Document duplicate checking method and device and medium

Abstract

The invention relates to the technical field of document duplicate checking. The document duplicate checking method comprises the steps that according to syntactic dependency structures in text information of a document to be subjected to duplicate checking and text information of a plurality of comparison documents, keywords and context dependency relationships of each sentence are recognized, a semantic map is constructed, and the semantic map is subjected to duplicate checking; mapping the keywords and the context dependency relationship to corresponding concept nodes in the semantic graph, constructing a semantic sub-graph of the document to be subjected to duplicate checking and a plurality of semantic sub-graphs of the comparison document, and comparing the structural similarity between the semantic sub-graph of the document to be subjected to duplicate checking and the semantic sub-graphs of the comparison document to obtain a structural similarity score; and based on a node sequence in the semantic subgraph of the document to be subjected to duplicate checking, constructing a disturbance semantic subgraph, calculating a disturbance score fluctuation range, screening a comparison document of which the structural similarity score is higher than a preset threshold value and the disturbance score fluctuation range is lower than a disturbance tolerance threshold value, and outputting duplicate checking difference information. The method and the device have the effect of improving the document duplicate checking accuracy.

Description

Technical Field

[0001] This invention relates to the technical field of document plagiarism detection, and in particular to a document plagiarism detection method, apparatus and medium. Background Technology

[0002] In existing technologies, common methods for calculating document similarity and plagiarism detection include: **Word frequency statistics:** This method counts the frequency of each word in a document and determines document similarity by comparing the similarity of word frequencies between different documents. For example, it calculates the percentage of identical words in two documents. **Keyword matching:** This method pre-defines keywords and checks the occurrence of these keywords in the document to be checked and the comparison documents to measure document similarity. **Text vector space model-based methods:** These methods convert documents into vector form and evaluate document similarity by calculating metrics such as the angle or distance between vectors.

[0003] However, existing document similarity calculation and plagiarism detection methods have significant drawbacks. They often focus only on the surface features of documents, such as word frequency and keyword matching, while ignoring the internal syntactic structure and semantic relationships. This leads to an inability to accurately detect document similarity when faced with rewritten or rearranged documents that are semantically similar, resulting in missed or false positives and failing to effectively achieve accurate document plagiarism detection. Summary of the Invention

[0004] To improve the accuracy of document plagiarism detection, this invention provides a document plagiarism detection method, apparatus, and medium.

[0005] The above-mentioned objective of this invention is achieved through the following technical solution: A document plagiarism detection method, the document plagiarism detection method comprising: Obtain the text information of the document to be checked for plagiarism and the text information of multiple comparison documents; Based on the syntactic dependency structure of the text information of the document to be checked and the text information of multiple comparison documents, identify the keywords and contextual dependency relationships of each sentence; Constructing a semantic graph; Keywords and contextual dependencies are mapped to corresponding concept nodes in the semantic graph to construct a semantic subgraph of the document to be checked and multiple semantic subgraphs of the comparison documents; The structural similarity score is obtained by comparing the structural similarity between the semantic subgraph of the document to be checked and the semantic subgraphs of multiple comparison documents. Based on the node order in the semantic subgraph of the document to be checked for plagiarism, a perturbation semantic subgraph is constructed, and the structural similarity between the perturbation semantic subgraph and the semantic subgraphs of multiple comparison documents is re-compared to calculate the fluctuation range of the perturbation score. The system filters out comparison documents whose structural similarity scores are higher than a preset threshold and whose perturbation scores fluctuate within a range lower than the perturbation tolerance threshold, and outputs information on the differences in plagiarism detection.

[0006] By adopting the above technical solution, the text information of the document to be checked and the text information of multiple comparison documents are obtained. The keywords and contextual dependencies of each sentence are identified and mapped to the constructed semantic graph. This further forms a semantic subgraph of the document to be checked and multiple comparison documents. Based on this, the structural similarity between the semantic subgraphs is compared and the stability analysis results of the perturbation semantic subgraph are combined to select comparison documents with true semantic consistency. This method can analyze the deep relationship between documents from the perspective of syntactic dependency structure and semantic level, no longer relying on single word frequency, keywords or surface vector representation. This effectively solves the problem of missed detection of rewritten text plagiarism due to ignoring syntactic structure and semantic association in existing technologies, and improves the accuracy and robustness of document similarity calculation and plagiarism detection.

[0007] Preferred method: Based on the syntactic dependency structure of the text information of the document to be checked and the text information of multiple comparison documents, identify the keywords and contextual dependency relationships of each sentence, including: Perform bidirectional path traversal on the syntactic dependency relation structure and calculate the semantic connection strength of each candidate keyword on the forward and reverse paths respectively; Based on semantic connection strength, words whose semantic connection stability is higher than the preset keyword recognition threshold are selected as keywords; Determine the corresponding contextual dependency relationship based on the connection pattern between the keyword and other words in the dependency path.

[0008] By employing the above technical solution, a bidirectional path traversal operation is performed on the syntactic dependency relation structure. The semantic connection strength of candidate keywords on the forward and reverse paths is calculated respectively. Based on the bidirectional observation results, a semantic connection stability evaluation index is constructed. From this, words with stability higher than the preset keyword recognition threshold are selected as keywords. Then, combined with the connection patterns between these keywords and other words in the dependency path, the contextual dependency relationship is determined. This approach can fully utilize the structural features and connection rules in the syntactic path to identify keywords with stable dominant semantic positions and their dependency relationships. This overcomes the problem in existing technologies that rely solely on word frequency or surface statistical methods and cannot effectively characterize the semantic master-slave structure within sentences. This improves the accuracy of semantic relationship extraction in document plagiarism detection under semantic reconstruction or word order adjustment.

[0009] Preferred steps for constructing a semantic graph include: Acquire terminology, semantic relationship samples between terms, and semantic rule base information corresponding to the target application domain; Constructing a set of concept nodes in a semantic graph based on terminology; Extract the term pairs that constitute the semantic relations from the semantic relation samples, and map the two terms in the term pairs to the corresponding source concept node and target concept node in the concept node set, respectively. Based on information from the semantic rule base, verify whether the semantic relationships represented by terms and words satisfy the preset semantic rule constraints. When the semantic relationship satisfies the preset semantic rule constraints, construct a semantic edge connecting the source concept node and the target concept node, and form a semantic edge set by multiple semantic edges that satisfy the constraints. A semantic graph that satisfies semantic constraints is constructed by combining a set of concept nodes and a set of semantic edges.

[0010] By adopting the above technical solution, we first acquire terminology, semantic relationship samples, and semantic rule base information corresponding to the target application domain, ensuring that the foundation for constructing the semantic graph closely aligns with specific contextual requirements. Secondly, we construct a set of concept nodes in the semantic graph based on the terminology. Then, we extract terminology pairs from the semantic relationship samples and map them to source and target concept nodes in the concept node set. We verify whether the semantic relationships meet the preset semantic rule constraints using the semantic rule base. Only when the constraints are met are semantic edges constructed, forming a set of semantic edges. Finally, we combine the set of concept nodes and the set of semantic edges to construct a semantic graph that satisfies the semantic constraints. This graph not only possesses rigorous semantic logic but also effectively expresses the semantic dependencies between concepts, thus solving the problem of lacking semantic standardization and semantic consistency verification when constructing graphs in existing technologies, and improving the accuracy and robustness of subsequent semantic mapping and structural similarity comparison.

[0011] Preferred method: Map keywords and contextual dependencies to corresponding concept nodes in the semantic graph, constructing a semantic subgraph of the document to be checked and multiple semantic subgraphs of the comparison documents, including: Based on the semantic information contained in the keywords, target nodes that match the descriptive information of concept nodes in the semantic graph are identified, and a mapping relationship is established between the keywords and the matched concept nodes to obtain a node mapping set; Based on the grammatical direction and dependency type indicated in the context dependency relation, determine the structural connection mode between any pair of concept nodes from the node mapping set, and retrieve the graph connection edges that conform to the structural connection mode in the semantic graph to construct the node connection relation set; Based on the set of node mappings and the set of node connection relationships, a graphical unit representing the semantic structure of a single sentence is generated. The graphical units are divided into sentences and assigned to the corresponding sentences of the document to be checked for plagiarism and multiple comparison documents, thus constructing a semantic subgraph of the document to be checked for plagiarism and a semantic subgraph of multiple comparison documents.

[0012] By adopting the above technical solution, firstly, based on the semantic information contained in the keywords, target nodes that match the descriptive information of concept nodes in the semantic graph are identified, and a mapping relationship between keywords and matching concept nodes is established to form a node mapping set. On this basis, according to the grammatical direction and dependency type indicated in the contextual dependency relationship, any pair of concept nodes is selected from the node mapping set to determine their structural connection mode, and graph connection edges that match the connection mode are retrieved in the semantic graph to form a node connection relationship set. Subsequently, the node mapping set and the node connection relationship set are combined to generate a graphical unit that can represent the semantic structure of a single sentence. Finally, according to the sentence division method, the graphical unit is assigned to the corresponding sentences of the document to be checked and multiple comparison documents, thereby constructing a semantic subgraph of the document to be checked and a semantic subgraph of multiple comparison documents. This process preserves the real semantic structural relationship between words while constructing the semantic subgraph, solving the problem that the semantic expression structure is one-sided and cannot cover the semantic association within the sentence in the existing technology, and providing a complete and semantically clear graphical foundation for subsequent similarity comparison.

[0013] Preferred method: Compare the structural similarity between the semantic subgraph of the document to be checked and the semantic subgraphs of multiple comparison documents to obtain a structural similarity score, including: Based on the semantic subgraph of the document to be checked for plagiarism and the semantic subgraphs of multiple comparison documents, a graph structure representation for structural similarity analysis is constructed. Based on the graph structure representation, calculate the node similarity and semantic edge connection similarity between the semantic subgraph of the document to be checked and each semantic subgraph of the comparison document; The structural similarity score is obtained by weighting the node similarity and the semantic edge connection similarity.

[0014] By adopting the above technical solution, a graph structure representation for structural similarity analysis is first constructed based on the conceptual nodes and semantic edge connections contained in the semantic subgraph of the document to be checked and the semantic subgraphs of multiple comparison documents. In this graph structure representation, the semantic content of each conceptual node and its position in the graph are clearly marked, and the connection direction, dependency type, and node connection order of the semantic edges are structurally expressed. Then, according to the graph structure representation, the node similarity and semantic edge connection mode similarity between the semantic subgraph of the document to be checked and each semantic subgraph of the comparison documents are calculated respectively. The node similarity is obtained by comparing the semantic description, lexical affiliation, and contextual consistency between conceptual nodes, and the semantic edge connection mode similarity is determined by comparing the connection direction, dependency path length, and syntactic function annotation of the semantic edges. On this basis, the two similarity values ​​are weighted and combined according to preset weights to obtain the structural similarity score. This process can effectively evaluate the dual similarity of the semantic subgraph in terms of node semantics and structural connections, avoid misjudgment or omission caused by single-dimensional matching, and improve the robustness and accuracy of similarity calculation when facing rewritten texts such as word order rewriting and grammatical reconstruction.

[0015] Preferred: A weighted combination of node similarity and semantic edge connection similarity is used to obtain a structural similarity score, including: The structural similarity score is obtained using the following formula: , of which S struct This refers to the structural similarity score, S node This refers to node similarity, S edge α refers to the similarity of semantic edge connection methods, β refers to the similarity weight of node similarity, α+β=1, and ε refers to the perturbation response coefficient, which ranges from 0 to 1.

[0016] By adopting the above technical solution, node similarity and semantic edge connection similarity are first used as two core dimensions to measure the structural consistency between the semantic subgraph of the document to be checked and the semantic subgraph of the comparison document. Then, node similarity weight and semantic edge similarity weight are introduced to reflect the contribution ratio of the two types of similarity in the overall structural matching process. Under the constraint that the weight parameters sum to one, the structural similarity is weighted and combined. Furthermore, to enhance the stability evaluation ability of the structural similarity index when facing node order perturbation, an additional perturbation response coefficient is introduced to adjust the response amplitude of structural similarity to perturbed semantic subgraph, thereby improving the sensitivity and adaptability of the similarity score to structural changes. The final structural similarity score integrates the node matching of semantic content expression, the consistency of syntactic connection mode, and the robustness performance after node order changes, providing a more stable and semantically deep structural comparison basis for subsequent plagiarism difference analysis.

[0017] Preferred method: Based on the node order in the semantic subgraph of the document to be deduplicated, a perturbed semantic subgraph is constructed. The structural similarity between the perturbed semantic subgraph and the semantic subgraphs of multiple comparison documents is re-compared, and the fluctuation range of the perturbation score is calculated, including: Based on node order, a perturbation semantic subgraph is constructed by exchanging node order within a limited range. The perturbation semantic subgraph is compared with the semantic subgraphs of multiple comparison documents to obtain multiple perturbation structural similarity scores; The range of perturbation score fluctuation is obtained by calculating the range of similarity scores of multiple perturbation structures.

[0018] By adopting the above technical solution, firstly, while keeping the set of concept nodes and the set of semantic edges unchanged, based on the node order in the semantic subgraph of the document to be deduplicated, a fixed node order exchange window is set. Within this window, adjacent node pairs are selected to perform an order exchange operation, thereby constructing multiple perturbed semantic subgraphs with only minor adjustments in node arrangement. This introduces structural perturbation while ensuring that the original semantic structure is not destroyed. Then, each perturbed semantic subgraph is compared with the semantic subgraphs of multiple comparison documents to obtain the corresponding perturbed structural similarity score set. The range of this score set is then calculated to evaluate the degree of structural similarity fluctuation caused by node order perturbation. Finally, a perturbation score fluctuation range index is formed to measure the stability of the original structural similarity score under order perturbation. This helps to identify the similar structure masking that may exist in the document while keeping the semantics unchanged, thereby improving the comprehensive judgment ability of plagiarism detection on semantic consistency and structural robustness.

[0019] The second objective of this invention is achieved through the following technical solution: A document plagiarism detection device, the document plagiarism detection device comprising: The text information acquisition module is used to acquire the text information of the document to be checked for plagiarism and the text information of multiple comparison documents; The syntactic dependency analysis module is used to identify the keywords and contextual dependency relationships of each sentence based on the syntactic dependency structure in the text information of the document to be checked and the text information of multiple comparison documents. The semantic graph construction module is used to construct semantic graphs; The semantic subgraph generation module is used to map keywords and contextual dependencies to corresponding concept nodes in the semantic graph, and to construct the semantic subgraph of the document to be checked and the semantic subgraph of multiple comparison documents; The structural similarity calculation module is used to compare the structural similarity between the semantic subgraph of the document to be checked and the semantic subgraphs of multiple comparison documents to obtain a structural similarity score; The perturbation fluctuation assessment module is used to construct a perturbation semantic subgraph based on the node order in the semantic subgraph of the document to be deduplicated, re-compare the structural similarity between the perturbation semantic subgraph and the semantic subgraphs of multiple comparison documents, and calculate the perturbation score fluctuation range. The similar document filtering and difference output module is used to filter comparison documents whose structural similarity scores are higher than a preset threshold and whose perturbation score fluctuation range is lower than the perturbation tolerance threshold, and output the plagiarism difference information.

[0020] By adopting the above technical solution, the text information of the document to be checked and the text information of multiple comparison documents are obtained. The keywords and contextual dependencies of each sentence are identified and mapped to the constructed semantic graph. This further forms a semantic subgraph of the document to be checked and multiple comparison documents. Based on this, the structural similarity between the semantic subgraphs is compared and the stability analysis results of the perturbation semantic subgraph are combined to select comparison documents with true semantic consistency. This method can analyze the deep relationship between documents from the perspective of syntactic dependency structure and semantic level, no longer relying on single word frequency, keywords or surface vector representation. This effectively solves the problem of missed detection of rewritten text plagiarism due to ignoring syntactic structure and semantic association in existing technologies, and improves the accuracy and robustness of document similarity calculation and plagiarism detection.

[0021] The above-mentioned objective three of the present invention is achieved through the following technical solution: A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the document plagiarism detection method described above.

[0022] The above-mentioned objective four of the present invention is achieved through the following technical solution: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the document plagiarism detection method described above.

[0023] In summary, the present invention has at least one of the following beneficial technical effects: 1. Obtain the text information of the document to be checked for plagiarism and the text information of multiple comparison documents, identify the keywords and contextual dependencies of each sentence, and map them to the constructed semantic graph. Further, form a semantic subgraph of the document to be checked for plagiarism and a semantic subgraph of multiple comparison documents. On this basis, compare the structural similarity between the semantic subgraphs and combine the stability analysis results of the perturbation semantic subgraph to screen out the comparison documents with true semantic consistency. It can analyze the deep relationship between documents from the perspective of syntactic dependency structure and semantic level, no longer relying on single word frequency, keywords or surface vector representation, thus effectively solving the problem of missed detection of rewritten text plagiarism due to ignoring syntactic structure and semantic association in the existing technology, and improving the accuracy and robustness of document similarity calculation for plagiarism detection. 2. First, node similarity and semantic edge connection similarity are used as two core dimensions to measure the structural consistency between the semantic subgraph of the document to be checked and the semantic subgraph of the comparison document. Then, node similarity weight and semantic edge similarity weight are introduced to reflect the contribution ratio of the two types of similarity in the overall structural matching process. Under the constraint that the weight parameters sum to one, the structural similarity is weighted and combined. Furthermore, in order to enhance the stability evaluation ability of the structural similarity index when faced with node order perturbation, an additional perturbation response coefficient is introduced to adjust the response amplitude of structural similarity to perturbed semantic subgraph, thereby improving the sensitivity and adaptability of the similarity score to structural changes. The final structural similarity score integrates the node matching of semantic content expression, the consistency of syntactic connection mode, and the robustness performance after node order changes, providing a more stable and semantically deep structural comparison basis for subsequent plagiarism difference analysis. 3. First, while keeping the set of concept nodes and the set of semantic edges unchanged, based on the node order in the semantic subgraph of the document to be checked, a fixed node order exchange window is set. Within this window, adjacent node pairs are selected and their order is swapped, thus constructing multiple perturbed semantic subgraphs with only minor adjustments in node arrangement. This introduces structural perturbation while ensuring that the original semantic structure is not destroyed. Then, each perturbed semantic subgraph is compared with the semantic subgraphs of multiple comparison documents to obtain the corresponding perturbed structural similarity score set. The range of this score set is then calculated to evaluate the degree of structural similarity fluctuation caused by node order perturbation. Finally, a perturbation score fluctuation range index is formed to measure the stability of the original structural similarity score under order perturbation. This helps to identify the similar structure masking that may exist in the document while keeping the semantics unchanged, thereby improving the comprehensive judgment ability of plagiarism detection on semantic consistency and structural robustness. Attached Figure Description

[0024] Figure 1 This is a flowchart of a document plagiarism detection method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a document plagiarism detection device according to an embodiment of the present invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to the accompanying drawings.

[0026] In one embodiment, such as Figure 1 As shown, this invention discloses a document plagiarism detection method, which specifically includes the following steps: S10: Obtain the text information of the document to be checked for plagiarism and the text information of multiple comparison documents.

[0027] Specifically, the process involves reading the electronic file of the document to be checked for plagiarism. An electronic file refers to a text document stored in a computer-recognizable format. The main text area is identified; this area contains the core content of the document, excluding auxiliary information such as tables of contents, footnotes, headers, and footers. The text is extracted in paragraph order based on paragraph markers (line breaks or formatting marks used to distinguish different paragraphs). Sentence boundaries within each paragraph are then identified using language segmentation rules. These rules are a set of rules for identifying sentence boundaries based on punctuation marks, functional words, and semantic connections within natural language text. Finally, all paragraph and sentence texts are combined in the extraction order to form the text information of the document to be checked for plagiarism. This text information refers to the main body of the document. The content area is extracted according to paragraph order and sentence boundaries to form a structured text set. The text information of multiple comparison documents is obtained from multiple comparison document files within a specified comparison range. Multiple comparison document files refer to several electronic files used as comparison objects. The main text area of ​​each comparison document file is read sequentially, and paragraph text is obtained according to the paragraph extraction method consistent with the text information of the document to be checked for plagiarism. Based on the same language punctuation rules, the paragraph text is divided into sentence text, which constitute the sentence-level text information of each comparison document. Sentence-level text information refers to the set of sentences extracted from the comparison document according to the language punctuation rules. The sentence-level text information of all comparison documents is combined in document order to form multiple comparison document text information, which maintains the same field type and content structure as the text information of the document to be checked for plagiarism.

[0028] S20: Based on the syntactic dependency structure of the text information of the document to be checked and the text information of multiple comparison documents, identify the keywords and contextual dependency relationships of each sentence.

[0029] Specifically, based on the text information of the document to be checked for plagiarism and the content of each sentence contained in the text information of multiple comparison documents, a preset syntactic dependency analysis tool is invoked to perform dependency analysis. The syntactic dependency analysis tool refers to a language processing component with lexical analysis, part-of-speech tagging, and syntactic relation recognition functions. Its function is to parse the input natural language sentence into a syntactic dependency structure composed of word nodes and dependency edges. A syntactic dependency structure is a graph structure where words in a sentence are used as word nodes, and directed dependency edges represent the grammatical control relationships between words. Each dependency edge connects a dependency child node to its parent dependency node. Dependency edges contain two types of information: dependency direction and dependency label. The dependency direction refers to the direction of the control relationship between word nodes, usually from the semantically dominant word to the semantically subordinate word, reflecting the grammatical master-subordinate relationship. The dependency label refers to the grammatical function category to which the dependency edge belongs, specifically including subject, predicate, object, attributive, adverbial, complement, etc., used to identify the role of the word node in the syntactic structure. For each sentence, dependency edge information between all word nodes and their corresponding parent nodes is extracted word by word. The starting word of the dependency edge is taken as the dependency child node, and the ending word as the dependency parent node. The dependency direction and dependency label are recorded. Based on this, grammatical roles are determined for all word nodes, identifying words in subject, predicate, object, attributive, or adverbial grammatical positions that have a content word part of speech, such as noun, verb, adjective, or adverb. Words meeting these conditions are marked as keywords. Keywords are word nodes that undertake the core semantic expression task and occupy a dominant grammatical position in the syntactic structure, used for subsequent semantic mapping and graph structure construction. Then, for each keyword, its contextual dependency relationship is identified based on the connection mode of the keyword in the syntactic dependency structure. The contextual dependency relationship refers to the dependency connection pattern between the keyword and other words in the syntactic dependency structure. This dependency connection pattern includes three items: the parent node word with which it has a dependency relationship, the dependency direction, and the dependency label, used to represent the grammatical control environment in which the keyword is located. Record the dependent parent node of each keyword, the dependency direction between the keyword and the parent node, and the syntactic function category of the dependency relationship as the contextual dependency relationship of the keyword.

[0030] S30: Construct a semantic graph.

[0031] Specifically, this involves acquiring terminology, semantic relationship samples between terms, and semantic rule base information corresponding to the target application domain. The target application domain refers to the professional technical field or knowledge classification scope to which the document to be checked belongs, such as law, medicine, education, and artificial intelligence. The target application domain is used to limit the source range of semantic resources required for semantic graph construction. Terminology refers to standardized expression units that have clear definitions and express specific concepts or operational meanings in the target application domain. Terminology usually consists of one or more words and includes categories such as noun terms, verb terms, and adjective terms. Semantic relationship samples refer to the set of terminology pairs and their semantic association tags generated by manual annotation or knowledge extraction in the target application domain. Each semantic relationship sample contains two terminology pairs and the semantic relationship category between them, such as "superior-subordinate," "causal," "chronological order," and "functional attribution." Semantic rule base information refers to the set of rules preset by domain experts to verify the validity of semantic relationships. This set of rules includes semantic relationship type definitions, concept node matching conditions, connection direction constraints, and context adaptation conditions. Each term in the terminology is treated as a candidate concept. A candidate concept is a term or lexical entity that has not yet been deambigued or semantically normalized. Term deduplication and semantic discrimination rules are applied to the candidate concepts. Term deduplication involves merging and normalizing terms with repetitive semantic expressions or spelling variations. Semantic discrimination rules are filtering rules that determine whether a term or lexical belongs to a specific semantic category and whether there are ambiguous terms. These rules are used to exclude semantically ambiguous or polysemous terms, ensuring semantic consistency. Terms with unique semantics and clear definitions are retained as concept nodes, forming a concept node set. The concept node set is the set of nodes in the semantic graph used to represent basic conceptual units. Each concept node corresponds to a term or lexical entity, supporting subsequent semantic connection operations. The node structure in the semantic graph is constructed using a set of concept nodes. The node structure refers to the logical representation of concept nodes in the semantic graph and its set of attributes, including node identifiers, semantic category labels, word meaning description information, and unique term codes, which are used to distinguish the semantic boundaries between different concepts. Term pairs that constitute semantic connections are extracted from semantic relationship samples. Each term pair contains a source term and a target term, indicating that the semantic connection direction is from the source term to the target term. The two terms in the term pair are matched in the set of concept nodes. If they can be mapped to the source concept node and the target concept node in the set of concept nodes respectively, the concept node mapping relationship of the term pair is considered to be valid.The semantic rule base information is used to find the semantic relationship type and connection conditions of the corresponding term pair. The semantic relationship type refers to the specific semantic connection category expressed between the source concept node and the target concept node, such as "contains," "uses," "evolutions," "belongs to," etc. The connection conditions refer to the logical constraints established by the semantic rule base for the semantic edge, such as "must come from the same domain," "must satisfy the hierarchical structure," "must conform to the directional rule," etc. When the term pair satisfies the semantic relationship type and connection conditions specified in the semantic rules, a semantic edge is established between the corresponding source concept node and the target concept node. A semantic edge is a directed edge connecting two concept nodes, used to represent the semantic relationship between them. The semantic edge contains attribute labels, connection direction, and relationship type information. Each semantic edge that meets the conditions is added to the semantic edge set. The semantic edge set is the set of all valid semantic edges in the semantic graph, recording the semantic connection methods between all concepts. The above operation is repeated until all term pairs that meet the semantic rule constraints have completed concept node mapping and semantic edge construction. Finally, the set of concept nodes is used as the vertex set, and the set of semantic edges is used as the edge set to form a semantic graph that meets the semantic rule constraints.

[0032] S40: Map keywords and contextual dependencies to corresponding concept nodes in the semantic graph, and construct a semantic subgraph of the document to be checked for plagiarism and multiple semantic subgraphs of the comparison documents.

[0033] Specifically, for each sentence in the text information of the document to be checked for plagiarism and the text information of multiple comparison documents, the semantic features and concept descriptions of each keyword are extracted. Semantic features refer to the set of semantic attributes consisting of the keyword's meaning, part of speech, semantic category, synonym set, hyponym set, and domain term tag, used to achieve semantic alignment in the semantic graph. Concept descriptions refer to the textual expression of the keyword's explanatory text, functional definition, scope of use, and semantic relationship background in its context, usually derived from professional terminology dictionaries, domain ontology, or expert annotations, used to assist in matching the semantic definition information of concept nodes. Based on the above semantic features and concept descriptions, target concept nodes that match the semantic content or description information of the keyword are searched in the concept node set of the semantic graph. Target concept nodes are concept nodes that form a semantic similarity match with the keyword in the semantic graph. Each node in the concept node set contains a concept identifier, term name, semantic category, unique term code, and word meaning definition, used to support high-precision concept recognition and semantic correspondence. Once the semantic similarity score threshold between keywords and target concept nodes is met, a one-to-one mapping relationship is established between keywords and target concept nodes. This mapping relationship refers to the semantic binding between a keyword and a concept node in the semantic graph, used to construct the node structure of the semantic subgraph. All successfully matched keywords and their target concept nodes are combined to form a node mapping set, which is an ordered set of mapping pairs between keywords and their matching concept nodes, serving as the node input set for the semantic subgraph. Further, for each keyword's contextual dependency relationship, information about other keywords associated with the keyword in the syntactic dependency structure is extracted. The dependency direction and dependency label information between the keyword and its dependency parent node are obtained. The dependency direction refers to the directional relationship of syntactic dependency, representing the connection direction between the controller and the controlled. The dependency label refers to the syntactic function type of the syntactic dependency relationship, such as subject-verb, verb-object, attributive-head, adverbial, etc. The dependency direction and dependency label between the current keyword and its context keywords are abstracted into a structural connection pattern. In the semantic edge set of the semantic graph, semantic edges that match the above structural connection patterns are retrieved. Each semantic edge in the semantic edge set contains a source concept node, a target concept node, a connection direction, a semantic relationship type, and edge attribute annotations. Structural connection pattern matching refers to retrieving semantic edges in the semantic graph by matching concept node pairs, semantic connection directions, and semantic relationship types simultaneously. After a successful match, the target concept node corresponding to the aforementioned keyword is used as a node in the semantic subgraph, and the successfully retrieved semantic edges are used as connecting edges in the semantic subgraph. Connecting edges are directed edges that represent the semantic structural relationship between nodes, forming the edge set of the semantic subgraph.Sequentially construct the semantic subgraph for each sentence in the document to be checked for plagiarism. Combine all the semantic subgraphs of the sentences and organize them in the order of the document structure to form a complete semantic subgraph of the document to be checked for plagiarism. Using the same node mapping strategy and semantic edge filtering rules, perform the above operations on each sentence in multiple comparison documents to construct the corresponding semantic subgraphs. Combine them in the order of the documents to form multiple semantic subgraphs of comparison documents. The semantic subgraphs of multiple comparison documents and the semantic subgraph of the document to be checked for plagiarism are consistent in terms of structural form, node attribute type and connection edge expression.

[0034] S50: Compare the structural similarity between the semantic subgraph of the document to be checked and the semantic subgraphs of multiple comparison documents to obtain a structural similarity score.

[0035] Specifically, structural similarity comparisons are performed on the semantic subgraphs of the document to be checked and multiple comparison documents. Structural similarity refers to the comprehensive similarity between two semantic subgraphs at both the semantic expression and structural connection levels, used to measure the correspondence between the document to be checked and the comparison documents in terms of conceptual expression and syntactic organization. First, concept nodes and semantic edge information are extracted from the semantic subgraph of the document to be checked and each comparison document's semantic subgraph. A concept node is a graph structure vertex in the semantic graph used to represent the semantic entities of terms; its attributes include node identifier, semantic category, word meaning definition, and unique term encoding. A semantic edge is a directed edge used to connect two concept nodes, expressing the semantic dependency relationship between them; its attributes include semantic relationship type, connection direction, semantic label, and contextual adaptation conditions. For each pair of semantic subgraphs, node similarity is first calculated. Node similarity refers to the degree of matching of semantic attributes between the concept nodes in the semantic subgraph of the document to be checked and the concept nodes in the semantic subgraph of the comparison documents. During the calculation, a node matching mechanism is used to align semantically similar concept nodes one by one, and node similarity is calculated based on the semantic distance between aligned node pairs. Semantic distance refers to the degree of difference between two concepts in the semantic space. The semantic distance is calculated using the cosine similarity or Jaccard coefficient between semantic embedding vectors. Cosine similarity assesses semantic direction consistency by calculating the cosine of the angle between two word vectors, while the Jaccard coefficient assesses semantic coverage by calculating the intersection and union ratio of the sets of hyponyms and synonyms of a concept. A higher node similarity value indicates that the corresponding nodes are semantically closer. Then, the semantic edge connection similarity is calculated. Semantic edge connection similarity refers to the similarity of semantic edges corresponding to structures in two semantic subgraphs along the dimension of connection features. The connectivity features include three aspects: connectivity direction, which refers to the directionality of the semantic edge from the source concept node to the target concept node; connectivity type, which refers to the category of semantic relationship expressed by the semantic edge, such as "superiority-subordinate," "causality," "containment," "belonging," etc.; and node pair positional relationship, which refers to the word order relationship between the two concept nodes connected by the semantic edge in the original syntactic dependency structure and the hierarchical relationship in the document structure. For each pair of semantic edges, structural matching is performed based on the above three dimensions, and a corresponding semantic edge similarity score is assigned. After normalizing the similarity results of all semantic edges, the semantic edge connection mode similarity is obtained. Finally, the node similarity and the semantic edge connection mode similarity are weighted and summed according to the set weight parameters to obtain the final structural similarity score.

[0036] S60: Based on the node order in the semantic subgraph of the document to be checked for plagiarism, construct a perturbation semantic subgraph, re-compare the structural similarity between the perturbation semantic subgraph and the semantic subgraphs of multiple comparison documents, and calculate the fluctuation range of the perturbation score.

[0037] Specifically, based on the original node order information of the concept nodes in the semantic subgraph of the document to be checked for plagiarism, a perturbation semantic subgraph set is constructed. The original node order information of the concept nodes refers to the linear sequence formed by the arrangement of concept nodes according to the sentence, paragraph, and semantic logical order in the original text during the construction of the semantic subgraph of the document to be checked for plagiarism. This node order not only reflects the semantic unfolding order in the text, but also affects the contextual connectivity of semantic edges, thus having a significant impact on structural similarity. In the process of constructing the perturbation semantic subgraph, the perturbation operation range is first defined. The perturbation operation range refers to the limited strategy that allows perturbation and adjustment of the concept node order within a certain range. This range is defined by the perturbation window size and the perturbation step size. The perturbation window size refers to the number of adjacent concept nodes participating in the order adjustment in each perturbation, and the perturbation step size refers to the interval unit by which the perturbation window slides in the concept node sequence. According to the preset perturbation strategy, two or more locally adjacent concept nodes are selected, and the node order exchange operation is performed to construct a new concept node arrangement sequence. After adjusting the node order, the set of concept nodes and the corresponding set of semantic edges are reorganized according to the adjusted node order to generate a perturbed semantic subgraph, called the perturbed semantic subgraph. The perturbed semantic subgraph maintains the same node and semantic edge set composition as the original semantic subgraph of the document to be checked for plagiarism, but the node arrangement order and some connection structures are adjusted due to changes in context, aiming to simulate the impact of different semantic expressions on structural similarity. For each constructed perturbed semantic subgraph, a structural similarity comparison operation with multiple comparison document semantic subgraphs is re-executed. The comparison method uses the same structural similarity calculation method as in S50, calculating the structural similarity score between each perturbed semantic subgraph and each comparison document semantic subgraph. The structural similarity scores of all perturbed semantic subgraphs under the same comparison document semantic subgraph are compiled to form a perturbed structural similarity sequence. Based on the perturbed structural similarity sequence, the perturbed score fluctuation range is calculated. The perturbed score fluctuation range refers to the numerical fluctuation range of the structural similarity score under the condition of node order perturbation, and this range is calculated by the difference between the maximum and minimum scores. The fluctuation range of the perturbation score is used to measure the stability of the semantic structure of the document. When the fluctuation range is small, it indicates that the semantic subgraph structure of the document to be deduplicated is relatively stable and the comparison of semantic organization methods has little impact on the results. When the fluctuation range is large, it indicates that the structure matching is highly dependent on the original expression order and there is sensitivity to the semantic expression method.

[0038] S70: Filter out comparison documents with structural similarity scores higher than a preset threshold and perturbation score fluctuation range lower than the perturbation tolerance threshold, and output plagiarism difference information.

[0039] Specifically, each semantic subgraph of the comparison document is traversed sequentially, and a dual threshold judgment operation is performed based on the fluctuation range of its corresponding structural similarity score and perturbation score. First, it is determined whether the structural similarity score corresponding to the semantic subgraph of the comparison document is greater than a preset threshold. Second, it is determined whether the fluctuation range of the corresponding perturbation score is less than the perturbation tolerance threshold. The preset threshold is the minimum critical value used to determine whether the structural similarity score meets the potential semantic duplication condition. This critical value is set based on the results of semantic graph matching experiments, the statistical distribution of real document duplication annotation samples, and the requirements of domain-specific plagiarism detection strategies. The value range is between 0 and 1, with higher values ​​indicating stricter structural matching standards. The perturbation tolerance threshold is the upper limit used to determine whether the fluctuation range of the perturbation score is within the acceptable range of semantic structural changes. This upper limit is determined based on the structural stability performance of the semantic subgraph under sequential perturbation conditions. It is usually set by analyzing the range distribution of perturbation scores through large-scale semantic perturbation experiments and combining it with fault tolerance strategies. When the structural similarity score is greater than the preset threshold and the fluctuation range of the perturbation score is less than the perturbation tolerance threshold, the comparison document corresponding to the semantic subgraph of the comparison document is marked as a candidate comparison document for plagiarism detection. For each candidate document to be compared for plagiarism detection, a plagiarism difference information extraction operation is performed. During this extraction, based on the established concept node mapping relationship and semantic edge matching structure between the semantic subgraph of the document to be checked and the semantic subgraph of the candidate document, the following three types of differences are identified: semantic path divergence information (where the concept node mapping is consistent but the path structure is different, specifically manifested as different semantic path construction methods between the same starting and ending nodes in the two semantic subgraphs); concept offset information (where, even with a concept node mapping relationship, the semantic role or contextual dependency of the keyword shifts, causing a change in semantic positioning); and semantic edge conflict information (where, in the two semantic subgraphs, the semantic edge relationship types between mapped concept node pairs are inconsistent, resulting in confusion between "causal relationship" and "subordinate relationship," or intersections of "chronological sequence" and "conditional transition," etc.). These three types of differences are encapsulated in a structured format to generate plagiarism difference information. The plagiarism detection discrepancy information refers to the structural semantic difference record obtained after detecting differences in three dimensions—concept nodes, semantic paths, and semantic edge relationships—during the semantic graph structure comparison process. This information includes concept node pair identifiers, path structure divergence descriptions, semantic role offset types, and relationship conflict classifications, supporting subsequent manual verification, machine interpretability report output, and visualization of plagiarism detection results.

[0040] In one embodiment, step S20, namely, identifying the keywords and contextual dependencies of each sentence based on the syntactic dependency structures in the text information of the document to be deduplicated and the text information of multiple comparison documents, includes: S201: Perform a bidirectional path traversal operation on the syntactic dependency relation structure and calculate the semantic connection strength of each candidate keyword on the forward and reverse paths respectively.

[0041] Specifically, a dependency graph is constructed for each sentence in the text information of the document to be checked for plagiarism and the text information of multiple comparison documents, based on the syntactic dependency relationship structure. A dependency graph is a directed graph structure built based on syntactic dependency relationships, where each node corresponds to a specific word, and each directed edge represents the syntactic dependency relationship between two words. The direction of the dependency edge is from the governing word to the dependent word, used to express the syntactic structural relationship between the controlling and the controlled. The dependency graph fully preserves the dependency path information generated during syntactic analysis, supporting subsequent path traversal operations on the syntactic structure of candidate keywords. After the dependency graph is constructed, for each candidate keyword, a forward path traversal operation is first performed. The forward path traversal operation means starting from the node where the candidate keyword is located, expanding the syntactic path layer by layer along the directed dependency edges, following the direction from the keyword to the dependent child node. On each forward path, the syntactic type, syntactic confidence, and path depth of each segment in the path are recorded sequentially. Syntactic type refers to the syntactic relation category to which the dependency edge belongs, such as "subject-predicate relation," "verb-object relation," "attributive-head relation," and "adverbial-head relation," used to label the function of the path segment in the grammatical structure. Syntactic confidence refers to the score given by the syntactic analysis tool on the reliability of the dependency relation between words; the value is usually between 0 and 1, with a higher value indicating a more reliable dependency relation. Path depth refers to the number of edges traversed from the candidate keyword node to the current dependency edge position, i.e., the path level. For each path segment in the path, the semantic connection weight of that segment is calculated. The semantic connection weight of a path segment is a weighted value reflecting the semantic contribution of the path segment in the entire path. It is calculated by multiplying the syntactic confidence of the path segment by the reciprocal of the path depth, reflecting the principle that the farther the path distance, the smaller the influence; and the higher the confidence, the greater the contribution. The semantic connection weights of all path segments along the same path are summed to obtain the total semantic connection strength of that path. Then, the semantic connection strengths of all forward paths starting from the candidate keyword are aggregated to form the forward path semantic connection strength of that candidate keyword. Subsequently, starting from the same candidate keyword node, a reverse path traversal operation is performed. The reverse path traversal operation refers to traversing backward along the dependency edges from the dependency parent node to the candidate keyword node, sequentially searching for paths on all syntactic dominance chains. During the reverse traversal, the syntactic type, syntactic confidence, and path depth of each path segment are recorded in the same way as the forward path, and the semantic connection weights of the path segments are calculated. The semantic connection weights of all reverse path segments are summed to obtain the reverse path semantic connection strength of that candidate keyword. Finally, the forward path semantic connection strength and the reverse path semantic connection strength of each candidate keyword are output separately. These two semantic connection strength values ​​together serve as a quantitative indicator of the degree of semantic adhesion between the keyword and its context in the syntactic dependency structure.

[0042] S202: Based on semantic connection strength, select words whose semantic connection stability is higher than the preset keyword recognition threshold as keywords.

[0043] Specifically, for each candidate keyword, the forward path semantic connection strength and the reverse path semantic connection strength corresponding to that candidate keyword are obtained. The forward path semantic connection strength and the reverse path semantic connection strength together constitute the bidirectional observation value of the candidate keyword's semantic connection strength, used to evaluate the candidate keyword's ability to form stable semantic connections with the context in the syntactic dependency structure. After obtaining the bidirectional observation value of semantic connection strength, a semantic connection stability evaluation index for the candidate keyword is calculated. The semantic connection stability evaluation index is a numerical evaluation parameter used to measure semantic stability, calculated by jointly using the forward path semantic connection strength and the reverse path semantic connection strength as inputs, and the mean and range. The mean is the arithmetic mean of the forward path semantic connection strength and the reverse path semantic connection strength, used to measure the connection strength level of the candidate keyword in the overall grammatical structure, reflecting its ability to form semantic adhesion with the context. The range is the absolute difference between the forward path semantic connection strength and the reverse path semantic connection strength, reflecting the semantic connection consistency of the candidate keyword in different directions, used to determine whether the word has directional stability. Using the mean and range as a joint evaluation index, a semantic connection stability evaluation index of the following form is constructed: Where μ represents the mean semantic connection strength, δ represents the range of semantic connection strength, and λ is a weighting coefficient used to adjust the influence of semantic fluctuations on the overall stability score. A higher semantic connection stability evaluation index indicates that the candidate keyword exhibits stable and strong connection characteristics in both bidirectional paths, and possesses greater semantic identification capabilities. The semantic connection stability evaluation index is numerically compared with a preset keyword identification threshold. The preset keyword identification threshold is a numerical critical value used to determine whether candidate words possess sufficient semantic connection stability. This threshold is set based on the language complexity of the target application domain, semantic consistency requirements, and the statistical results of the connection stability distribution of known keywords in the annotated text. The preset keyword identification threshold is used to control the keyword selection criteria and improve the semantic representativeness and structural reliability of keyword judgment. When the semantic connection stability evaluation index corresponding to a candidate keyword is greater than the preset keyword identification threshold, the candidate keyword is determined to have the stability characteristics to become a keyword and meets the keyword judgment conditions. The candidate keyword is then added to the keyword set.

[0044] S203: Determine the corresponding contextual dependency relationship based on the connection pattern between the keyword and other words in the dependency path.

[0045] Specifically, for each keyword in the text information of the document to be checked for plagiarism and the text information of multiple comparison documents, all connection path information of the keyword in the dependency graph is obtained, and the word nodes directly connected to the keyword in each dependency path are traversed sequentially. For each connection path, the following three structural elements are extracted: path type, connection direction, and path length, which are used to construct the semantic connection structure between the keyword and other words. Path type refers to the semantic category label of each syntactic connection in the dependency path, mainly including subject-predicate relation, verb-object relation, attributive-head relation, adverbial-head relation, complement-subject relation, etc., which are used to describe the subordinate relationship and functional positioning of the keyword and the context words in the syntactic structure. Connection direction refers to the position and direction of the keyword in the dependency path, including the forward connection direction from the keyword to the dependency child node and the reverse connection direction from the dependency parent node to the keyword node, which is used to clarify the control flow direction in the semantic structure. Path length refers to the number of dependency edges traversed from the keyword to the target word node in the path, which is used to measure the syntactic distance of the semantic connection. The smaller the value, the closer the semantic association between the keyword and the word. Based on the above structural elements, the semantic hierarchy between the keyword and its connected words is further determined. Semantic hierarchy refers to determining whether the keyword plays a controller (subject, predicate) or a controlled (object, modifier) ​​role in the syntactic structure, relying on the path type and connection direction information in the dependency path. For example, when the path type is "subject-predicate relationship" and the keyword is the subject node, the keyword is the agent of the action in that path; when the path type is "verb-object relationship" and the keyword is in the object position, it is the receiver of the action. Based on identifying the semantic hierarchy, the semantic role of the keyword in that path is further determined. A semantic role refers to the abstract semantic function that the keyword undertakes in the contextual semantic structure. Common semantic roles include "agent" (the agent of the action), "patient" (the receiver of the action), "modifier" (modifying the headword), and "adverbial" (modifying the time, place, manner, etc. of the action). By combining the path type and the semantic hierarchy, the semantic role of the keyword in the path can be uniquely determined. This process categorizes the connection patterns of the same keyword across multiple dependency paths, analyzing its semantic role types and structural features in different paths to complete the semantic role classification of the keyword. Semantic role classification involves structurally summarizing the semantic roles undertaken by the keyword in multiple paths to form a stable semantic feature expression of the keyword in the syntactic structure. Finally, the contextual dependency relationship of each keyword is output.Contextual dependency relations refer to the semantic connections between keywords and contextual words identified based on dependency path structures. Structurally, they include three parts: dependency type (the syntactic structure category corresponding to the path type), path structure (the semantic connection topology determined by the connection direction and path length), and semantic role (the semantic function of the keyword within the dependency path). The generated contextual dependency relations serve as the semantic connection basis for constructing the semantic subgraph of the document to be checked and multiple comparison documents. The semantic connection basis refers to the structured set of semantic connections formed based on the dependency path information between keywords and contextual words, used to construct connection edges between concept nodes in the semantic graph.

[0046] In one embodiment, step S30, namely the step of constructing a semantic graph, includes: S301: Obtain terminology, semantic relationship samples between terms, and semantic rule base information corresponding to the target application domain.

[0047] Specifically, the process begins by accessing the terminology management database based on pre-defined target application domain identifiers. Semantic resource entries related to the target application domain are then extracted from this database. Target application domain identifiers are unique identifiers, names, or sets of keywords used to identify the target technology area, such as "natural language processing," "power system dispatching," and "medical image diagnosis," which are used to quickly locate the target domain within the semantic resource database. The terminology management database is a structured database system that stores and manages cross-domain terms and their semantic descriptions, supporting searches based on domain tags, semantic categories, and usage frequency. Semantic resource entries are single semantic description records of terms stored in the terminology management database, including the terminology ontology, semantic annotations, usage scenarios, and relational annotations. After accessing the semantic resource entries, a domain keyword search method is used to filter out a set of terms that meet grammatical and semantic integrity requirements. This domain keyword search method utilizes core keywords contained in the target application domain identifiers, employing full-text matching, lemmatization, and semantic expansion retrieval to filter terms from the terminology management database. A terminology vocabulary set refers to a group of terms retrieved from a database, including common technical terms, professional terms, and standardized definitions within the target domain. Each term has a unique term code, semantic category label, and definition text description. Subsequently, the semantic relationship sample between terms is retrieved from the terminology management database or external knowledge sources by calling the semantic relationship interface. The semantic relationship interface is a functional interface used to retrieve and parse semantic connections between terms, supporting term pairs as input and returning corresponding semantic relationship categories and explanations. A semantic relationship sample is a semantic data unit composed of term pairs and semantic relationships. Common relationship types include synonymy, hyponymy / hypernymy, causal, and temporal relationships. Each semantic relationship sample contains explicit term pairs, relationship types, and semantic explanation information. After obtaining the semantic relationship sample, the corresponding semantic rule base information is loaded through the rule index path. The rule index path is a path identification method used to locate a specific rule set in the semantic management system, typically consisting of rule category, rule number, and storage location. Semantic rule base information refers to a set of semantic rules pre-annotated and abstracted by domain experts, containing three types of rules: term classification rules, used to define the conceptual hierarchy and category division of terms in the semantic graph; concept merging rules, used to merge semantically repetitive or similar term nodes to ensure consistency of concept representation; and semantic edge constraint rules, used to restrict the connection conditions of semantic edges in semantic relation samples, such as requiring that "causal relationship" must be established between action-type terms and result-type terms. Ultimately, the acquired term vocabulary set, semantic relation samples, and semantic rule base information serve as the input basis for constructing the semantic graph, supporting subsequent concept node construction and semantic edge generation.

[0048] S302: Construct a set of concept nodes in a semantic graph based on terminology.

[0049] Specifically, the first step is to perform field standardization on the acquired terminology. Field standardization addresses textual inconsistencies by performing three operations according to a unified standard: standardizing capitalization (converting all terms to a uniform capitalization format to avoid duplication due to case differences), removing special characters (symbols, punctuation, and control characters not essential for semantic expression to ensure the purity of the terminology text), and normalizing language style (converting synonyms, abbreviations, or colloquial expressions into standardized written expressions to ensure semantic consistency). After field standardization, a professional terminology database matching method based on target application domain identifiers is used to verify the validity and semantic classification of the terminology. This method compares the standardized terminology with a professional terminology database corresponding to the target application domain to confirm its validity and assign a semantic classification label and domain attribute to each valid term. Semantic classification labels refer to the concept categories to which terms belong in the semantic graph, such as "equipment terms," ​​"action terms," ​​and "time terms," ​​used to distinguish different semantic levels. Domain attributes refer to the specific application domain to which the terms belong, such as "power system" and "natural language processing," used to enhance the domain adaptability of the semantic graph. After completing validity verification and semantic classification labeling, each term is encapsulated into a concept node structure. A concept node structure is a data unit composed of a term name, semantic classification label, domain attribute, and unique term code, used to represent a unique concept entity in the semantic graph. Specifically, the term name displays the textual expression of the concept node, the semantic classification label describes the category characteristics of the concept, the domain attribute indicates the application scenario to which the term applies, and the unique term code enables unique identification and rapid indexing of the concept node. Finally, all concept node structures are uniformly added to a concept node set. The concept node set is the total set of concept nodes in the semantic graph, used to represent the basic semantic entity units in the semantic graph. The concept node set maintains a one-to-one correspondence between terms and concept nodes, ensuring that the semantic graph can fully cover the terminology expressions of the target application domain and providing a conceptual foundation for the subsequent construction of semantic edge sets.

[0050] S303: Extract the term pairs that constitute the semantic relations from the semantic relation samples, and map the two terms in the term pairs to the corresponding source concept node and target concept node in the concept node set, respectively.

[0051] Specifically, each semantic relation sample is first parsed to extract its semantic triple structure. A semantic triple structure is a data representation consisting of three parts: "term starting point—semantic relation type—term ending point." The term starting point represents the initiator of the relation, the term ending point represents the receiver, and the semantic relation type represents the semantic relationship category between the two, such as "synonym," "hypernym / hypernym," "causality," or "temporal sequence." By parsing the semantic triple structure, the term pairs constituting the relation can be directly obtained, i.e., the combination of the semantic starting point term and the semantic ending point term. After obtaining the term pairs, a concept node mapping operation is performed on each term in the pair. Concept node mapping involves searching for node entities in the concept node set that match the term. The search criteria include two dimensions: term encoding and semantic classification label. Term encoding refers to a unique identifier for a term within the semantic resource system, used for precise matching of concept nodes. Semantic classification label refers to the semantic category to which a term belongs in the semantic graph. When a term encoding is missing or a semantic variant exists, matching can be achieved through an equivalent semantic classification label. If both terms in a term pair successfully match in the set of concept nodes, the concept node corresponding to the starting term in the semantic triple structure is set as the source concept node, and the concept node corresponding to the ending term in the semantic triple structure is set as the target concept node. The source concept node is the concept node that initiates the semantic relation in the semantic direction, and the target concept node is the concept node that receives the semantic relation in the semantic direction. The relationship between the source and target concept nodes is determined by the type of semantic relation in the semantic triple. While establishing the binding relationship between the source and target concept nodes, it is also necessary to record the original order information and semantic direction information of the term pair in the semantic relation sample. The original order information refers to the sequential position of the term pair in the semantic relation sample, used to maintain consistency with the input sample. The semantic direction information refers to the directional identifier from the source concept node to the target concept node in the semantic relation, used to ensure that the semantic relation remains consistent with the original semantic dependency structure when constructing semantic edges. Finally, through the above mapping and recording operations, each term pair in the semantic relation sample is accurately bound to the source concept node and the target concept node in the concept node set, providing structured input data for the subsequent construction of the semantic edge set.

[0052] S304: Based on semantic rule base information, verify whether the semantic relationships represented by terms and words satisfy the preset semantic rule constraints.

[0053] Specifically, the process begins by iterating through the term pairs extracted from the semantic relation samples. For each term pair, a retrieval operation is performed within the semantic rule base. The semantic rule base is a knowledge set comprised of multiple semantic rule entries. Each semantic rule entry is a structured unit recording semantic relation constraints, including allowed semantic relation types, the semantic category to which the terms belong and their combination, the direction of semantic connection, and applicable contextual conditions. After retrieving the semantic rule entries, a matching and verification operation is performed between the term pairs and the semantic rule entries. The matching verification operation refers to comparing term pairs item by item according to the constraints defined in the semantic rule entries. Specifically, it includes the following four aspects: First, verifying whether the semantic relationship represented by the term pair belongs to the allowed semantic relationship types recorded in the semantic rule entries. Second, verifying whether the source term and target term in the term pair belong to the term category combinations defined in the semantic rule entries. The term category refers to the category into which term terms are divided according to semantic features, such as "entity category", "action category", and "attribute category". Third, verifying whether the semantic connection direction in the term pair is consistent with the direction required in the semantic rule entries. The semantic connection direction refers to the directional definition from the source term to the target term, used to limit the directional relationship of the semantic dependency structure. Fourth, verifying whether the applicable context conditions set in the semantic rule entries are consistent with the semantic relationship sample context in which the term pair is located. The applicable context conditions refer to the set of conditions that restrict the establishment of semantic relationships in a specific semantic scenario, such as semantic rules that are limited to "medical text", "legal text", or "technical document". When a term pair satisfies the matching requirements under all the above semantic rule entries, the term pair is determined to be a valid term pair and used as the input basis for constructing the semantic edge set. If the term pair does not meet the requirements under any semantic rule entry, the term pair is marked as a term pair that does not meet the rule constraints and will not be used to generate the semantic edge set in subsequent processing.

[0054] S305: When the semantic relationship satisfies the preset semantic rule constraints, construct a semantic edge connecting the source concept node and the target concept node, and form a semantic edge set by multiple semantic edges that satisfy the constraints.

[0055] Specifically, after verifying the semantic rule constraints for each pair of terms, it is necessary to first filter out all term pairs that satisfy the semantic rule constraints. For each pair of terms that satisfies the semantic rule constraints, the source concept node and target concept node corresponding to the term pair need to be located in the concept node set. The source concept node is the concept node mapped from the source term in the term pair, and the target concept node is the concept node mapped from the target term in the term pair. A directed semantic edge needs to be constructed between the source concept node and the target concept node. A directed semantic edge is a semantic connection structure with a direction, pointing from the source concept node to the target concept node. The following information is appended to directed semantic edges: First, the semantic relationship type, i.e., the specific semantic relationship category represented by the term pair, such as synonymy, hyponymy / hypernymy, causal relationship, or temporal relationship; second, the connection direction, i.e., the directional identifier from the source concept node to the target concept node, used to clarify the flow of semantic dependencies; third, the context condition identifier, which is an identifier used to mark the context environment required for the semantic relationship to be established, mapped from the applicable context conditions defined in the semantic rule base information; and fourth, a unique relationship number, which is a unique identification code assigned to each semantic edge, used to track the source of the semantic edge and the corresponding semantic rule entry. Each directed semantic edge that has been constructed and satisfies the semantic rule constraints is sequentially stored in a semantic edge set. The semantic edge set is a collection structure composed of multiple semantic edges, used to systematically store the semantic relationships between concept nodes. Ultimately, the set of semantic edges and the set of concept nodes together form the basic structure of the semantic graph. The semantic graph is a graph structure composed of the set of concept nodes and the set of semantic edges, used to represent the semantic relationships between terms in the target application domain, and serves as the input basis for subsequent semantic subgraph mapping and structural similarity calculation operations.

[0056] S306: Construct a semantic graph that satisfies semantic constraints by combining a set of concept nodes and a set of semantic edges.

[0057] Specifically, each concept node in the concept node set is initialized as a vertex element and stored in the graph structure. A vertex element is a basic node unit used to represent a semantic entity in the graph structure, which originates from the standardized and encapsulated concept node information in the concept node set. After initialization, each semantic edge in the semantic edge set is added to the graph structure as a directed edge connecting the source concept node and the target concept node. A directed edge is an edge element in the graph structure that has a direction attribute, with its direction pointing from the source concept node to the target concept node. The process of adding semantic edges requires configuration based on several attributes contained within the semantic edge: First, the attribute labels of directed edges are set according to the semantic relationship type, which refers to the specific semantic category represented by the semantic edge, such as synonymy, hyponymy / hypernymy, causal relationship, or temporal relationship. Second, the pointing order of directed edges is set according to the connection direction, which refers to the unidirectional identifier from the source concept node to the target concept node. Third, the applicable scope and semantic restrictions of directed edges are configured according to the context condition identifier, which refers to the identifier mapped from the context adaptation conditions in the semantic rule base information, used to ensure that the semantic relationship is valid in the applicable context. After completing the addition and attribute configuration of semantic edges, connectivity analysis needs to be performed on the entire semantic graph. Connectivity analysis refers to detecting whether there is path reachability between all concept nodes in the graph structure, and whether any semantic path follows the logical deduction order and semantic consistency requirements defined in the semantic rule base. A semantic path is a sequence of paths composed of concept nodes and semantic edges, used to describe the multi-step semantic association between concepts. Through connectivity analysis, it can be confirmed whether different combinations of semantic relations in the semantic graph form closed links. A closed link refers to a path structure in which both the starting and ending points can form a logically closed path structure under the constraints of semantic rules, which is used to ensure the reachability and consistency of semantic reasoning. Finally, by combining the set of concept nodes and the set of semantic edges, the completed semantic graph is a graph structure that satisfies the constraints of semantic rules. The semantic graph has semantic integrity, logical coherence, and reasoning closure, which is used to support subsequent semantic subgraph mapping and structural similarity calculation operations.

[0058] In one embodiment, step S40 involves mapping keywords and contextual dependencies to corresponding concept nodes in a semantic graph, constructing a semantic subgraph of the document to be deduplicated and multiple semantic subgraphs of the comparison documents, including: S401: Based on the semantic information contained in the keywords, identify the target nodes that match the descriptive information of the concept nodes in the semantic graph, and establish a mapping relationship between the keywords and the matched concept nodes to obtain a node mapping set.

[0059] Specifically, the first step is to extract semantic information from each keyword in the document to be checked or multiple comparison documents. This semantic information includes part-of-speech tagging, semantic roots, hypernyms and hyponyms, domain terminology tags, and contextual semantic vectors. Part-of-speech tagging refers to the word class attribute assigned to keywords using natural language processing tools, such as noun, verb, or adjective. Semantic roots are the core semantic units obtained after word form reconstruction of keywords, used to unify different forms of lexical expression. Hypernyms and hyponyms are related words with hierarchical semantic relationships to keywords, including hypernyms and hyponyms. Domain terminology tags are domain affiliation identifiers assigned to keywords using a domain-specific terminology database, used to represent the application domain to which the keywords belong. Contextual semantic vectors are vectorized representations trained based on a context window, used to quantify the semantic distribution characteristics of keywords in context. After obtaining the semantic information of the keywords, a multi-dimensional feature matching model is constructed. This model is a computational model that vectorizes and combines multiple semantic attributes of keywords to form a unified matching space. This model is used to match and compare the semantic information of keywords with the descriptive information of concept nodes in a semantic graph. The descriptive information refers to the set of attributes stored in the concept node, including term name, semantic category, domain attributes, and unique term encoding. During the matching process, a semantic similarity scoring mechanism is employed. This mechanism generates a numerical similarity score by comprehensively calculating the closeness of multi-dimensional features between keywords and concept nodes. Specifically, the implementation includes: First, using word embedding vector distance calculation to measure the spatial distance between the semantic vector of the keyword context and the vector representation of the concept node's descriptive information. The distance calculation method can use the cosine similarity formula. Second, executing a word sense disambiguation strategy to select the word sense that best fits the context from multiple possible semantic interpretations. This strategy involves selecting a semantic interpretation based on context windows, syntactic dependencies, or semantic rules. Third, performing contextual consistency determination. This determination confirms the reasonableness of the match by comparing whether the text segment containing the keyword matches the domain attributes of the concept node. After calculating the similarity score, the score is compared with a preset matching threshold. The preset matching threshold is a pre-defined numerical standard used to determine whether the semantic similarity reaches an acceptable matching level. When the similarity score between a keyword and a certain concept node exceeds the preset matching threshold, the mapping relationship between the keyword and the concept node is confirmed. Finally, all confirmed valid mapping relationships between keywords and target concept nodes are recorded and formed into a node mapping set. The node mapping set is a data set composed of multiple mapping pairs. Each mapping pair consists of keywords from the document to be checked or multiple comparison documents and the corresponding target concept node in the semantic graph. This node mapping set serves as the input for subsequent construction of the semantic subgraph.

[0060] S402: Based on the grammatical direction and dependency type indicated in the context dependency relation, determine the structural connection method between any pair of concept nodes from the node mapping set, and retrieve the graph connection edges that conform to the structural connection method in the semantic graph to construct a set of node connection relations.

[0061] Specifically, the process begins by reading the contextual dependency relationships between each keyword and its connected words. These relationships include grammatical direction information and dependency type information. Grammatical direction information refers to the directional relationship between the keyword and its connected words in the dependency path, including forward relationships from the keyword to dependent child nodes and backward relationships from dependent parent nodes to the keyword, used to describe the directionality of semantic dependencies. Dependency type information refers to the syntactic dependency categories between the keyword and its connected words in the dependency path, such as subject-verb, verb-object, and attributive-head relationships, used to annotate the grammatical structural features between words. After obtaining the grammatical direction and dependency type information, the dependency path structure between the keyword and its connected words is extracted. The dependency path structure is a structured representation composed of the keyword, connected word nodes, and dependency relationship elements between them (including grammatical direction, dependency type, and path length), used to describe the dependency relationship patterns between the keyword and other words in a sentence. Subsequently, based on the node mapping set, the source concept node corresponding to the keyword and the target concept node corresponding to the connected words are retrieved. Source concept nodes are concept nodes mapped from keywords, while target concept nodes are concept nodes mapped from words connected to keywords. Based on grammatical direction information, the structural connection method between source and target concept nodes is determined. The structural connection method refers to the direction and type of connection between concept nodes. Direction can be unidirectional or bidirectional, and the type is determined by dependency type information, which specifies the semantic relation category, such as subject-predicate, verb-object, or attributive-head. After determining the source concept node, target concept node, and structural connection method, these three are combined to form a retrieval triple. A retrieval triple is a structured query condition composed of the source concept node, target concept node, and structural connection method, used to locate semantic connection edges that meet the requirements in the semantic graph. Based on the retrieval triple, a structural pattern matching operation is performed in the semantic graph. The structural pattern matching operation involves searching and comparing the connection patterns of all semantic edges in the semantic graph according to the source node, target node, and semantic relation conditions set in the retrieval triple, and filtering out semantic edges that satisfy the semantic relation category, connection direction, and context conditions. Successfully matched semantic edges are called graph connection edges. Graph connection edges refer to directed edges that actually exist in the semantic graph and satisfy specific semantic rule constraints between concept nodes. Finally, all successfully matched graph connection edges are organized and stored as a set of node connection relationships. The set of node connection relationships is a data set consisting of multiple concept node pairs and their corresponding graph connection edges. It is used to represent the edge information in the semantic subgraph and serves as the basic input for subsequently constructing the semantic subgraph of the document to be checked and multiple comparison document semantic subgraphs.

[0062] S403: Generate a graphical unit representing the semantic structure of a single sentence based on the set of node mappings and the set of node connection relationships.

[0063] Specifically, the mapping relationship between keywords and their corresponding concept nodes is first obtained from the node mapping set. The mapped concept nodes are then stored as semantic nodes in the graph unit of a single sentence. A semantic node is a concept node determined by semantic matching of keywords. It is used to represent the basic semantic elements in the semantic graph unit. Each semantic node contains a unique node identifier, semantic category, and domain attributes.

[0064] Subsequently, the graph connection edges between each pair of concept nodes are read from the set of node connection relationships, and these graph connection edges are added to the graph unit as semantic edges. A semantic edge is a directed connection structure formed by the semantic relationship between a source concept node and a target concept node, used in the graph unit to represent the contextual dependency relationship between semantic nodes. Semantic edges include attributes such as edge type, connection direction, and semantic relationship number. The edge type refers to the category of semantic relationship corresponding to the semantic edge, such as subject-predicate, verb-object, or attributive-head relationship, used to annotate the role of the semantic edge in the grammatical and semantic structure. The connection direction refers to the pointing order of the semantic edge between the source and target concept nodes, divided into unidirectional or bidirectional connections, used to clarify the directionality of semantic dependencies. After the semantic nodes and semantic edges are determined, they are combined sequentially according to the grammatical structure order of the original sentence to form a semantic graph structure. A semantic graph structure is a directed graph composed of multiple semantic nodes and semantic edges in syntactic dependency order, with each semantic graph structure corresponding to the semantic expression pattern of a sentence. Finally, the semantic graph structure formed by combining semantic nodes and semantic edges is used as the semantic graph unit. A semantic graph unit is a graphical representation that fully represents the semantic mapping and contextual structure of a single sentence. It includes three core elements: node identifier, edge type, and connection direction. It can reflect the role distribution and contextual dependencies of keywords in the semantic hierarchy. The generated semantic graph unit will serve as the basic input for subsequently constructing the semantic subgraph of the document to be checked for plagiarism and the semantic subgraphs of multiple comparison documents.

[0065] S404: Divide the graphic units into sentences and assign them to the corresponding sentences of the document to be checked for plagiarism and multiple comparison documents, and construct the semantic subgraph of the document to be checked for plagiarism and the semantic subgraph of multiple comparison documents.

[0066] Specifically, firstly, based on the sentence division order in the text information of the document to be checked and the text information of multiple comparison documents, the semantic graph units generated in the previous step are sequentially mapped and matched with their corresponding original sentences. A semantic graph unit is a graph structure that fully represents the semantic structure of a single sentence, generated by combining a set of node mappings and a set of node connection relationships, containing semantic nodes and semantic edges. After completing the mapping between semantic graph units and sentences, the semantic graph units belonging to each sentence in the document to be checked are arranged sequentially according to the order of the sentences in the document, and a semantic graph structure merging operation is performed. The semantic graph structure merging operation means connecting the semantic graph units of adjacent sentences through document order indexes while keeping the unique identifiers of semantic nodes unchanged, and retaining all semantic edges in a unified graph structure, thereby forming a semantic subgraph of the document to be checked. The semantic subgraph of the document to be checked is a graph structure formed by merging all semantic graph units in the document to be checked, fully representing the mapping relationships of the document at the sentence-level semantic and overall semantic organization levels. Subsequently, the semantic graph units belonging to each sentence in each comparison document are arranged sequentially according to the order of the sentences in the comparison document, and the same semantic graph structure merging operation is performed to generate semantic subgraphs for each comparison document. A semantic subgraph for a comparison document refers to a semantic graph structure composed of all semantic graph units in a particular comparison document, representing the organizational pattern of the comparison document at the sentence-level and overall semantic levels. After construction, each semantic subgraph contains the following elements: concept nodes, which are semantic nodes generated by the mapping relationship between keywords and concept nodes in the semantic graph, representing the basic semantic entities in the sentence; graph connection edges, which are semantic edges determined by context dependencies and retrieved from the semantic graph, representing the structural connection method between concept nodes; and relative position numbers, which are the sequence numbers assigned to each semantic graph unit according to the order of the sentences in the document, used to record the document-level order of each semantic node and semantic edge in the semantic subgraph.

[0067] In one embodiment, step S50, namely comparing the structural similarity between the semantic subgraph of the document to be deduplicated and multiple semantic subgraphs of the comparison documents to obtain a structural similarity score, includes: S501: Based on the semantic subgraph of the document to be checked for plagiarism and the semantic subgraphs of multiple comparison documents, construct a graph structure representation for structural similarity analysis.

[0068] Specifically, a unified graph structure encoding operation is performed on the semantic subgraph of the document to be checked and the semantic subgraphs of multiple comparison documents, converting each semantic subgraph into a standard graph structure representation for structural similarity analysis. Graph structure encoding refers to organizing the structural elements and attribute information contained in the semantic subgraph into a graph-theoretic data structure to support structural analysis and computation. During the graph structure encoding process, the set of concept nodes and the set of semantic edges in each semantic subgraph are first extracted. The set of concept nodes represents the core semantic entities constituting the semantic subgraph, and the set of semantic edges represents the semantic relationships between concept nodes. Based on the order of appearance of concept nodes in the original sentence and the connection relationships in the semantic subgraph, two core graph representation matrices are constructed: an adjacency matrix and a node attribute matrix. The adjacency matrix is ​​a two-dimensional matrix structure used to express the connection relationships between concept nodes, where each row and column corresponds to a concept node. When there is a semantic edge connection between two concept nodes, the edge weight is recorded at the corresponding matrix position; if there is no connection, it is recorded as zero. A node attribute matrix is ​​a matrix structure used to express the semantic attribute features of each concept node. Each row corresponds to a concept node, and each column corresponds to a semantic feature dimension, including semantic classification labels, domain attribute identifiers, lexical embedding vectors, or terminology-level labels. When constructing the adjacency matrix, the connection type, path length, and semantic relationship label of each semantic edge are encoded as additional information in the edge weight and placed in the corresponding positions of the adjacency matrix. The connection type refers to the category of semantic relationship represented by the semantic edge, such as synonymy, causality, or hierarchical relationship, used to distinguish different semantic functions. The path length refers to the number of edges contained in the semantic path constituting the semantic connection, used to reflect the semantic propagation distance between two concept nodes. The semantic relationship label is the text or encoded mark used to identify the specific relationship definition of the semantic edge. Through the joint expression of the adjacency matrix and the node attribute matrix, a complete graph structure representation integrating structural and semantic attribute information is formed.

[0069] S502: Based on the graph structure representation, calculate the node similarity and semantic edge connection similarity between the semantic subgraph of the document to be checked and each semantic subgraph of the comparison document.

[0070] Specifically, the semantic subgraph of the document to be checked for plagiarism and the semantic subgraphs of multiple comparison documents are first obtained, along with the sets of concept nodes and semantic edges. The set of concept nodes refers to the collection of all concept nodes in the semantic subgraph. Each concept node is encapsulated by a term, containing information such as term name, semantic category, domain attribute, and unique term code. The set of semantic edges refers to the collection of all semantic edges in the semantic subgraph. Each semantic edge connects a source concept node and a target concept node, and includes the semantic relationship type, connection direction, contextual condition identifier, and unique relationship number. After obtaining these sets, a unified graph structure encoding operation is performed. Graph structure encoding involves converting the node and edge information in the semantic subgraph into a matrix form to support subsequent similarity calculations. Specifically, this includes constructing an adjacency matrix: The adjacency matrix is ​​a two-dimensional matrix generated based on the set of concept nodes, according to the order of node appearance in the sentence and the connection relationships in the set of semantic edges. In this matrix, each row and column corresponds to a concept node. If a semantic edge exists between two concept nodes, the edge weight is filled into the corresponding matrix cell; otherwise, zero is filled in. The adjacency matrix is ​​used to represent the connection structure between concept nodes. Constructing the node attribute matrix: The node attribute matrix is ​​a two-dimensional matrix based on the semantic attribute features of concept nodes, where each row corresponds to a concept node and each column corresponds to a semantic attribute. Semantic attribute features include part-of-speech tagging, semantic categories, domain term labels, and contextual semantic vectors. The node attribute matrix describes the semantic information of concept nodes. Generating edge weight information: Edge weight information refers to the specific numerical values ​​filled into the adjacency matrix. These values ​​are determined by the connection type, path length, and semantic relation label of the semantic edge. The connection type refers to the semantic relation category expressed by the semantic edge, such as "superordinate-inferior," "causal," or "belonging." The path length refers to the number of dependency steps between the source and target concept nodes. The semantic relation label refers to the relation identifier defined by the semantic edge in the semantic rule base information. By mapping the connection type, path length, and semantic relation label to numerical values ​​and then weighting and combining them, the final edge weights can be obtained. After the above steps are completed, the semantic subgraph of the document to be checked is transformed into a semantic graph structure representation containing an adjacency matrix and a node attribute matrix. Similarly, the semantic subgraphs of multiple comparison documents are also transformed into their corresponding semantic graph structure representations. The semantic graph structure representation refers to the graphical encoding result composed of an adjacency matrix, a node attribute matrix, and edge weight information, integrating structural and semantic attribute information to support subsequent structural similarity analysis.

[0071] S503: The structural similarity score is obtained by weighting the node similarity and the semantic edge connection similarity.

[0072] Specifically, firstly, node similarity and semantic edge connection similarity are calculated for the semantic subgraph of the document to be checked and for each semantic subgraph of the comparison document. Node similarity refers to the degree of similarity between concept nodes in the semantic subgraph of the document to be checked and concept nodes in the semantic subgraph of the comparison document at the semantic attribute level. The calculation of node similarity is based on the semantic embedding vector representation of concept nodes, using the cosine similarity function or the Jaccard coefficient function. The cosine similarity function measures the consistency of semantic direction by calculating the cosine value of the angle between two vectors, while the Jaccard coefficient function measures the degree of semantic coverage by calculating the ratio of the intersection to the union of two sets of concept attributes. After obtaining the pairwise similarity between all concept nodes, a node similarity matrix is ​​constructed. The node similarity matrix is ​​a two-dimensional matrix filled with the semantic similarity values ​​between corresponding nodes, using the concept nodes in the semantic subgraphs of the document to be checked and the semantic subgraphs of the comparison document as row and column indices, respectively. By performing the maximum matching algorithm or the average matching strategy on the node similarity matrix, the overall similarity score at the node level is obtained, which serves as the summary value of node similarity. Then, the semantic edge connection similarity is calculated. Semantic edge connection similarity refers to the degree of matching between the semantic edges in the semantic subgraph of the document to be checked and the semantic edges in the semantic subgraph of the comparison document in terms of structural connection features. Semantic edge connection features include three dimensions: connection direction, semantic relationship type, and path length. Connection direction refers to the directional nature of the semantic edge between the source concept node and the target concept node, such as from concept A to concept B. Semantic relationship type refers to the semantic category represented by the semantic edge, such as "causal relationship," "hyper-hyper-parallel relationship," or "parallel relationship." Path length refers to the dependency connection distance between two concept nodes, usually expressed as the number of edges. For each pair of matching semantic edges, a comparison is performed item by item according to the above three feature dimensions, a structural similarity score is calculated, and the semantic edge connection similarity value is obtained by summing them. After obtaining the node similarity and semantic edge connection similarity, they are combined and calculated according to the set weighting parameters to obtain the structural similarity score.

[0073] In one embodiment, in step S503, the node similarity and semantic edge connection similarity are weighted and combined to obtain a structural similarity score, including: S5031: The structural similarity score is obtained using the following formula: , of which S struct This refers to the structural similarity score, S node This refers to node similarity, S edge α refers to the similarity of semantic edge connection methods, β refers to the similarity weight of node similarity, α+β=1, and ε refers to the perturbation response coefficient, which ranges from 0 to 1.

[0074] Specifically, the node similarity (Snode) and semantic edge similarity (Sedge) between the semantic subgraph of the document to be checked and each semantic subgraph of the comparison document are obtained. Node similarity is calculated using the cosine similarity or Jaccard coefficient of semantic embedding vectors by comparing the semantic labels, categories, and contextual consistency of each concept node in the two semantic subgraphs. Semantic edge similarity is calculated using a structural pattern comparison function based on factors such as the connection direction, dependency path structure, and semantic role consistency of the edges between concept nodes. Then, a node similarity weight α and a semantic edge similarity weight β are set, with their sum being 1, to balance the contribution of the two types of similarity to structural similarity. A perturbation response coefficient ε is then selected, which is set based on the statistical results of the fluctuation of structural similarity between the previously constructed perturbed semantic subgraph and the original semantic subgraph. Finally, the results are calculated according to the formula... Perform weighted combination operations and output structural similarity scores for subsequent plagiarism detection.

[0075] In one embodiment, in step S60, based on the node order in the semantic subgraph of the document to be deduplicated, a perturbation semantic subgraph is constructed, and the structural similarity between the perturbation semantic subgraph and multiple comparative document semantic subgraphs is re-compared to calculate the perturbation score fluctuation range, including: S601: Based on node order, construct a perturbation semantic subgraph by performing node order swapping operations within a limited range.

[0076] Specifically, the process first obtains the original order information of all concept nodes in the semantic subgraph of the document to be checked. This original order information refers to the arrangement order of each concept node recorded according to the order in which sentences appear in the document during the construction of the semantic subgraph, representing the linear arrangement structure of the concept nodes. Following preset perturbation operation rules, a fixed node order exchange window length is set. Based on this window length, adjacent concept nodes are selected sequentially in the concept node sequence, and node order exchange operations are performed while maintaining semantic edges, generating a perturbed version of the semantic subgraph. The perturbation operation rules refer to a set of strategies used to restrict the structural perturbation method. This set of strategies includes three main parameters: the node order exchange window length (the number of concept nodes participating in the order exchange in each round of perturbation), the exchange step size (the movement step of the sliding window in the node sequence), and whether repeated node exchange operations are allowed (whether the same node pair is exchanged multiple times in the same semantic subgraph). Typically, setting the node order exchange window length to 2 allows for exchange operations only between adjacent concept nodes, introducing sufficient structural perturbation changes without disrupting semantic edge connections, while controlling the perturbation magnitude. After each node order swap operation, a new semantic subgraph is generated. This new subgraph is identical to the original semantic subgraph of the document to be checked for plagiarism in terms of both the set of concept nodes and the set of semantic edges, differing only in the order of the concept nodes. This perturbation process is repeated, continuously sliding the window and swapping the order of nodes at different positions, gradually constructing a perturbed semantic subgraph set containing multiple perturbation versions. The perturbed semantic subgraph set is a dataset composed of multiple semantic subgraphs with identical structures but different node orders, used to simulate the stability of the semantic structure under conditions of permutation changes.

[0077] S602: Compare the structural similarity of the perturbation semantic subgraph with the semantic subgraphs of multiple comparison documents to obtain multiple perturbation structural similarity scores.

[0078] Specifically, for the set of perturbed semantic subgraphs constructed by sequential perturbation of the semantic subgraph of the document to be deduplicated, each perturbed semantic subgraph is selected sequentially, and structural feature comparison operations are performed between the semantic subgraphs and multiple comparison document semantic subgraphs. In each comparison, based on the matching between the perturbed semantic subgraph and the target comparison document semantic subgraph in terms of node composition, connection patterns, and semantic path structure, the following three types of structural indicators are extracted: concept node coverage rate, which represents the coverage ratio of matched concept nodes in the perturbed semantic subgraph in the target comparison document semantic subgraph; semantic connectivity, which measures the complete preservation of effective connection paths existing in the perturbed semantic subgraph in the comparison document semantic subgraph; and semantic path direction consistency, which evaluates the structural consistency between the two semantic subgraphs in terms of connection order and direction. After normalizing the above three structural indicators, a perturbed structure evaluation function is constructed. The perturbed structure evaluation function is defined as follows: ,in, This represents the perturbation structure similarity score between the i-th perturbation semantic subgraph and the j-th comparison document semantic subgraph. This represents the concept node coverage metric. Indicates semantic connectivity metrics. The semantic path direction consistency index is represented by ρ1, ρ2, and ρ3, which are weighting coefficients in the perturbation structure score calculation, satisfying ρ1 + ρ2 + ρ3 = 1. By calculating the perturbation structure similarity score for each perturbation semantic subgraph and each comparison document semantic subgraph, a set of perturbation structure score sequences grouped by document is obtained. This perturbation structure score sequence serves as the basis for subsequent perturbation stability analysis and fluctuation range calculation. S603: Calculate the range of similarity scores for multiple perturbation structures to obtain the fluctuation range of perturbation scores.

[0079] Specifically, for each semantic subgraph of the comparison document, the perturbation structure similarity scores of all perturbation semantic subgraphs corresponding to that subgraph are collected, forming a perturbation structure similarity score sequence. The perturbation structure similarity score sequence refers to the set of multiple perturbation structure similarity scores obtained by comparing the same semantic subgraph of the comparison document with multiple perturbation semantic subgraphs, arranged sequentially according to the perturbation version. After obtaining the perturbation structure similarity score sequence, a range calculation operation is performed. The range calculation operation involves extracting the maximum and minimum perturbation structure similarity scores from the perturbation structure similarity score sequence and calculating the numerical difference between them. The range calculation formula is as follows: ,in: This indicates the range of fluctuation in the disturbance score. This represents the perturbation structure similarity score corresponding to the 1st to nth perturbation semantic subgraph in the perturbation structure similarity score sequence.

[0080] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0081] In one embodiment, a document plagiarism detection device is provided, which corresponds one-to-one with a document plagiarism detection method described in the above embodiments. For example... Figure 2 As shown, this document plagiarism detection device includes a text information acquisition module, a syntactic dependency analysis module, a semantic graph construction module, a semantic subgraph generation module, a structural similarity calculation module, a perturbation fluctuation assessment module, and a similar document screening and difference output module. Detailed descriptions of each functional module are as follows: The text information acquisition module is used to acquire the text information of the document to be checked for plagiarism and the text information of multiple comparison documents. The syntactic dependency analysis module is used to identify keywords and contextual dependency relationships in each sentence based on the syntactic dependency structure of the document to be checked and multiple comparison documents. The semantic graph construction module is used to construct semantic graphs. The semantic subgraph generation module maps keywords and contextual dependencies to corresponding concept nodes in the semantic graph, constructing a semantic subgraph of the document to be checked and semantic subgraphs of multiple comparison documents. The structural similarity calculation module compares the structural similarity between the semantic subgraph of the document to be checked and the semantic subgraphs of multiple comparison documents to obtain a structural similarity score. The perturbation fluctuation assessment module is used to construct a perturbation semantic subgraph based on the node order in the semantic subgraph of the document to be deduplicated, re-compare the structural similarity between the perturbation semantic subgraph and the semantic subgraphs of multiple comparison documents, and calculate the perturbation score fluctuation range. The similar document filtering and difference output module is used to filter comparison documents whose structural similarity scores are higher than a preset threshold and whose perturbation score fluctuation range is lower than the perturbation tolerance threshold, and output the plagiarism difference information.

[0082] For specific limitations regarding a document plagiarism detection device, please refer to the limitations of a document plagiarism detection method described above, which will not be repeated here. Each module in the aforementioned document plagiarism detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0083] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain the text information of the document to be checked for plagiarism and the text information of multiple comparison documents. Based on the syntactic dependency structure of the document to be checked and the text information of multiple comparison documents, the keywords and contextual dependency relationships of each sentence are identified. Constructing a semantic graph, Keywords and contextual dependencies are mapped to corresponding concept nodes in the semantic graph, constructing a semantic subgraph of the document to be checked and multiple semantic subgraphs of the comparison documents. Structural similarity scores are obtained by comparing the structural similarity between the semantic subgraph of the document to be checked and the semantic subgraphs of multiple comparison documents. Based on the node order in the semantic subgraph of the document to be deduplicated, a perturbed semantic subgraph is constructed. The structural similarity between the perturbed semantic subgraph and the semantic subgraphs of multiple comparison documents is then re-compared, and the fluctuation range of the perturbation score is calculated. The system filters out comparison documents whose structural similarity scores are higher than a preset threshold and whose perturbation scores fluctuate within a range lower than the perturbation tolerance threshold, and outputs information on the differences in plagiarism detection.

[0084] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain the text information of the document to be checked for plagiarism and the text information of multiple comparison documents. Based on the syntactic dependency structure of the document to be checked and the text information of multiple comparison documents, the keywords and contextual dependency relationships of each sentence are identified. Constructing a semantic graph, Keywords and contextual dependencies are mapped to corresponding concept nodes in the semantic graph, constructing a semantic subgraph of the document to be checked and multiple semantic subgraphs of the comparison documents. Structural similarity scores are obtained by comparing the structural similarity between the semantic subgraph of the document to be checked and the semantic subgraphs of multiple comparison documents. Based on the node order in the semantic subgraph of the document to be deduplicated, a perturbed semantic subgraph is constructed. The structural similarity between the perturbed semantic subgraph and the semantic subgraphs of multiple comparison documents is then re-compared, and the fluctuation range of the perturbation score is calculated. The system filters out comparison documents whose structural similarity scores are higher than a preset threshold and whose perturbation scores fluctuate within a range lower than the perturbation tolerance threshold, and outputs information on the differences in plagiarism detection.

[0085] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A document plagiarism detection method, characterized in that, The document deduplication method includes: Obtain the text information of the document to be checked for plagiarism and the text information of multiple comparison documents; Based on the syntactic dependency structure of the text information of the document to be checked and the text information of multiple comparison documents, identify the keywords and contextual dependency relationships of each sentence; Constructing a semantic graph; Keywords and contextual dependencies are mapped to corresponding concept nodes in the semantic graph to construct a semantic subgraph of the document to be checked and multiple semantic subgraphs of the comparison documents; The structural similarity score is obtained by comparing the structural similarity between the semantic subgraph of the document to be checked and the semantic subgraphs of multiple comparison documents. Based on the node order in the semantic subgraph of the document to be checked for plagiarism, a perturbation semantic subgraph is constructed, and the structural similarity between the perturbation semantic subgraph and the semantic subgraphs of multiple comparison documents is re-compared to calculate the fluctuation range of the perturbation score. The system filters out comparison documents whose structural similarity scores are higher than a preset threshold and whose perturbation scores fluctuate within a range lower than the perturbation tolerance threshold, and outputs information on the differences in plagiarism detection.

2. The document plagiarism detection method according to claim 1, characterized in that, Based on the syntactic dependency structure of the text information of the document to be checked and the text information of multiple comparison documents, the keywords and contextual dependency relationships of each sentence are identified, including: Perform bidirectional path traversal on the syntactic dependency relation structure and calculate the semantic connection strength of each candidate keyword on the forward and reverse paths respectively; Based on semantic connection strength, words whose semantic connection stability is higher than the preset keyword recognition threshold are selected as keywords; Determine the corresponding contextual dependency relationship based on the connection pattern between the keyword and other words in the dependency path.

3. The document plagiarism detection method according to claim 1, characterized in that, The steps to construct a semantic graph include: Acquire terminology, semantic relationship samples between terms, and semantic rule base information corresponding to the target application domain; Constructing a set of concept nodes in a semantic graph based on terminology; Extract the term pairs that constitute the semantic relations from the semantic relation samples, and map the two terms in the term pairs to the corresponding source concept node and target concept node in the concept node set, respectively. Based on information from the semantic rule base, verify whether the semantic relationships represented by terms and words satisfy the preset semantic rule constraints. When the semantic relationship satisfies the preset semantic rule constraints, construct a semantic edge connecting the source concept node and the target concept node, and form a semantic edge set by multiple semantic edges that satisfy the constraints. A semantic graph that satisfies semantic constraints is constructed by combining a set of concept nodes and a set of semantic edges.

4. The document plagiarism detection method according to claim 1, characterized in that, Keywords and contextual dependencies are mapped to corresponding concept nodes in the semantic graph, constructing a semantic subgraph of the document to be checked and multiple semantic subgraphs of the comparison documents, including: Based on the semantic information contained in the keywords, target nodes that match the descriptive information of concept nodes in the semantic graph are identified, and a mapping relationship is established between the keywords and the matched concept nodes to obtain a node mapping set; Based on the grammatical direction and dependency type indicated in the context dependency relation, determine the structural connection mode between any pair of concept nodes from the node mapping set, and retrieve the graph connection edges that conform to the structural connection mode in the semantic graph to construct the node connection relation set; Based on the set of node mappings and the set of node connection relationships, a graphical unit representing the semantic structure of a single sentence is generated. The graphical units are divided into sentences and assigned to the corresponding sentences of the document to be checked for plagiarism and multiple comparison documents, thus constructing a semantic subgraph of the document to be checked for plagiarism and a semantic subgraph of multiple comparison documents.

5. The document plagiarism detection method according to claim 1, characterized in that, The structural similarity scores are obtained by comparing the semantic subgraphs of the document to be deduplicated with the semantic subgraphs of multiple comparison documents, including: Based on the semantic subgraph of the document to be checked for plagiarism and the semantic subgraphs of multiple comparison documents, a graph structure representation for structural similarity analysis is constructed. Based on the graph structure representation, calculate the node similarity and semantic edge connection similarity between the semantic subgraph of the document to be checked and each semantic subgraph of the comparison document; The structural similarity score is obtained by weighting the node similarity and the semantic edge connection similarity.

6. The document plagiarism detection method according to claim 5, characterized in that, The structural similarity score is obtained by weighting node similarity and semantic edge connection similarity, including: The structural similarity score is obtained using the following formula: , of which S struct This refers to the structural similarity score, S node This refers to node similarity, S edge α refers to the similarity of semantic edge connection methods, β refers to the similarity weight of node similarity, α+β=1, and ε refers to the perturbation response coefficient, which ranges from 0 to 1.

7. The document plagiarism detection method according to claim 1, characterized in that, Based on the node order in the semantic subgraph of the document to be deduplicated, a perturbed semantic subgraph is constructed. The structural similarity between the perturbed semantic subgraph and the semantic subgraphs of multiple comparison documents is re-compared, and the fluctuation range of the perturbation score is calculated, including: Based on node order, a perturbation semantic subgraph is constructed by exchanging node order within a limited range. The perturbation semantic subgraph is compared with the semantic subgraphs of multiple comparison documents to obtain multiple perturbation structural similarity scores; The range of perturbation score fluctuation is obtained by calculating the range of similarity scores of multiple perturbation structures.

8. A document plagiarism detection device, characterized in that, The document plagiarism detection device includes: The text information acquisition module is used to acquire the text information of the document to be checked for plagiarism and the text information of multiple comparison documents; The syntactic dependency analysis module is used to identify the keywords and contextual dependency relationships of each sentence based on the syntactic dependency structure in the text information of the document to be checked and the text information of multiple comparison documents. The semantic graph construction module is used to construct semantic graphs; The semantic subgraph generation module is used to map keywords and contextual dependencies to corresponding concept nodes in the semantic graph, and to construct the semantic subgraph of the document to be checked and the semantic subgraph of multiple comparison documents; The structural similarity calculation module is used to compare the structural similarity between the semantic subgraph of the document to be checked and the semantic subgraphs of multiple comparison documents to obtain a structural similarity score; The perturbation fluctuation assessment module is used to construct a perturbation semantic subgraph based on the node order in the semantic subgraph of the document to be deduplicated, re-compare the structural similarity between the perturbation semantic subgraph and the semantic subgraphs of multiple comparison documents, and calculate the perturbation score fluctuation range. The similar document filtering and difference output module is used to filter comparison documents whose structural similarity scores are higher than a preset threshold and whose perturbation score fluctuation range is lower than the perturbation tolerance threshold, and output the plagiarism difference information.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the document deduplication method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the document deduplication method as described in any one of claims 1 to 7.

Images