Creference intention classification method based on dynamic heterogeneous graph fusion network

Abstract

The invention discloses a reference intention classification method based on a dynamic heterogeneous graph fusion network. The method comprises the following steps: dynamically extracting reference sentences and related contexts thereof, and combining chapter information to form a text T; utilizing a SciBERT pre-training model to extract a feature vector of the text T; constructing a heterogeneous graph to express a complex relationship among nodes such as papers, chapters and authors; node features are extracted through a dynamic heterogeneous graph network, and node feature vectors of a heterogeneous graph are generated; fusing the node feature vector with a text feature vector of the text T to form a comprehensive feature; and finally, predicting the reference intention by using a classifier, and optimizing the model through a loss function. According to the method, the semantic features with distinction are extracted from the internal structure of the paper and the cross-literature relation, high-precision citation intention recognition is achieved in a complex citation network, and the method is particularly suitable for citation intention classification tasks of academic literatures.

Description

Technical Field

[0001] The present invention relates to the technical field of scientometrics and academic literature analysis, and in particular to a citation intention classification method based on a dynamic heterogeneous graph fusion network. Background Art

[0002] Understanding citation intent has important practical significance for academic literature analysis. First, it helps in literature evaluation in scientometrics, accurately identifying the citation type of a paper, and more objectively measuring the influence of the paper in the academic community. Second, citation intent classification also has extensive value in applications such as academic recommendation systems, knowledge graph construction, and academic search engine optimization.

[0003] With the rapid development of machine learning and natural language processing (NLP), the number of automated methods for citation intent classification has gradually increased. Among them, heterogeneous graph classification methods have attracted widespread attention due to their advantages in integrating multimodal information and capturing complex relationships. This type of method improves text classification by constructing heterogeneous graphs between different types of nodes (such as papers, authors, citations, etc.). However, current heterogeneous graph text classification methods are mostly applied to general document classification or content classification tasks, lacking a dynamic mechanism specifically designed for citation intent classification, and are difficult to adapt to the ever-changing context and semantic requirements in citation relationships. These methods have poor generalization capabilities in specific citation intent recognition tasks and are difficult to accurately characterize the true intentions of academic authors.

[0004] For example, the patent document with publication number CN114564573A discloses a method for predicting academic cooperation relationships based on heterogeneous graph neural networks. Its heterogeneous graph contains author nodes, work nodes, belonging nodes, and content nodes, which are mainly used to express the topological structure of scholars' cooperation relationships and interest nodes. It uses DeepWalk and Text-CNN to pre-train node features, and aggregates nodes through a multi-head attention mechanism combined with meta-paths. This solution is very basic and cannot capture the deep semantic information of heterogeneous graph structures. The patent document with publication number CN114722192A discloses a multi-label classification method for disciplinary inspection clues based on heterogeneous graph neural networks. By constructing a heterogeneous graph structure of sentence and word nodes, the entire text is decomposed into the connection relationship between words and sentences, and the syntactic structure information is aggregated using GNN (graph neural network). This structure is suitable for general multi-label text classification tasks and classification is achieved based on the grammatical association between word and sentence nodes. The model uses Bi-LSTM (bidirectional long short-term memory network) and attention mechanism to aggregate the context of sentence nodes, but is limited to pre-defined syntactic contexts and does not dynamically adapt to the specific semantics of quoted sentences.

[0005] It can be seen that although previous heterogeneous graph classification methods can use the heterogeneity of nodes and edges to capture surface relationships, it is difficult to effectively obtain the deep semantic information in the citation intent. The citation intent classification task usually requires identifying the potential intention of academic authors when citing a certain document. This process involves a deep understanding of the citation sentence, context, and multi-layer node relationships. However, these heterogeneous graph methods are often limited to the representation of shallow connections and simple relationships, and it is difficult to fully capture the semantic depth and delicacy required for citation intent. Therefore, when faced with complex contextual semantics in citation intent, the recognition effect of such methods is often unsatisfactory, which limits the performance in practical applications. Obviously, existing heterogeneous graph text classification methods still face significant challenges in the citation intent classification task. To this end, it is necessary to design a new solution to better adapt to the dynamics and complexity of citation intent classification, thereby improving the accuracy and applicability of citation intent classification. Summary of the invention

[0006] In order to address the shortcomings of the prior art, the present invention provides a citation intent classification method based on a dynamic heterogeneous graph fusion network. The method takes the citation sentence as the center, selects the most relevant context sentences from the paper text, introduces chapter information as an auxiliary feature for citation intent identification, and constructs a heterogeneous graph to express the complex relationship between nodes such as papers, chapters and authors. The feature representation of citation intent classification is enriched, and the characteristics of heterogeneous graphs that dynamically adjust the aggregation of adjacent nodes and assign adaptive weights to different node relationships are utilized. This enables the model to extract more discriminative semantic features from the internal structure of the paper and the relationship across documents, and can more accurately identify citation intent in a complex citation network. The method is particularly suitable for citation intent classification tasks in academic documents.

[0007] The embodiment of the present invention provides the following solution:

[0008] An embodiment of the present invention provides a method for classifying reference intent based on a dynamic heterogeneous graph fusion network, the method comprising:

[0009] Step 1: For each quoted sentence in the dataset, dynamically extract its context sentences, select the most relevant context sentence set based on the semantic similarity between the quoted sentence and the context sentences, and combine the chapter information to form the chapter-enhanced context text T;

[0010] Step 2: Use the SciBERT pre-trained model to encode the chapter-enhanced context text T into a text feature vector;

[0011] Step 3: Construct a heterogeneous graph. The nodes in the heterogeneous graph include paper nodes, chapter nodes, and author nodes. The node features are represented by text features or entity information. The edge relationships are used to express the reference relationships, chapter association relationships, or author relationships between nodes.

[0012] Step 4: Extract representations of the heterogeneous graph, iteratively update node features using a dynamic heterogeneous graph network, adaptively adjust the weights of different node relationships, and dynamically aggregate the features of neighboring nodes to generate a node feature vector of the heterogeneous graph;

[0013] Step 5: Fuse the node feature vector of the heterogeneous graph with the feature vector of the chapter-enhanced context text T to form the final classification feature, which is input into the classifier for citation intent classification prediction;

[0014] Step 6: Use the loss function to measure the difference between the classification prediction results and the true labels and optimize the model parameters.

[0015] In an optional embodiment, step one includes the following process:

[0016] S1.1. Get the reference sentence C from the dataset cite ;

[0017] S1.2, extract the three sentences before and after the quoted sentence, and form a candidate sentence set C = {C i};

[0018] S1.3. For each candidate sentence C i , calculate its difference with the reference sentence C cite Similarity(C cite ,C i ):

[0019]

[0020] Among them, v cite and v i They are the vector representations of the quoted sentence and the candidate sentence respectively;

[0021] S1.4. The quoted sentence and the most relevant context sentence are merged to form text C containing quoted context information:

[0022] Select and quote sentence C from the candidate sentence set cite Most relevant sentence C rel :

[0023]

[0024] Use the concatenation operation to quote sentence C cite and the most relevant sentence C rel Combine and concatenate to form text C containing the reference context information;

[0025] S1.5. The chapter name is merged with the text C containing the reference context information to form a chapter-enhanced context text T.

[0026] In an optional embodiment, step 2 includes the following process:

[0027] S2.1. For the enhanced context text of the chapter T = {w1, w2, ..., w n}, and use the SciBERT model to generate the corresponding word-level feature representation {h1,h2,...,h n}:

[0028] {h1,h2,…,h n}=SciBERT({w1,w2,…,w n})

[0029] Among them, w i represents the i-th word in the text, h i For word w i Embedding vector representation in the model;

[0030] S2.2. Take the hidden state h corresponding to the [CLS] tag [CLS] The feature vector representation of the context text T as chapter enhancement:

[0031] T=h [CLS] ∈R d

[0032] Where R d is the hidden layer dimension of the SciBERT model;

[0033] S2.3. Normalize the text feature vector to obtain the text feature representation h T .

[0034] In an optional embodiment, the heterogeneous graph in step 3 includes nodes and edges, wherein:

[0035] Node: the i-th paper P in the dataset i Use Node It means that its eigenvector is Chapter s u s Use Node It means that the eigenvector is The kth author A k Use Node It means that the eigenvector is

[0036] Edge: For the i-th paper P in the dataset i and cited papers, the jth paper P j The reference relationship between For the i-th paper P i and Chapter s sThe reference relationship between For the kth author A k and the i-th paper P i The reference relationship between express;

[0037] Then the heterogeneous graph is G = (V, E), where the node set Edge Set

[0038] In an optional embodiment, if there is no author information in the data set, the author node set is an empty set, so the constructed heterogeneous graph is G'=(V',E'), where

[0039] In an optional embodiment, step 4 includes the following process:

[0040] S4.1. Input the heterogeneous graph G into the HGT model. Any type of node v∈V has a corresponding feature vector X v Represents; suppose there are N nodes and the feature dimension is d, then:

[0041]

[0042] S4.2. The features of each node are propagated through its neighboring nodes. For each node v, the feature information of its neighboring nodes N(v) is updated by the following formula:

[0043]

[0044] in, is the feature vector of the neighbor node v in the lth layer, W (l) is the weight matrix of the lth layer, C vu is the normalization factor;

[0045] S4.3. Calculate the attention coefficient e of node v to its neighbor node u in each layer by the following formula vu :

[0046]

[0047] Among them, ReLU() is the activation function, a is the learned attention vector, || represents the vector concatenation processing, is the feature vector of node v in the lth layer; is the feature vector of neighbor node y in the lth layer; W (l) is the weight matrix of the lth layer, used to transform node features; e vu is the attention coefficient of node v to its neighbor node u, indicating the influence of the neighbor node;

[0048] S4.4. Normalize the attention coefficient through the Softmax function to obtain the influence weight α of node u on node v vu :

[0049]

[0050] By weighted aggregation of neighboring node features, the feature vector representation of node v is updated:

[0051]

[0052] S4.5. For node v, calculate the semantic attention weights under different meta-paths

[0053]

[0054] Among them, a m is the weight vector of semantic attention under the mth path, M is the total number of meta-paths, and m represents the path ordinal number;

[0055] S4.6. Update the node feature vector representation of node v:

[0056]

[0057] In an optional embodiment, step five includes the following process:

[0058] S5.1. Use element-level multiplication operation to fuse the node feature vector of the heterogeneous graph with the text feature vector of the text T to obtain the fused feature F:

[0059] F=H v ⊙h T

[0060] S5.2, classification through fully connected layer and Softmax activation function:

[0061] P(y|F)=Softmax(W f F+b f )

[0062] Among them, W f F is the classifier weight matrix, b f is the bias setting, and P(y|F) is the predicted category probability distribution.

[0063] In an optional embodiment, the loss function described in step six adopts a cross entropy loss function.

[0064] The beneficial effects of the present invention based on its technical solution are:

[0065] (1) The citation intention classification method based on a dynamic heterogeneous graph fusion network provided by the present invention introduces a multi-level node structure of "paper-chapter-author". The chapter information provides a higher level of contextual information for the citation intention, which enables the model to have an in-depth understanding of the background of the citation intention. By incorporating specific structures in academic documents (such as chapters and authors) into the heterogeneous graph, the unique contextual patterns of academic documents can be captured, further improving the domain adaptability of the model, especially capturing the citation intention in academic documents.

[0066] (2) The citation intent classification method based on a dynamic heterogeneous graph fusion network provided by the present invention designs a dynamic context extraction strategy at the text level, extracts the most relevant context sentences centered on the citation sentence, and embeds chapter information as an auxiliary feature to enhance the recognition of citation intent. The dynamic context can flexibly adapt to different citation scenarios in practical applications, while the chapter information can provide clear contextual guidance for citation intent. This strategy combining dynamics with high-level context has not been implemented in general multi-label text classification, and is particularly suitable for the requirements for context understanding in citation intent classification.

[0067] (3) The citation intent classification method based on a dynamic heterogeneous graph fusion network provided by the present invention not only aggregates multi-node information in a heterogeneous graph, but also uses HGT (heterogeneous graph transformation) to flexibly aggregate heterogeneous node information and dynamically adjust weights during the node aggregation process. By considering the heterogeneity between different nodes, HGT can capture the complex semantic relationship between nodes at a deeper level, providing more detailed feature expression and higher semantic capture accuracy for citation intent classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0069] Figure 1 It is a schematic diagram of the process of the present invention.

[0070] Figure 2 The figure is a schematic diagram of the processing process of an embodiment of the present invention. DETAILED DESCRIPTION

[0071] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention belong to the scope of protection of the embodiments of the present invention.

[0072] Reference Figure 1 and Figure 2 , an embodiment of the present invention provides a reference intention classification method based on a dynamic heterogeneous graph fusion network, the method comprising:

[0073] Step 1: For each quoted sentence in the data set, dynamically extract its context sentences, select the most relevant context sentence set based on the semantic similarity between the quoted sentence and the context sentences, and combine the chapter information to form the chapter-enhanced context text T. It includes the following processes:

[0074] S1.1. Get the reference sentence C from the dataset cite ;

[0075] S1.2, extract the three sentences before and after the quoted sentence, and form a candidate sentence set C = {C i};

[0076] S1.3. For each candidate sentence C i , calculate its difference with the reference sentence C cite Similarity(C cite ,C i ):

[0077]

[0078] Among them, v cite and v i They are the vector representations of the quoted sentence and the candidate sentence respectively;

[0079] S1.4. The quoted sentence and the most relevant context sentence are merged to form text C containing quoted context information:

[0080] Select and quote sentence C from the candidate sentence set cite Most relevant sentence C rel :

[0081]

[0082] Once the most relevant sentence C is determined rel , you can use the concatenation operation to reference sentence C cite and the most relevant sentence C relCombine. Concatenation can be achieved by directly concatenating the vector representations of the two sentences into one long vector:

[0083] C=[C cite ; C rel ]

[0084] Among them, the symbol [C cite ; C rel ] indicates that the sentence will be quoted C cite and the representation of the most relevant sentence C rel The concatenated text C containing the reference context information is used as the input for subsequent tasks.

[0085] S1.5. Merge the chapter name with the text C containing the reference context information. This concatenation operation can be achieved by directly concatenating the representation S of the chapter name and the representation C containing the reference context information into a long vector. The specific formula is as follows:

[0086] T = [S; C]

[0087] Among them, S represents the chapter name, C represents the text C containing the reference context information, and the concatenated chapter enhanced context text T is the final input.

[0088] Through concatenation, the chapter name S and the text C containing the citation context information are merged into a new text, namely the chapter enhanced context text T, which contains the chapter topic information and the enhanced context of the citation sentence. The concatenated text can provide richer semantic information for subsequent tasks and help the model better understand the intention of literature citation.

[0089] Step 2: Use the SciBERT pre-trained model to encode the enhanced context text T into a text feature vector as the initial feature representation of the corresponding node in the subsequent heterogeneous graph construction. This includes the following process:

[0090] S2.1. For the chapter enhanced context text T = {w1, w2, ..., w n}, and use the SciBERT model to generate the corresponding word-level feature representation {h1,h2,...,h n}:

[0091] {h1,h2,…,h n}=SciBERT({w1,w2,…,wn})

[0092] Among them, w i represents the i-th word in the text, h i For word w i Embedding vector representation in the model;

[0093] S2.2. Take the hidden state h corresponding to the [CLS] tag [CLS] The feature vector representation of the context text T as chapter enhancement:

[0094] T=h [CLS] ∈R d

[0095] Where R d is the hidden layer dimension of the SciBERT model;

[0096] S2.3. Normalize the text feature vector to obtain the text feature representation h T .

[0097] Step 3: Construct a heterogeneous graph. The nodes in the heterogeneous graph include paper nodes, chapter nodes, and author nodes. Node features are represented by text features or entity information, and edge relationships are used to express reference relationships, chapter association relationships, or author relationships between nodes.

[0098] Node: the i-th paper P in the dataset i Use Node It means that its eigenvector is Chapter s u s Use Node It means that the eigenvector is The kth author A k Use Node It means that the eigenvector is

[0099] Edge: For the i-th paper P in the dataset i and cited papers, the jth paper P j The reference relationship between For the i-th paper P i and Chapter s s The reference relationship between For the kth author A k and the i-th paper P i The reference relationship between express;

[0100] Then the heterogeneous graph is G = (V, E), where the node set Edge Set

[0101] If there is no author information in the data set, the author node set is an empty set, so the constructed heterogeneous graph is G'=(V',E'), where

[0102] Step 4: Extract representations of heterogeneous graphs, use dynamic heterogeneous graph networks (such as HGT models) to iteratively update node features, adaptively adjust the weights of different node relationships, and dynamically aggregate the features of neighboring nodes to generate node feature vectors of heterogeneous graphs:

[0103] S4.1. Input the heterogeneous graph G into the HGT model. Any type of node v∈V has a corresponding feature vector X v Represents; suppose there are N nodes and the feature dimension is d, then:

[0104] X v ∈R N×d

[0105] S4.2. The features of each node are propagated through its neighboring nodes. For each node v, the feature information of its neighboring nodes N(v) is updated by the following formula:

[0106]

[0107] in, is the feature vector of the neighbor node v in the lth layer, W (l) is the weight matrix of the lth layer, C vu It is a normalization factor based on the degree of neighbor nodes and is usually defined as: Among them, deg(v) and deg(u) are the degrees of node v and neighbor node u respectively;

[0108] S4.3. Calculate the attention coefficient e of node v to its neighbor node u in each layer (i.e., a layer in the graph neural network) by the following formula: vu :

[0109]

[0110] Among them, ReLU() is the activation function, a is the learned attention vector, || represents the vector concatenation processing, is the feature vector of node v in the lth layer; is the feature vector of neighbor node u in the lth layer; W (l) is the weight matrix of the lth layer, used to transform node features; e vu It is the attention coefficient of node v to its neighbor node u, indicating the influence of the neighbor node.

[0111] S4.4. Normalize the attention coefficient through the Softmax function to obtain the influence weight α of node u on node v vu :

[0112]

[0113] By weighted aggregation of neighboring node features, the feature vector representation of node v is updated:

[0114]

[0115] S4.5. For node v, calculate the semantic attention weight β under different meta-paths v (m) :

[0116]

[0117] Among them, a m is the weight vector of semantic attention under the mth path, M is the total number of meta-paths, and m represents the path ordinal number;

[0118] S4.6. Update the node feature vector representation of node v:

[0119]

[0120] The multi-layer graph neural network gradually extracts and learns higher-order graph structure information by stacking multiple layers. Each layer uses the node features of the previous layer for weighted aggregation and updates the node representation through certain transformations (such as linear transformations or activation functions), where the features of each node are aggregated and updated according to the features of its neighboring nodes. Each layer of the GAT (graph attention network) used in this embodiment dynamically adjusts the influence of neighboring nodes on the update of target node features by calculating the attention coefficient between the node and the neighboring node.

[0121] Step 5: Fuse the node feature vector of the heterogeneous graph with the text feature vector of the text T to form the final classification feature, and input it into the classifier for citation intent classification prediction. This includes the following process:

[0122] S5.1. Use element-level multiplication operation to fuse the node feature vector of the heterogeneous graph with the text feature vector of the text T to obtain the fused feature F:

[0123] F=H v ⊙h T

[0124] S5.2, classification through fully connected layer and Softmax activation function:

[0125] P(y|F)=Softmax(W f F+b f )

[0126] Among them, W f F is the classifier weight matrix, b f is the bias setting, and P(y|F) is the predicted category probability distribution.

[0127] Step 6: Use the loss function to measure the difference between the classification prediction result and the true label and optimize the model parameters. The loss function uses the cross entropy loss function:

[0128]

[0129] Among them, N is the number of samples, C is the number of categories, and y ij is the true label of sample i, is the model's predicted probability for the jth category of sample i.

[0130] This example selects two publicly available datasets, namely ACL-ARC and SciCite. The ACL-ARC dataset focuses on citation context annotation in the field of computational linguistics and covers six different citation features: background (51%), purpose (19%), comparison / contrast (18%), motivation (5%), extension (4%), and future work (4%).

[0131] The SciCite dataset covers 6,600 papers from the fields of computer science and medicine, classified into background information (58%), methods (29%), and comparison of results (13%). This classification scheme merges multiple categories into the background category and is applicable to scientific literature in different disciplines. The interdisciplinary nature of this dataset makes it an ideal choice for studying multi-field scientific discourse analysis.

[0132] To verify the effectiveness of the method of the present invention, the method of the present invention (DHGFN) was compared with the current mainstream classification methods, including:

[0133] (1) SciBERT (BERT model based on scientific literature): SciBERT is a pre-trained language model based on BERT, which is unsupervised pre-trained on a large-scale corpus of scientific publications. It performs well in processing scientific domain-specific terminology and complex syntactic structures. SciBERT is selected as the baseline model to evaluate the performance of language models pre-trained for scientific texts in the citation intent classification task, so as to verify whether DHGFN can provide more advantageous performance improvements after combining structural information.

[0134] (2) DMTL (Deep Multi-Task Learning): This method combines citation intent classification with three auxiliary tasks: prediction of the chapter in which the citation appears, judgment of whether a sentence needs to be cited, and tasks based on the title information of the cited paper. By sharing the underlying parameters, DMTL effectively enhances the performance of citation intent classification using a multi-task learning framework. DMTL was chosen mainly because it introduces contextual associations in multi-task learning, which can better mine the contextual information of citations and provide a comparative reference for DHGFN to verify the role of multi-dimensional information in citation intent classification.

[0135] (3) GAT (Graph Attention Network): GAT uses the graph attention mechanism to encode each node in the graph, and assigns different attention weights to the features of neighboring nodes. This allows for more flexible encoding of node relationships in the reference network. GAT can capture fine-grained dependencies between nodes and is suitable for scenarios with complex reference graph structures. This is why we chose GAT, which aims to help evaluate the impact of the reference network structure on classification results.

[0136] (4) GraphCite (graph embedding citation intent classification model): This method combines the text embedding of citation phrases and the structural information of the citation network. It extracts text semantic features through SciBERT and encodes the citation graph structure using the graph attention network (GAT). It pays special attention to the relationship between paper and author nodes to generate a richer representation of citation intent. Since it successfully integrates text and graph structure features, choosing GraphCite as a baseline helps to verify DHGFN's improvement in modeling the interaction between text and structural information.

[0137] (5) MTCIC (Multi-task Citation Intent Classification Model) (SOTA) (State-of-the-art model): This method jointly learns multiple citation-related tasks through a multi-task learning framework, including citation intent classification, citation chapter prediction, and citation value classification. MTCIC is designed to enhance the generalization ability of the model through the interaction between multiple tasks, and has demonstrated its strong performance in the citation intent classification task. Because it combines heterogeneous information and performs multi-task learning, it has become an important control model for evaluating DHGFN's ability to process multimodal information.

[0138] Experiment 1:

[0139] The results of classification prediction using the ACL-ARC dataset are shown in the following table:

[0140]

[0141]

[0142] As can be seen from the above table, DHGFN has achieved significant improvements in the citation intent classification task. First, DHGFN has achieved the best performance in all indicators. Compared with the most advanced model MTCIC, the precision (P) has increased by 1.84 percentage points, the recall (R) has increased by 5.48 percentage points, and the F1 (the harmonic mean of precision and recall) value has increased by 4.81 percentage points. Such an improvement is significant in the field of natural language processing, indicating that DHGFN has stronger performance and generalization ability in the citation intent classification task.

[0143] Compared with other baseline models, DHGFN has a more obvious performance advantage. SciBERT, a language model pre-trained based on scientific text, achieved an F1 value of 63.58%, which is significantly lower than DHGFN. This shows that relying solely on pre-trained language models cannot fully capture the complex semantics and structural information in citation intent. The DMTL model achieved an F1 value of 73.71% by combining multiple auxiliary tasks, but it still lags behind DHGFN by about 6.88 percentage points. This may be because DMTL failed to effectively coordinate the shared information between tasks in multi-task learning, resulting in limited performance improvement of the main task. The GraphCite model improved its F1 value to 77.34% by fusing text embedding and graph embedding, which is close to the level of MTCIC. This shows that combining text and graph structure information plays an important role in improving model performance.

[0144] Experiment 2:

[0145] The results of classification prediction using the Sci-Cite dataset are shown in the following table:

[0146]

[0147]

[0148] As can be seen from the table above, DHGFN achieved the best performance in all indicators, with a precision (P) of 90.90%, a recall (R) of 89.65%, and an F1 value of 90.26%. This result is significantly better than other baseline models and existing state-of-the-art models, reflecting the excellent performance of DHGFN in the citation intent classification task.

[0149] Comparing DHGFN with its closest competitor GraphCite, the key indicator F1 has improved by about 1.72 percentage points. This improvement shows that DHGFN has an advantage in capturing the complex semantic and structural features of citation intent. Specifically, DHGFN more effectively integrates text embedding and graph structure information, thereby improving the overall performance of the model. Secondly, the performance gap between DHGFN and DMTL is also worthy of attention. DHGFN improved the precision and recall by 2.28 and 2.90 percentage points respectively, and the F1 value increased by 2.62 percentage points. This further proves that DHGFN has stronger generalization ability and stability in the task of citation intent classification.

[0150] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0151] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (modules, systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0152] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0153] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0154] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0155] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for classifying citation intent based on a dynamic heterogeneous graph fusion network, characterized in that: The method comprises: Step 1: For each quoted sentence in the dataset, dynamically extract its context sentences, select the most relevant context sentence set based on the semantic similarity between the quoted sentence and the context sentences, and combine the chapter information to form the chapter-enhanced context text T; Step 2: Use the SciBERT pre-trained model to encode the chapter-enhanced context text T into a text feature vector; Step 3: Construct a heterogeneous graph. The nodes in the heterogeneous graph include paper nodes, chapter nodes, and author nodes. The node features are represented by text features or entity information. The edge relationships are used to express the reference relationships, chapter association relationships, or author relationships between nodes. Step 4: Extract representations of the heterogeneous graph, use a dynamic heterogeneous graph network to iteratively update node features, adaptively adjust the weights of different node relationships, and dynamically aggregate the features of neighboring nodes to generate a node feature vector of the heterogeneous graph; Step 5: Fuse the node feature vector of the heterogeneous graph with the feature vector of the chapter-enhanced context text T to form the final classification feature, which is input into the classifier for citation intent classification prediction; Step 6: Use the loss function to measure the difference between the classification prediction result and the true label and optimize the model parameters.

2. The method for classifying reference intentions based on a dynamic heterogeneous graph fusion network according to claim 1 is characterized in that: Step 1 includes the following process: S1.

1. Get the reference sentence C from the dataset cite ; S1.2, extract the three sentences before and after the quoted sentence, and form a candidate sentence set C = {C i }; S1.

3. For each candidate sentence C i , calculate its difference with the reference sentence C cite The cosine similarity (C cite ,C i ): Among them, v cite and v i They are the vector representations of the quoted sentence and the candidate sentence respectively; S1.

4. The quoted sentence and the most relevant context sentence are merged to form text C containing quoted context information: Select and quote sentence C from the candidate sentence set cite Most relevant sentence C rel : Use the concatenation operation to quote sentence C cite and the most relevant sentence C rel Combine and concatenate to form text C containing the reference context information; S1.

5. The chapter name is merged with the text C containing the reference context information to form a chapter-enhanced context text T.

3. The method for classifying reference intentions based on a dynamic heterogeneous graph fusion network according to claim 1 is characterized in that: Step 2 includes the following process: S2.

1. For the chapter enhanced context text T = {w1, w2, ..., w n }, and use the SciBERT model to generate the corresponding word-level feature representation {h1,h2,...,h n }: {h1,h2,…,h n }=SciBERT({w1,w2,…,wn}) Among them, w i represents the i-th word in the text, h i For word w i Embedding vector representation in the model; S2.

2. Take the hidden state h corresponding to the [CLS] tag [CLS] The feature vector representation of the context text T as chapter enhancement: T=h [CLS] ∈R d Where R d is the hidden layer dimension of the SciBERT model; S2.

3. Normalize the text feature vector to obtain the text feature representation h T .

4. The method for classifying reference intentions based on a dynamic heterogeneous graph fusion network according to claim 1 is characterized in that: The heterogeneous graph described in step 3 includes nodes and edges, wherein: Node: the i-th paper P in the dataset i Use Node It means that its eigenvector is Chapter s u s Use Node It means that the eigenvector is The kth author A k Use Node It means that the eigenvector is Edge: For the i-th paper P in the dataset i and cited papers, the jth paper P j The reference relationship between For the i-th paper P i and Section sI s The reference relationship between For the kth author A k and the i-th paper P i The reference relationship between express; Then the heterogeneous graph is G = (V, E), where the node set Edge Set 5. The method for classifying reference intentions based on a dynamic heterogeneous graph fusion network according to claim 3 or 4, characterized in that: If there is no author information in the data set, the author node set is an empty set, so the constructed heterogeneous graph is G'=(V',E'), where 6. The method for classifying reference intentions based on a dynamic heterogeneous graph fusion network according to claim 4 is characterized in that: Step 4 includes the following process: S4.

1. Input the heterogeneous graph G into the HGT model. Any type of node v∈V has a corresponding feature vector X v Represents; suppose there are N nodes and the feature dimension is d, then: X v ∈R N×d S4.

2. The features of each node are propagated through its neighboring nodes. For each node v, the feature information of its neighboring nodes N(v) is updated by the following formula: in, is the feature vector of the neighbor node v in the lth layer, W (l) is the weight matrix of the lth layer, C vu is the normalization factor; S4.

3. Calculate the attention coefficient e of node v to its neighbor node u in each layer by the following formula vu : Among them, ReLU() is the activation function, a is the learned attention vector, || represents the vector concatenation processing, is the feature vector of node v in the lth layer; is the feature vector of neighbor node u in the lth layer; W (l) is the weight matrix of the lth layer, used to transform node features; e vu is the attention coefficient of node v to its neighbor node u, indicating the influence of the neighbor node; S4.

4. Normalize the attention coefficient through the Softmax function to obtain the influence weight α of node u on node v vu : By weighted aggregation of neighbor node features, the feature vector representation of node v is updated: S4.

5. For node v, calculate the semantic attention weights under different meta-paths Among them, a m is the weight vector of semantic attention under the mth path, M is the total number of meta-paths, and m represents the path ordinal number; S4.

6. Update the node feature vector representation of node v:

7. The method for classifying reference intentions based on a dynamic heterogeneous graph fusion network according to claim 6 is characterized in that: Step 5 includes the following process: S5.

1. Use element-level multiplication operation to fuse the node feature vector of the heterogeneous graph with the text feature vector of the text T to obtain the fused feature F: F=H v ⊙h T S5.2, classification through fully connected layer and Softmax activation function: P(y|F)=Softmax(W f F+b f ) Among them, W f F is the classifier weight matrix, b f is the bias setting, and P(y|F) is the predicted category probability distribution.

8. The method for classifying reference intentions based on a dynamic heterogeneous graph fusion network according to claim 7 is characterized in that: The loss function described in step 6 adopts the cross entropy loss function.

Images