of nodes, number of edges, average degree, average weighted degree, network density, network diameter, average path length, connected components, and modularity. The number of nodes and edges measures network size; average degree and average weighted degree measure the average degree of association between knowledge units; network density measures connection sparsity; network diameter and average path length measure the length of knowledge flow paths; connected components measure network connectivity; and modularity measures the strength of the community structure. For the interdisciplinary knowledge unit referencing subnets, due to their directed, weighted, and bipartite network characteristics, the referencing knowledge unit nodes and the cited knowledge unit nodes belong to different node sets, making it difficult for traditional clustering coefficients to directly reflect the local clustering relationships in this bipartite network. Therefore, a network projection method is used to construct the referencing node projection network and the cited knowledge unit node projection network, respectively. The projected network of citing nodes is used to represent whether different citing knowledge units jointly cite the same cited knowledge unit, while the projected network of cited knowledge unit nodes is used to represent whether different cited knowledge units are jointly cited by the same citing knowledge unit. By calculating the density of the projected network or the average clustering coefficient, the degree of overlap in knowledge needs among citing knowledge units and the co-citation relationship among cited knowledge units are analyzed, thereby supplementing the characterization of the local clustering features of the cross-disciplinary knowledge unit citation subnet.

[0053] Interdisciplinary influence analysis of knowledge units is used to identify core cited knowledge units and active citing knowledge units in interdisciplinary knowledge unit citation networks. Using nodes and directed edges in the interdisciplinary knowledge unit citation network as the calculation objects, the activity and coreness of nodes are measured by in-degree, out-degree, and weighted degree. Specifically, in-degree represents the number of directed edges pointing to a node, used to measure the number of times a cited knowledge unit is cited by different citing knowledge units; out-degree represents the number of directed edges from a node to other nodes, used to measure the activity level of citing external knowledge units by citing knowledge units; weighted in-degree represents the sum of the weights of directed edges pointing to a node, used to measure the cumulative intensity of citations of the cited knowledge unit; and weighted out-degree represents the sum of the weights of directed edges from a node to other nodes, used to measure the cumulative intensity of citations of external knowledge by citing knowledge units. The edge weights of the directed edges are obtained by aggregating the citation frequencies of the same citing knowledge unit, the same cited knowledge unit, and the same citation function category. Furthermore, the PageRank algorithm is used to calculate the PageRank value of the citing or cited knowledge unit from the perspective of the global network structure, i.e., the importance of the node in the network. Based on high-importance nodes, fundamental theories, key methods, or important data located on the core path of knowledge flow are identified. Simultaneously, the Hyperlink-Induced Topic Search (HITS) algorithm is used to calculate the authority value and hub value of the citing knowledge unit. The authority value reflects the degree of recognition of the cited knowledge unit as an external knowledge source, while the hub value reflects the citing knowledge unit's ability to absorb and organize external knowledge.

[0054] Interdisciplinary citation pattern analysis of knowledge units is used to reveal the citation combination rules between knowledge units with different lexical function types. Based on the lexical function categories of the citing and cited knowledge units, citation relationships are divided into six patterns: "Research problem citing research method," "Research problem citing data," "Research problem citing theory or concept," "Research method citing research method," "Research method citing data," and "Research method citing theory or concept." In specific statistics, each directed edge in the interdisciplinary knowledge unit citation subnet is used as the statistical object. The lexical function categories of the source node and the target node of the directed edge are read and combined into a citation pattern. For example, when the source node's lexical function category is "research problem" and the target node's lexical function category is "research method," the directed edge is classified into the "research problem citing research method" pattern. For each citation pattern, the number of unique patterns, the total number of citations, and the average citation intensity are counted. The number of unique patterns is the number of unique "source node entity name - target node entity name" combinations that are not repeated within that pattern; the total number of citations is the sum of the edge weights of all directed edges within that pattern; and the average citation intensity is the ratio of the total number of citations to the number of unique patterns. Based on the above statistical results, the system is able to compare the distribution differences of interdisciplinary knowledge citation patterns under different combinations of lexical functions.

[0055] Interdisciplinary knowledge community identification is used to discover sets of citing or cited knowledge units that are closely connected and semantically related in an interdisciplinary knowledge unit citation network. The Louvain community detection algorithm is used to divide the interdisciplinary knowledge unit citation network into multiple knowledge communities. For each knowledge community, network metrics, citation function category distribution, and document position distribution of citing and cited knowledge unit nodes within the community are statistically analyzed. Specifically, citing knowledge units with the highest out-degree, weighted out-degree, PageRank value, or HITS hub value within the community are identified as core citing nodes; cited knowledge units with the highest in-degree, weighted in-degree, PageRank value, or HITS authority value within the community are identified as core cited knowledge units; the number or weight of directed edges corresponding to various citation function categories within the community is counted, and the citation function category with the highest proportion is identified as the main citation function of the community; the number or weight of directed edges corresponding to various document positions within the community is counted, and the document position with the highest proportion is identified as the main document position of the community. Subsequently, by combining the core citing nodes, core cited knowledge units, main citation functions, and main chapter positions, the themes and interdisciplinary integration types of knowledge communities are explained. Through this process, stable combinations of interdisciplinary knowledge integration are discovered at the network structure level, and different types of interdisciplinary knowledge integration patterns are further summarized.

[0056] The interdisciplinary knowledge community dynamic evolution analysis is used to characterize the changing trends of knowledge integration patterns from a time perspective. This analysis constructs a dynamic monitoring framework based on a fixed community structure: using the core knowledge communities identified in network analysis as a benchmark, and according to the citation time attributes in the literature citation relationship dataset, the interdisciplinary knowledge unit citation network is divided into different time slices; by statistically analyzing the changes in citation intensity within each knowledge community under different time slices, activity evolution curves of different interdisciplinary integration patterns are plotted, thereby capturing the macro-migration path of the research focus of the field. That is, when the citation intensity of a certain knowledge community shows a continuous increase in adjacent time slices and the average annual growth rate is significantly higher than that of other communities, it indicates that the interdisciplinary knowledge integration hotspot is in a period of rapid development. The process involves several stages: the emergence stage, the continuation stage (when citation intensity remains high and stable across multiple time slices), and the decline stage (when citation intensity declines continuously after reaching its peak). The process then tracks the annual replacement of core cited knowledge units within each knowledge community, analyzing the specific content and direction of knowledge unit flow. Specifically, when a core cited knowledge unit with high citation frequency within a knowledge community is replaced in an adjacent time slice, it indicates a shift in the content evolution direction of this interdisciplinary knowledge integration hotspot. This reveals how an external knowledge unit gradually evolves from peripheral exploratory citations into a core infrastructure, or how a theoretical concept is reactivated and experiences explosive growth under new technological conditions.

[0057] Example The following uses academic literature in the field of neuroscience as an example to illustrate the implementation process of this invention. This embodiment is only used to explain the technical solution of this invention and does not constitute a limitation on the scope of protection of this invention.

[0058] Experimental Dataset The data in this embodiment comes from PubMed and PubMed Central databases. PubMed is a core biomedical literature retrieval platform, and PubMed Central stores a large number of full-text biomedical literatures. The two can be linked through document identifiers. First, a topic search query "(Neuroscience[Title / Abstract] OR Neurosciences[Title / Abstract]) AND 'free full text'[sb]" was constructed in the PubMed database, with the search period ending on October 1, 2023, yielding document identifiers for 23,879 neuroscience-related documents. Subsequently, the PubMed Central database interface was used for matching and full-text download, ultimately obtaining 17,265 full-text neuroscience documents in XML format.

[0059] The aforementioned full-text documents are subjected to structured parsing to extract information such as title, abstract, keywords, main text, references, citation sentences, publication year, and chapter position. This information is then stored as a citation reference bibliographic dataset, a citation context dataset, and a bibliographic citation relationship dataset. The citation context dataset is formed by extracting citation sentences with reference identifiers and extending them one sentence before and after as context windows. In this embodiment, to reduce interference from general background citations on the identification of substantive knowledge associations, citations in the introduction and review sections are removed during the preprocessing stage, while citation sentences in the methods, results, and discussion sections are retained. After screening and cleaning, a citation sentence dataset containing 103,669 records is finally constructed for subsequent citation function determination and extraction of cited knowledge units.

[0060] Implementation of Knowledge Unit and Semantic Function Recognition Module In the knowledge unit and semantic function recognition module, this embodiment constructs three tasks: the citing knowledge unit and its lexical function recognition task, the cited knowledge unit and its lexical function recognition task, and the citation function recognition task. The citing knowledge unit and its lexical function recognition task takes the title, abstract, and keywords of the cited document as input to identify the research question and research method; the cited knowledge unit and its lexical function recognition task takes the cited sentence and its context as input to identify the cited research method, data, theory, or concept; and the citation function recognition task takes the cited sentence and its context as input to identify citation functions such as background, use, extension, and comparison.

[0061] This embodiment employs a recognition framework that integrates retrieval-enhanced generation technology with a large language model. First, a manually annotated sample knowledge base and a domain terminology knowledge base are constructed. The manually annotated sample knowledge base is derived from manually annotated bibliographic texts and citations, while the domain terminology knowledge base incorporates medical subject headings as supplementary professional terminology. These knowledge bases are stored in vector form. When recognizing each text to be processed, similar samples and related terminology explanations are retrieved. The search results, the text to be recognized, category definitions, judgment rules, and output format are then combined and fed into a structured prompt word template, which is then input into the large language model to complete the recognition.

[0062] In this embodiment, after multiple rounds of verification and optimization, the citation function recognition task has achieved a weighted average citation function recognition result. The value reached 86.21%; entity name recognition of cited knowledge units. The F1 score for citation function recognition reached 81.43%, and the BERTS score reached 92.48%. The F1 score for entity name recognition of citation knowledge units reached 83.44%, and the BERTS score reached 93.58%. The multi-round validation and optimization process for the citation function recognition task is shown in Table 1. These results demonstrate that this module can reliably identify citation or cited knowledge units and their lexical functions from academic texts in a specialized field, providing structured semantic data for subsequent network construction.

[0063] Table 1. Multi-round validation and optimization process of the citation function recognition task Prompt word version Optimization plan and explanation RAG training set knowledge base data volume P0 The initial framework includes basic modules (Persona, Context, Task, Category Definitions, Input, Output), but the definitions are brief and lack specific decision rules and priorities, making the model prone to inconsistent classifications. 500 P5 Introduced priority and extension rules. A new Priority module was added, establishing the decision order as Extend > Compare > Use > Background > Other. In Category Definitions, an overriding rule "overrides Use" was added to the "Extend" class, along with supplementary trigger words (such as modified, adapted). Key rules and disambiguation logic were strengthened. A new Key Rules module was added, integrating the core decision logic from previous iterations (such as focusing on the core sentence and contextual disambiguation). 600 P10 Preferred rules (If unclear... choose Use / Extend) have been added to the "Use" and "Extend" categories to improve the clarity of the classification. 700 P11 The "Background" class is precisely defined. To address the issue of the model misclassifying "only reporting others' work" as "use," a key explanation is added to the definition of "Background": "Used when the citation only reports what others did, without this study adopting it." 800 P12 Expand the scope of the "Extended" class by adding extended class data. To handle cases like "providing stronger evidence", add rule 8 to Key Rules: "Providing stronger / more direct evidence than prior work counts as Extend (2)." Also, since extended class data accounts for a small proportion in the training set, supplementary extended data is added. 820 P13 The optimal version. Added key sentence structure rule 9: "'In our case / here we / in this study' → treat as Use / Extend, not Compare.", effectively resolving a common misjudgment. With this addition, the prompt word structure is complete, the rule system is stable, and optimal performance is achieved on the test set. 920 P14 Experimental Adjustment 1 (Results Decreased). Based on P13, we attempted to add more granular negation rules for the "comparison" category (such as adding "not necessarily implying..."), which caused the model to be overly conservative in some sentences that should have belonged to the "comparison" category, resulting in a decrease in recall. 920 P15 Trial Adjustment 2 (Decreased Effectiveness). An attempt was made to further expand the range of trigger words for "use" (e.g., adding "based on" and "following") in order to capture more edge use cases, but this resulted in some "background" sentences being misjudged as "use", leading to a decrease in precision. 1020 P16 Trial Adjustment 3 (Decreased Effectiveness). An attempt was made to modify the Priority order (e.g., moving Compare priority before Extend) to accommodate specific comparison sentence structures, but this severely disrupted the stable adjudication logic established on page 13, resulting in poorer overall classification consistency and a decline in all indicators. 1020 Table 1 shows the multi-round validation and optimization process of the citation function recognition task. This table illustrates the optimization process from the initial prompt word version to the optimal version and subsequent trial versions, including the prompt word versions, optimization schemes and explanations, and the amount of data in the knowledge base for generating the retrieval enhancement training set. This table demonstrates that the present invention, through multi-round validation and optimization, continuously corrects the category definition, judgment rules, classification priorities, and sample knowledge base, thereby improving the accuracy and stability of citation function recognition.

[0064] Implementation of interdisciplinary knowledge unit reference network building modules In the interdisciplinary knowledge unit citation network construction module, knowledge units are matched with the recognition results of the semantic function recognition module, the literature citation relationship dataset, and subject classification information. Specifically, research questions and methods in neuroscience literature are used as citing knowledge units, while research methods, data, theories, or concepts in cited literature or citation context are used as cited knowledge units. Directed edges are generated in the direction of "citing knowledge units pointing to cited knowledge units".

[0065] Each knowledge unit node is configured with a knowledge unit entity name, lexical function code, and subject classification code. Each citation edge is configured with attributes such as citation function, publication year, chapter position, and citation frequency. In this embodiment, background citations are used for general background descriptions and are not included in the final network construction; use, expansion, and comparison citations are used to construct a semantically sensitive knowledge unit citation network. Repeated citation relationships with the same source node, target node, and citation function are aggregated, and citation frequency is used as the edge weight. The node table and edge representation are shown in Tables 2 and 3, for example.

[0066] Table 2 Node Representation Examples id label type category field encoding 66000000001 neuroprotective effects_0_0 source 0 0 66000000002 minocycline_1_0 source 1 0 66000000003 laser axotomy_1_0 source 1 0 66000000004 chronic implantation_0_0 source 0 0 Table 3 Examples of edge representation Source Target Citation function secType citedYear weight Source_entity Target_entity 66000006382 68000118086 1 Methods 2021 1 neural inter-frequency power correlations_0_0 analytic wavelet_1_1 66000006383 68000118086 1 Methods 2021 1 statistical significance test_1_0 analytic wavelet_1_1 66000006382 68000118087 1 Methods 2021 1 neural inter-frequency power correlations_0_0 WPS coefficients_2_1 66000006383 68000118087 1 Methods 2021 1 statistical significance test_1_0 WPS coefficients_2_1 Table 2 shows an example of node representation. This table displays the basic fields of node data in a cross-disciplinary knowledge unit reference network, including node number, node label, node type, lexical function category, and subject classification code. The node label, composed of the knowledge unit entity name, lexical function code, and subject classification code, represents the content, semantic function, and subject attribute of the knowledge unit.

[0067] Table 3 shows an example of edge representation. This table displays the basic fields of edge data in the interdisciplinary knowledge unit citation network, including source node, target node, citation function, document position, cited document year, edge weight, source node entity, and target node entity. This table illustrates that this invention represents the citation behavior of a citing knowledge unit pointing to a cited knowledge unit as a directed edge, and characterizes the interdisciplinary citation relationships between knowledge units through attributes such as citation function, document position, year, and weight.

[0068] Based on the above method, the initial overall network constructed in this embodiment contains 49,009 knowledge unit nodes and 185,745 weighted directed edges. Considering that low-frequency reference relationships may interfere with the core structure analysis, an edge weight threshold is set for filtering. The resulting semantically sensitive interdisciplinary overall knowledge unit reference network contains 16,037 high-frequency knowledge unit nodes and 41,746 strongly related edges. This network simultaneously retains both internal knowledge reference paths within the neuroscience field and external interdisciplinary knowledge reference paths. Furthermore, reference relationships where the source node belongs to neuroscience and the target node belongs to non-neuroscience are further filtered according to disciplinary attributes, forming an interdisciplinary knowledge unit reference subnet, such as... Figure 6 As shown. Figure 6The network structure of the interdisciplinary knowledge unit referencing subnet is illustrated, where nodes represent knowledge units and edges represent referencing relationships between referencing knowledge units and referenced knowledge units. This diagram allows observation of the connections between target subject knowledge units and external subject knowledge units, the distribution of core nodes, and the local knowledge structure, demonstrating the invention's ability to visually represent interdisciplinary knowledge referencing paths.

[0069] Implementation of the interdisciplinary knowledge unit reference network analysis module In the interdisciplinary knowledge unit citation network analysis module, the network topology of the overall network and interdisciplinary subnets is first analyzed, calculating indicators such as the number of nodes, number of edges, network density, average degree, average weighted degree, connected components, and modularity. The citation relationships between knowledge units can be visualized through the network. Subsequently, using in-degree, out-degree, weighted degree, webpage ranking algorithms, and hyperlink-induced topic search algorithms, core cited knowledge units and actively citing knowledge units are identified.

[0070] In the analysis of interdisciplinary citation patterns of knowledge units, based on the lexical function combination of the citing and cited knowledge units, citation patterns are statistically analyzed, including research questions citing research methods, research questions citing data, research questions citing theories or concepts, research methods citing research methods, research methods citing data, and research methods citing theories or concepts. The number of unique patterns, total number of citations, and average citation intensity for each type of pattern are calculated. The statistics on the number of citation patterns are as follows: Figure 7 As shown. Figure 7 This study displays the quantity and volume of interdisciplinary knowledge citation patterns under different combinations of lexical functions, such as research questions citing research methods, research questions citing data, research questions citing theories or concepts, research methods citing research methods, research methods citing data, and research methods citing theories or concepts. This graph allows for comparison of the distribution differences of different knowledge citation patterns, thereby revealing the main patterns of interdisciplinary knowledge utilization in the target discipline. In the identification of interdisciplinary knowledge communities, a community discovery algorithm is used to identify closely connected sets of citing and cited knowledge units. Combined with the core citing nodes, core cited nodes, main citation functions, and main chapter positions within the community, the types of interdisciplinary knowledge fusion are explained. In the dynamic evolution analysis, time slices are formed according to the publication year, and the annual changes in citation intensity for different knowledge communities or citation patterns are statistically analyzed to obtain interdisciplinary knowledge fusion hotspots and their evolutionary trends.

[0071] Experimental Results and Effectiveness Through the above implementation process, this invention can transform unstructured knowledge content in the full text of academic literature into a cross-disciplinary knowledge unit citation network with lexical functions, subject attributes, citation functions, temporal attributes, and textual position attributes. Compared with methods that only use literature, keywords, or topics as analysis objects, this embodiment can further identify citation paths between specific research questions, research methods, data, theories, or concepts, and reveal the cross-disciplinary knowledge citation structure and patterns in the field of neuroscience from multiple dimensions such as network structure, node influence, citation patterns, knowledge communities, and dynamic evolution.

Claims

1. A system for constructing and analyzing interdisciplinary knowledge unit citation networks for academic literature, characterized in that, It includes a data acquisition and preprocessing module, a knowledge unit and semantic function recognition module, an interdisciplinary knowledge unit reference network construction module, and an interdisciplinary knowledge unit reference network analysis module; The data acquisition and preprocessing module is used to acquire and preprocess the full-text dataset of academic literature in the target field, and generate a dataset of cited bibliographic records, a dataset of citation context, and a dataset of literature citation relationships. The knowledge unit and semantic function recognition module is used to perform semantic function recognition on the preprocessed data, and to identify the citing knowledge unit and its lexical function based on the citing bibliographic dataset, and to identify the cited knowledge unit and its lexical function and citation function based on the citation context dataset. The interdisciplinary knowledge unit citation network construction module is used to combine semantic recognition results and subject classification information into composite nodes, extract knowledge unit citation relationships from the literature citation relationship dataset as edges, construct an interdisciplinary knowledge unit citation network, and then extract interdisciplinary knowledge unit citation subnets based on the subject attributes of the citing knowledge unit and the cited knowledge unit. The interdisciplinary knowledge unit reference network analysis module performs network topology analysis, knowledge unit interdisciplinary influence analysis, knowledge unit interdisciplinary reference pattern analysis, interdisciplinary knowledge community identification, and interdisciplinary knowledge community dynamic evolution analysis on the constructed interdisciplinary knowledge unit reference network and interdisciplinary knowledge unit reference subnets, thereby obtaining the target domain's interdisciplinary knowledge reference structure, core knowledge units, typical reference patterns, knowledge community structure, and its evolution trend.

2. The system according to claim 1, characterized in that, The full-text dataset of academic literature in the target field includes bibliographic information, text, and references cited in the literature.

3. The system according to claim 2, characterized in that, The preprocessing includes structured parsing of the full-text dataset of academic literature in the target field and two data cleaning processes. The structured parsing includes extracting title, abstract, and keyword fields from the cited references to form a cited reference bibliographic dataset, and extracting citation sentences with reference identifiers from the text of the cited references, and recording the reference number, chapter, publication year, and context information corresponding to the citation sentences to form a citation context dataset. The first data cleaning process removes records that are missing, have abnormal formatting, or whose reference identifiers cannot be parsed from the full-text data of academic literature in the target field; The second data cleaning process removes records with missing fields, invalid reference identifiers, duplicate citation records, and data from the citation bibliographic dataset, citation context dataset, and bibliographic citation relationship dataset that cannot establish a correspondence between citations and cited knowledge units.

4. The system according to claim 1, characterized in that, The specific steps for recognizing the knowledge unit and semantic function are as follows: The bibliographical examples of the citing knowledge units and their lexical functions, as well as the citation sentence examples of the citation function, the cited knowledge units and their lexical functions, are manually annotated, and the manually annotated examples are divided into training set, validation set and test set. A knowledge base of manually annotated examples is built based on the training set and combined with an existing domain terminology knowledge base for vectorized storage; The domain terminology knowledge base is used to store the professional terms and definitions of the target domain. When new citation data or citation context data is input, the system retrieves the top-ranked references with the highest cosine similarity from both the manually annotated sample knowledge base and the domain terminology knowledge base. Each set of examples and terms is given, and then reordered, retaining the top-ranked terms with the highest reorder scores. Examples and terms, ; The newly input citation bibliographic data, the retained examples and terms and their definitions, the judgment rules and the output format are all filled into the structured prompt word template, and then input into the large language model for recognition to obtain structured fields of citing knowledge units and their lexical functions, cited knowledge units and their lexical functions and citation functions.

5. The system according to claim 4, characterized in that, The knowledge unit and semantic function recognition module also employs accuracy. Recall rate and The evaluation criteria include the citing knowledge unit and its lexical function, the cited knowledge unit and its lexical function, and the citation function, among which precision is considered. Recall represents the proportion of correctly identified results from a large language model among all identified results. This represents the proportion of correct recognition results from the large language model compared to the recognition results on the validation or test set of manually labeled data. The value represents the accuracy. and recall rate The harmonic mean is calculated using the formulas shown in (1), (2), and (3): （1） （2） （3） in, This represents the number of citing knowledge units and their lexical functions, as well as the number of cited knowledge units and their lexical functions, correctly identified by the large language model. This represents the number of citing knowledge units and their lexical functions, as well as the number of cited knowledge units and their lexical functions, that were incorrectly identified or redundantly generated by the large language model. This represents the number of citing knowledge units and their lexical functions, and cited knowledge units and their lexical functions that the large language model failed to identify. ROUGE-1, ROUGE-2, ROUGE-L, and BERTScore are used to evaluate the lexical character similarity between cited and citing knowledge units (CKIs) identified by the large language model and those manually annotated. Specifically, ROUGE-1 evaluates the unarylexical overlap between CKIs identified by the large language model and those manually annotated, calculated based on term overlap or longest common subsequence at different granularities; ROUGE-2 evaluates the bigram overlap between CKIs identified by the large language model and those manually annotated; and ROUGE-L evaluates the lexical character similarity between CKIs identified by the large language model and those manually annotated. The longest common subsequence matching between the cited knowledge units and citing knowledge units identified by the large language model and the manually annotated cited knowledge units and citing knowledge units is used to measure the similarity between the cited knowledge units and citing knowledge units identified by the large language model and the manually annotated cited knowledge units and citing knowledge units at the lexical character level. BERTScore, based on the context semantic vector generated by the BERT model, evaluates the semantic similarity between the cited knowledge units and citing knowledge units identified by the large language model and the manually annotated cited knowledge units and citing knowledge units by calculating the semantic similarity between them. The calculation formulas are shown in (4), (5) and (6): （4） （5） （6） in, This represents the set of lexical units obtained after word segmentation of the citing and cited knowledge units identified by the large language model. This represents the set of lexical units obtained after word segmentation of manually annotated citing and cited knowledge units. Represents a set The first in The context semantic vector corresponding to each word element Represents a set The first in The context semantic vector corresponding to each word element; Represents a set The number of lexical units in the text Represents a set The number of lexical units in the text; Representing vectors with vector Cosine similarity between them; The semantic vectors in the citing and cited knowledge units recognized by the large language model. The maximum semantic similarity obtained by matching the semantic vectors in the manually annotated citing and cited knowledge units; This represents the semantic vectors in manually annotated citing and cited knowledge units. The maximum semantic similarity obtained by matching the semantic vectors of the citing and cited knowledge units identified by the large language model; This represents the semantic recall rate calculated based on the citing and cited knowledge units output by the large language model. This represents the semantic precision calculated based on manually annotated citing and cited knowledge units. express and The harmonic mean; Based on accuracy Recall rate The evaluation metrics include F1 score, ROUGE-1, ROUGE-2, ROUGE-L, and BERTScore. The large language model's recognition results are compared with manually annotated results to identify error types, including entity boundary errors, category misclassification, missed recognitions, multiple recognitions, and output format errors. The structured prompt word template is corrected based on the identified error types. The manually annotated bibliographic examples and citation examples used for verification in this round are added to the manually annotated example knowledge base for the next iteration. This process continues until the accuracy of the large language model's recognition results compared with the manually annotated results is achieved. Recall rate If the F1 score, ROUGE-1, ROUGE-2, ROUGE-L and BERTScore evaluation metrics no longer improve, then the optimization round that reaches the highest evaluation metric is determined as the optimal round. The structured prompt word version of the optimal round, the supplemented manually annotated sample knowledge base, the original domain terminology knowledge base and the large language model are combined to form the final knowledge unit and semantic function recognition model.

6. The system according to claim 1, characterized in that, The network model of the interdisciplinary knowledge unit referencing network is defined as a directed weighted graph. ,in, This represents the set of citing knowledge unit nodes and the set of cited knowledge unit nodes. This represents the set of directed edges formed between knowledge units extracted from the literature citation relation dataset based on citation relationships. Represents the set of edge weights. Represents the set of node attributes and edge attributes; The citing or cited knowledge unit node includes the entity name of the citing or cited knowledge unit and its lexical function classification code and subject classification code; the citing knowledge unit node serves as the source node, and the cited knowledge unit node serves as the target node; the lexical function classification code is pre-set according to the lexical function of the citing or cited knowledge unit in scientific research, and includes the research question code, research method code, research method code, data code, and theory or concept code in the citing knowledge unit; the subject classification code is set according to the second-level subject field in the OpenAlex subject classification system, and is used to identify the subject field to which the knowledge unit belongs, including the target subject code and the external subject classification code; the entity name of the citing or cited knowledge unit and its lexical function classification code and subject classification code are concatenated to form a node label.

7. The system according to claim 6, characterized in that, The extraction process of the interdisciplinary knowledge unit reference subnet is as follows: Traverse all directed edges in the interdisciplinary knowledge unit reference network, and determine whether they belong to interdisciplinary reference relationships based on the subject classification codes of the source node and the target node. If the source node belongs to the target subject and the target node belongs to a non-target subject, then retain the directed edge and its two endpoints. If the edge represents a reference within the target subject or a reference between external subjects, then remove it. After filtering, delete isolated nodes to form an interdisciplinary knowledge unit reference subnet.

8. The system according to claim 7, characterized in that, The network topology analysis analyzes the connectivity, sparsity, clustering, and structural distribution characteristics of the interdisciplinary knowledge unit referencing network and its subnets by analyzing the number of nodes, edges, average degree, average weighted degree, network density, network diameter, average path length, connected components, and modularity. This analysis is used to characterize the referencing structure of the interdisciplinary knowledge unit referencing network. Simultaneously, a network projection method is employed to construct the referencing node projection network and the cited knowledge unit node projection network of the interdisciplinary knowledge unit referencing subnets. The referencing node projection network indicates whether different referencing knowledge units jointly refer to the same cited knowledge unit, while the cited knowledge unit node projection network indicates whether different cited knowledge units are jointly referenced by the same referencing knowledge unit. By calculating the projection network density or average clustering coefficient, the degree of overlap in knowledge needs among referencing knowledge units and the co-reference relationships among cited knowledge units are analyzed, further characterizing the local clustering features of the interdisciplinary knowledge unit referencing subnets. The interdisciplinary influence analysis of knowledge units is used to identify core cited knowledge units and active citing knowledge units in the interdisciplinary knowledge unit citation network. The activity and coreness of nodes are measured by the in-degree, out-degree, weighted in-degree, and weighted out-degree of the nodes and directed edges in the interdisciplinary knowledge unit citation network. Specifically, in-degree represents the number of directed edges pointing to a node, used to measure the number of times a cited knowledge unit node is cited by different citing knowledge unit nodes; out-degree represents the number of directed edges from a node to other nodes, used to measure the activity level of citing cited knowledge unit nodes from external disciplines by citing knowledge unit nodes; weighted in-degree represents the sum of the weights of directed edges pointing to a node, used to measure the cumulative intensity of citations of a cited knowledge unit node; and weighted out-degree represents the cumulative intensity of citations of a node by a given node. The sum of the weights of directed edges pointing from a node to other nodes is used to measure the cumulative citation strength of the citing knowledge unit node to external subject knowledge. The weights of the directed edges are obtained by aggregating the citation frequencies of the same citing knowledge unit, the same cited knowledge unit, and the same citation function category. A webpage ranking algorithm is used to calculate the importance of the citing or cited knowledge unit node in the network, and the core knowledge units and basic theories, key methods, or important data on the core path of knowledge flow are judged based on the importance of the nodes. At the same time, a hyperlink-induced topic search algorithm is used to calculate the authority value and hub value of the citing knowledge unit node. The authority value is used to reflect the degree of recognition of the cited knowledge unit node as an external knowledge source, and the hub value is used to reflect the ability of the citing knowledge unit node to absorb and organize external knowledge. The cross-disciplinary citation pattern analysis of knowledge units is used to reveal the citation combination rules between knowledge units of different lexical function categories. Based on the lexical function categories of the citing and cited knowledge units, citation relationships are divided into six typical citation patterns: research problem citing research method, research problem citing data, research problem citing theory or concept, research method citing research method, research method citing data, and research method citing theory or concept. Each directed edge in the cross-disciplinary knowledge unit citation subnet is used as a statistical object. The lexical function categories of the source node and the target node of the directed edge are read and combined into a citation pattern. Based on each citation pattern, the number of unique patterns, the total number of citations, and the average citation intensity are calculated. The number of unique patterns is the number of combinations of non-repeating citing or cited knowledge unit entity names under this pattern; the total number of citations is the sum of the edge weights of all directed edges under this pattern; and the average citation intensity is the ratio of the total number of citations to the number of unique patterns. The interdisciplinary knowledge community identification method is used to discover sets of citing or cited knowledge units that are closely connected and have strong semantic relationships in the interdisciplinary knowledge unit citation network. A community discovery algorithm is used to divide the interdisciplinary knowledge unit citation network into multiple knowledge communities. The citing knowledge units with the highest out-degree, weighted out-degree, PageRank value, or hub value within each knowledge community are identified as core citing knowledge unit nodes. The cited knowledge units with the highest in-degree, weighted in-degree, PageRank value, or authority value within each knowledge community are identified as core cited knowledge unit nodes. The number of directed edges or the sum of edge weights corresponding to various citation function categories within a knowledge community are counted, and the citation function category with the highest proportion is identified as the main citation function of that community. The number of directed edges or the sum of edge weights corresponding to various chapter positions within a knowledge community are counted, and the chapter position with the highest proportion is identified as the main chapter position of that community. Subsequently, the theme and interdisciplinary integration type of the knowledge community are interpreted by combining the core citing knowledge unit nodes, core cited knowledge unit nodes, main citation functions, and main chapter positions. The interdisciplinary knowledge community dynamic evolution analysis is used to characterize the changing trends of knowledge integration patterns from a time dimension. This analysis constructs a dynamic monitoring framework based on a fixed community structure: using the core knowledge communities identified in network analysis as a benchmark, and according to the citation time attributes in the literature citation relationship dataset, the interdisciplinary knowledge unit citation network is divided into different time slices. By statistically analyzing the changes in citation intensity within each knowledge community under different time slices, the activity evolution curves of different interdisciplinary integration patterns are plotted, thereby capturing the macro-migration path of the research focus of the field. Then, the annual replacement of core cited knowledge units within each knowledge community is tracked, and the specific content and direction of knowledge unit flow are analyzed to reveal how an external knowledge unit gradually evolves from peripheral exploratory citations into a core infrastructure, or how a theoretical concept is reactivated and explosively grows under new technological conditions.

Images