A reference intention classification method based on a dynamic heterogeneous graph fusion network

Abstract

The application discloses a kind of reference intention classification methods based on dynamic heterogeneous graph fusion network.Method includes: dynamically extracting reference sentence and its related context, combine chapter information to form text T;Using SciBERT pre-training model, the feature vector of text T is extracted;Complex relationship between nodes such as paper, chapter and author is expressed by constructing heterogeneous graph;Node features are extracted by dynamic heterogeneous graph network, and the node feature vector of heterogeneous graph is generated;The text feature vector of text T is fused with node feature vector, to form comprehensive features;Finally, the reference intention is predicted using classifier, and the model is optimized by loss function.The application extracts distinguishing semantic features from the internal structure of the paper and the cross-document relationship, enabling high-precision reference intention recognition in complex citation networks, and is particularly suitable for reference intention classification tasks in academic literature.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of scientific metrology and academic literature analysis, and particularly relates to a citation intention classification method based on a dynamic heterogeneous graph fusion network. BACKGROUND

[0002] Understanding citation intention has important practical significance for academic literature analysis. First, it helps literature evaluation in scientific metrology. Accurate identification of the citation type of a paper can more objectively measure the influence of the paper in the academic community. Second, citation intention classification also has wide value in academic recommendation systems, knowledge graph construction, academic search engine optimization, and other applications.

[0003] With the rapid development of machine learning and natural language processing (NLP), automatic methods for citation intention classification have gradually increased. Heterogeneous graph classification methods have received widespread attention due to their advantages in multi-modal information integration and complex relationship capture. These methods improve text classification results 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, lack dynamic mechanisms specifically designed for citation intention classification, and are difficult to adapt to the changing context and semantic requirements in citation relationships. These methods have poor generalization ability in specific citation intention identification tasks and are difficult to accurately depict the real intentions of academic authors.

[0004] For example, the patent document with publication number CN114564573A discloses an academic cooperation relationship prediction method based on a heterogeneous graph neural network. Its heterogeneous graph contains author nodes, work nodes, affiliation nodes, and content nodes, mainly used to express the topological structure of scholars' cooperation relationships and interest nodes. It uses DeepWalk and Text-CNN to pretrain node features, and combines meta-paths through a multi-head attention mechanism to aggregate nodes. This scheme is very basic and cannot capture deep semantic information of the heterogeneous graph structure. The patent document with publication number CN114722192A discloses a discipline inspection clue multi-label classification method based on a heterogeneous graph neural network. It constructs a heterogeneous graph structure of sentence and word nodes, decomposes the entire text into the connection relationship of words and sentences, and uses GNN to aggregate syntactic structure information. This structure is suitable for general multi-label text classification tasks and realizes classification based on the grammatical relevance of word and sentence nodes. The model uses Bi-LSTM and an attention mechanism to aggregate the context of sentence nodes, but is limited to pre-defined syntactic context and does not dynamically adapt to the specific semantics of citation sentences.

[0005] It can be seen that although the previous heterogeneous graph classification method can capture the surface relationship by using the heterogeneity of nodes and edges, it is difficult to effectively obtain the deep semantic information in the citation intention. The citation intention classification task usually needs to identify the potential intention of the academic author when citing a certain literature, which involves deep understanding of the citation sentence, context and the relationship between multiple layers of nodes. However, these heterogeneous graph methods are often limited to shallow connection and simple relationship representation, and it is difficult to completely capture the semantic depth and subtlety required by the citation intention. Therefore, when facing the complex context semantics in the citation intention, the recognition effect of such methods is often not satisfactory, which limits the performance in practical applications. Obviously, the existing heterogeneous graph text classification method still faces significant challenges in the citation intention classification task. Therefore, it is necessary to design a new solution to better adapt to the dynamics and complexity of citation intention classification, so as to improve the accuracy and applicability of citation intention classification. SUMMARY

[0006] In order to solve the problems of the prior art, the present application provides a citation intention classification method based on a dynamic heterogeneous graph fusion network, which selects the most relevant context sentences from the paper text based on the citation sentence, introduces chapter information as an auxiliary feature for citation intention recognition, constructs a heterogeneous graph to express the complex relationship between paper nodes, chapter nodes and author nodes, enriches the feature representation of citation intention classification, and uses the characteristics of dynamic adjustment of the heterogeneous graph to aggregate adjacent nodes and give adaptive weights to different node relationships, so that the model can extract more discriminative semantic features from the internal structure of the paper and the relationship between the papers, and can more accurately identify the citation intention in the complex citation network, especially suitable for the citation intention classification task in academic literature.

[0007] The embodiment of the present application provides the following scheme:

[0008] The embodiment of the present application provides a citation intention classification method based on a dynamic heterogeneous graph fusion network, which comprises:

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

[0010] Step 2: Use the SciBERT pre-training 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, and the edge relationship is used to express the citation relationship, chapter association relationship or author relationship between nodes;

[0012] Step four, perform feature extraction on the heterogeneous graph, use the dynamic heterogeneous graph network to iteratively update the node features, adaptively adjust the weights of different node relationships, and dynamically aggregate the features of neighbor nodes to generate the node feature vector of the heterogeneous graph;

[0013] Step five, fuse the node feature vector of the heterogeneous graph with the chapter-enhanced context text T to form the final classification feature, and input the classifier for reference intention classification prediction;

[0014] Step six, measure the difference between the classification prediction result and the true label through the loss function, and optimize the model parameters.

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

[0016] S1.1, obtain the reference sentence C from the data set cite ;

[0017] S1.2, extract each 3 sentences before and after the reference sentence to form a candidate sentence set C = {C i};

[0018] S1.3, for each candidate sentence C i , calculate the cosine similarity Similarity(C cite ,C i ) between it and the reference sentence C cite :

[0019]

[0020] Where v cite and v i are the vector representations of the reference sentence and the candidate sentence respectively;

[0021] S1.4, fuse the reference sentence and the most relevant context sentence to form a text C containing reference context information:

[0022] Select the most relevant sentence C rel from the candidate sentence set to the reference sentence C cite :

[0023]

[0024] Use the concatenation operation to combine the reference sentence C cite and the most relevant sentence C rel , and the concatenated text C containing reference context information is formed;

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

[0026] In an alternative embodiment, step two comprises the following process:

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

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

[0029] where w i represents the i-th word in the text, and h i is the embedding vector representation of the word w i in the model;

[0030] S2.2, take the hidden state h [CLS] corresponding to the [CLS] mark as the feature vector representation of the chapter-enhanced context text T:

[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 alternative embodiment, the heterogeneous graph of step three comprises nodes and edges, wherein:

[0035] Node: the i-th paper P i in the dataset is represented by node , and its feature vector is the s-th chapter u s is represented by node , and its feature vector is the k-th author A k is represented by node , and its feature vector is

[0036] Edge: for the i-th paper P i in the dataset and the cited paper, the j-th paper P j , the citation relationship between them is represented by edge ; for the i-th paper P i and the s-th chapter u sThe reference relationship between the authors and the papers is represented by edges ; the reference relationship between the kth author A k and the ith paper P i is represented by edges ;

[0037] The heterogeneous graph is G=(V, E), where the node set V and the edge set E

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

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

[0040] S4.1, input the heterogeneous graph G into the HGT model, and each node v∈V has a corresponding feature vector X v ; assuming that there are N nodes and the feature dimension is d, then:

[0041]

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

[0043]

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

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

[0046]

[0047] wherein ReLU() is an activation function, a is a learned attention vector, || represents vector splicing processing, is the feature vector of node v in the lth layer of the graph layer; is the feature vector of neighbor node y in the lth layer of the graph layer; W (l) is the weight matrix of the lth layer, used to transform the 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 coefficients by a Softmax function to obtain the influence weight a of node u on node v vu :

[0049]

[0050] Update the feature vector representation of node v by weighted aggregation of the features of the neighbor nodes:

[0051]

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

[0053]

[0054] Where 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, fuse the node feature vector of the heterogeneous graph with the text feature vector of the text T by element-level multiplication operation to obtain the fused feature F:

[0059] F = H v ⊙h T

[0060] S5.2, classify by a fully connected layer and a Softmax activation function:

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

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

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

[0064] The present application has the beneficial effects based on the technical solutions:

[0065] (1) The citation intention classification method based on the dynamic heterogeneous graph fusion network provided by the application introduces a "paper-section-author" multi-level node structure, the section information provides higher-level context information for the citation intention, can make the model have in-depth understanding of the background of the citation intention, by incorporating the specific structure (such as section and author) in the academic literature into the heterogeneous graph, the context mode specific to the academic literature can be captured, and the domain adaptability of the model is further improved, especially the citation intention in the academic literature can be captured.

[0066] (2) The citation intention classification method based on the dynamic heterogeneous graph fusion network provided by the application designs a dynamic context extraction strategy at the text level, extracts the most relevant context sentences centering on the citation sentence, and embeds the section information as an auxiliary feature to enhance the recognition of the citation intention. The dynamic context can be flexibly adapted to different citation situations in actual application, and the section information can provide clear context guidance for the citation intention. This dynamic and high-level context combination strategy has not been realized in general multi-label text classification, and is especially suitable for the requirements of context understanding in citation intention classification.

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

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present specification or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present specification, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0069] Figure 1 The flowchart of the present application.

[0070] Figure 2 The process flowchart of the embodiment of the present application. DETAILED DESCRIPTION

[0071] With reference to the drawings and the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art belong to the scope of protection of the embodiments of the present application.

[0072] With reference to Figure 1 and Figure 2 , the embodiments of the present application provide a reference intention classification method based on a dynamic heterogeneous graph fusion network, which comprises the following steps:

[0073] Step 1, for each reference sentence in the data set, dynamically extract its context sentences, filter out the most relevant context sentence set based on the semantic similarity between the reference sentence and the context sentence, and combine the chapter information to form the chapter-enhanced context text T. Including the following processes:

[0074] S1.1, obtaining the reference sentence C from the data set cite ;

[0075] S1.2, extracting each 3 sentences before and after the reference sentence to form a candidate sentence set C = {C i};

[0076] S1.3, for each candidate sentence C i , calculate the cosine similarity Similarity(C cite ,C i ) between it and the reference sentence C cite :

[0077]

[0078] Wherein, v cite and v i are the vector representations of the reference sentence and the candidate sentence respectively;

[0079] S1.4, fuse the reference sentence and the most relevant context sentence to form a text C containing reference context information:

[0080] Select the most relevant sentence C rel from the candidate sentence set to the reference sentence C cite :

[0081]

[0082] Once the most relevant sentence C rel is determined, the reference sentence C cite and the most relevant sentence C relCombine them. Concatenation can be achieved by directly joining the vector representations of the two sentences into a single long vector:

[0083] C = [C cite C rel ]

[0084] Among them, the symbol [C cite C rel [] indicates that the quoted sentence is represented by C. cite The representation of the most relevant sentence C rel Perform concatenation. The resulting text C, containing contextual information, serves as input for subsequent tasks.

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

[0086] T = [S; C]

[0087] Where 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] By concatenating the chapter title S and the text C containing citation context information, a new text is formed, namely the chapter-enhanced context text T. This text contains the chapter's topic information and enhanced context of the cited sentences. The concatenated text can provide richer semantic information for subsequent tasks, helping the model better understand the intent behind the citations.

[0089] Step 2: Using the SciBERT pre-trained model, encode the enhanced context text T into text feature vectors, which serve as the initial feature representations for the corresponding nodes in subsequent heterogeneous graph construction. This includes the following processes:

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

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

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

[0093] S2.2, take the hidden state h corresponding to the [CLS] mark [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 get the text feature representation h T .

[0097] Step three, build 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, and the edge relationship is used to express the citation relationship between nodes, chapter association relationship or author relationship.

[0098] Node: the i-th paper P in the dataset i is represented by node , and its feature vector is The s-th chapter u s is represented by node , and its feature vector is The k-th author A k is represented by node , and its feature vector is

[0099] Edge: for the i-th paper P in the dataset i and the cited paper, the j-th paper P j The citation relationship between them is represented by edge ; for the i-th paper P i and the s-th chapter u s The citation relationship between them is represented by edge ; for the k-th author A k and the i-th paper P i The citation relationship between them is represented by edge ;

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

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

[0102] Step four, isomerism graph is characterized by extracting, using dynamic isomerism graph network (such as HGT model) to update node features iteratively, adaptively adjust the weight of different node relationship, and dynamically aggregate the features of neighbor nodes, generate node feature vector of isomerism graph:

[0103] S4.1, the isomerism graph G is input into the HGT model, and the node v of any type has a corresponding feature vector X v ; assuming that 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 neighbor nodes, and for each node v, the feature information of its neighbor nodes N(v) is updated through the following formula:

[0106]

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

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

[0109]

[0110] wherein, ReLU() is an activation function, a is a learned attention vector, || represents vector splicing processing, is the feature vector of node v in the lth layer of graph layer; is the feature vector of neighbor node u in the lth layer of graph 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 neighbor nodes.

[0111] S4.4, the attention coefficient is normalized by the Softmax function to obtain the influence weight a vu of node u to node v:

[0112]

[0113] The feature vector representation of node v is updated by weighted aggregation of the features of the neighbor nodes:

[0114]

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

[0116]

[0117] where 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 graph layers. Each layer will use the node features of the previous layer for weighted aggregation and update the node representation through certain transformations (such as linear transformation or activation function), where the features of each node are aggregated and updated according to the features of its neighbor nodes. The GAT (graph attention network) used in this embodiment dynamically adjusts the influence of neighbor nodes on the feature update of the target node by calculating the attention coefficient between the node and its neighbor nodes at each layer.

[0121] Step five, fuse the node feature vector of the heterogeneous graph with the text feature vector of the text T to form the final classification features and input the classifier for reference intent classification prediction. Including the following processes:

[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, classify through a fully connected layer and a Softmax activation function:

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

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

[0127] Step six, measure the difference between the classification prediction result and the true label by the loss function, optimize the model parameters. The loss function adopts cross-entropy loss function:

[0128]

[0129] Where N is the number of samples, C is the number of categories, y ij is the true label of sample i, is the prediction probability of the jth category of the model for sample i.

[0130] Two publicly available datasets, ACL-ARC and SciCite, are selected in this embodiment. The ACL-ARC dataset focuses on the annotation of citation context in the field of computational linguistics and covers six different citation functions: background (51%), use (19%), comparison / contrast (18%), motivation (5%), extension (4%), and future work (4%).

[0131] The SciCite dataset covers 6600 papers from the fields of computer science and medicine, classified into background information (58%), methods (29%), and results comparison (13%). This classification scheme combines multiple categories into the background category and is suitable for scientific literature in different disciplines. The interdisciplinary nature of this dataset makes it an ideal choice for studying multi-disciplinary scientific discourse analysis.

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

[0133] (1) SciBERT (SciBERT based on BERT model): SciBERT is a pre-trained language model based on BERT, which is pre-trained unsupervisedly through a large-scale scientific publication corpus. It performs well in handling scientific domain-specific terminology and complex syntactic structures. SciBERT is chosen as the baseline model to evaluate the performance of the language model pre-trained specifically for scientific text in the citation intent classification task, so as to verify whether DHGFN can provide more advantageous performance improvement after combining structural information.

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

[0135] (3) GAT (Graph Attention Network): GAT uses a graph attention mechanism to encode each node in the graph. It assigns different attention weights to the features of neighboring nodes, allowing for more flexible encoding of node relationships in citation networks. GAT can capture fine-grained dependencies between nodes, making it suitable for scenarios where the citation graph structure is complex. This is the reason we choose GAT, aiming to help evaluate the impact of citation network structure on classification results.

[0136] (4) GraphCite: This method combines both text embedding of citation phrases and structural information of citation networks. It extracts text semantic features using SciBERT and encodes the citation graph structure using a graph attention network (GAT). It focuses on the relationships between papers and authors, resulting in a richer representation of citation intent. Since it successfully integrates text and graph structure features, selecting GraphCite as a baseline helps to test the improvements of DHGFN in modeling the interaction between text and structure information.

[0137] (5) MTCIC (Multi-Task Citation Intent Classification) (SOTA): This method uses a multi-task learning framework to jointly learn multiple citation-related tasks, including citation intent classification, citation chapter prediction, and citation value classification. MTCIC is designed to enhance the model's generalization ability through interactions between multiple tasks, demonstrating its strong performance in the citation intent classification task. Since it combines heterogeneous information and performs multi-task learning, it becomes an important contrast model for evaluating the ability of DHGFN to handle multi-modal information.

[0138] Experiment 1:

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

[0140]

[0141]

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

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

[0144] Experiment Two:

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

[0146]

[0147]

[0148] From the above table, DHGFN achieved the best performance on 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 the most advanced model, demonstrating the outstanding performance of DHGFN in the citation intent classification task.

[0149] The DHGFN improves about 1.72 percentage points compared with the closest competitor GraphCite in the F1, a key indicator. This improvement indicates that the DHGFN has an advantage in capturing the complex semantic and structural features of citation intent. Specifically, the DHGFN more effectively integrates text embedding and graph structure information, thereby improving the overall performance of the model. Second, the performance gap between the DHGFN and the DMTL is also worth noting. The DHGFN improves by 2.28 and 2.90 percentage points in precision and recall, respectively, and by 2.62 percentage points in F1. This further demonstrates that the DHGFN has stronger generalization ability and stability in the citation intent classification task.

[0150] Those skilled in the art will appreciate that embodiments of the application can be supplied as methods, systems, or computer program products. Accordingly, the application can be embodied in the form of complete hardware embodiments, complete software embodiments, or embodiments combining software and hardware aspects. Furthermore, the application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) having computer usable program code embodied thereon.

[0151] The present application is described in reference to the flowcharts and / or block diagrams of the methods, apparatus (modules, systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, 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 apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus generate a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in the flow or multiple flows and / or blocks Figure 1 The functions specified in the flow or multiple flows and / or blocks

[0152] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in the flow or multiple flows and / or blocks Figure 1 The functions specified in the flow or multiple flows and / or blocks

[0153] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide a means for implementing the functions specified in the flowcharts and / or block diagrams.Figure 1 one or more processes and / or functions specified in one or more blocks Figure 1 one or more processes and / or functions specified in one or more blocks

[0154] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the preferred embodiments by those of skill in the art once they have the benefit of the present disclosure. Therefore, the appended claims are intended to encompass within their scope all possible alternatives and modifications to the preferred embodiments. 1

[0155] It is apparent that a person skilled in the art can make a variety of changes and modifications to the application without departing from the spirit and scope thereof. Thus, if these modifications and changes fall within the scope of the claims and their equivalents, it is intended to include them in the application.

Claims

1. A method for reference intention classification based on a dynamic heterogeneous graph fusion network, characterized in that, The method comprises: Step one, for each citation sentence in the data set, dynamically extract its context sentences, based on the semantic similarity between the citation sentence and the context sentence, filter out the most relevant context sentence set, combined with chapter information, form chapter enhanced context text T ; Step two, use SciBERT pre-trained model to chapter-enhanced contextual text encoded as a text feature vector; Step three, constructing a heterogeneous graph, the nodes in the heterogeneous graph including paper nodes, chapter nodes and author nodes, the node features being represented by text features or entity information, and the edge relationships being used to express the citation relationship, chapter association relationship or author relationship between the nodes; the heterogeneous graph comprises nodes and edges, wherein: Node: the i-th data set i The paper Node , whose eigenvector is ; the i-th section s The section Node , whose eigenvector is ; the i-th author k The author Node , whose eigenvector is ; Edge: for the reference relationship between the i-th paper and the j-th paper in the dataset i Paper and cited paper, the reference relationship between the i-th paper and the j-th paper is represented by edge j Paper The reference relationship between the i-th paper and the j-th chapter is represented by edge Paper i The reference relationship between the i-th paper and the j-th chapter is represented by edge Paper s The reference relationship between the i-th paper and the j-th chapter is represented by edge Paper The reference relationship between the i-th paper and the j-th chapter is represented by edge k Author The reference relationship between the i-th paper and the j-th chapter is represented by edge i Paper The reference relationship between the i-th paper and the j-th chapter is represented by edge Paper The isomorphic graph is wherein the node set , the edge set ; Step four, performing feature extraction on the heterogeneous graph, iteratively updating the node features using a dynamic heterogeneous graph network, adaptively adjusting the weights of different node relationships, and dynamically aggregating the features of neighbor nodes to generate a node feature vector of the heterogeneous graph; comprising the following processes: S4.1, input the heterogeneous graph G into the HGT model, an arbitrary type of node with the corresponding feature vector represent; provided with N nodes, the feature dimension is d then: ， S4.2, the features of each node are propagated through its neighbor nodes, for each node v whose neighbor nodes are updated by the following equation: ， wherein, is the feature vector of the l layer neighbor node, v is the weight matrix of the l layer, is a normalization factor;​ S4.3, calculate the attention coefficient e of each layer in the node v to its neighbor nodes u vu :​ ， wherein, ReLU () is an activation function, a is a learned attention vector, and || denotes vector concatenation, is a node In the first layer, the feature vector of the layer; is a neighbor node In the first layer, the feature vector of the layer; is a weight matrix of the first layer, used to transform the node feature; is a node The attention coefficient of its neighbor node , indicating the influence of the neighbor node; S4.4, normalize the attention coefficients by a Softmax function to get the neighbor nodes the influence weight of the node on the node : ， updating a feature vector representation of the node by weighted aggregation of features of neighboring nodes of the node ， S4.

5. For a node v, compute semantic attention weights under different meta-paths : ， wherein, 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, updating nodes v of the node feature vector representation: ； Step five, the node feature vectors of the heterogeneous graph are fused with the feature vectors of the chapter-enhanced context text to form final classification features, which are input into a classifier for reference intention classification prediction; Step five, the node feature vectors of the heterogeneous graph are fused with the feature vectors of the chapter-enhanced context text to form final classification features, which are input into a classifier for reference intention classification prediction; Step six, measuring the difference between the classification prediction result and the real label through a loss function, and optimizing the model parameters.

2. The method of claim 1, wherein the dynamic heterogeneous graph fusion network-based reference intention classification method is characterized in that: Step one comprises the following processes: S1.1, obtaining a reference sentence from a data set ; S1.2, extract each 3 sentences before and after the reference sentence to form a candidate sentence set C ={ C i}; S1.3, for each candidate sentence C i , compute its cosine similarity to the reference sentence​ ， wherein, and are the vector representations of the quote sentence and the candidate sentence, respectively; S1.4, fuse the quote sentence and the most relevant context sentence to form a text containing quote context information : selecting a sentence from a set of candidate sentences most relevant sentence : ， using a concatenation operation and the most relevant sentence combined, forming a text containing quote context information ; S1.5, the section name with text containing quote context information fused, forming a section-enhanced context text .

3. The method of claim 1, wherein the dynamic heterogeneous graph fusion network-based reference intention classification method is characterized in that: Step two comprises the following processes: S2.1, For chapter-enhanced context text , generate corresponding word-level feature representation with SciBERT model : ， wherein, represents the i-th word in the text, is the word embedding vector representation in the model;​ S2.2, take the hidden state corresponding to the [CLS] token Contextual text as chapter augmentation Feature vector representation: ， wherein is the hidden layer dimension of the SciBERT model; S2.3, normalizing the text feature vector to obtain a text feature representation .

4. The dynamic heterogeneous graph fusion network-based reference intention classification method according to claim 1 or 3, characterized in that: If there is no author information in the data set, the author node set is empty, so the heterogeneous graph constructed is wherein , .

5. The method of claim 1, wherein the dynamic heterogeneous graph fusion network-based reference intention classification method is characterized in that: Step five comprises the following processes: S5.1, the node feature vector of the heterogeneous graph is fused with the text feature vector of the text T by using an element-level multiplication operation to obtain a fused feature F: ， S5.2, classification is performed through a fully connected layer and a Softmax activation function: ， wherein, is a classifier weight matrix, is a bias setting, is a predicted class probability distribution.

6. The method of claim 5, wherein the dynamic heterogeneous graph fusion network-based reference intention classification method is characterized in that: The loss function in step six adopts a cross-entropy loss function.

Images