Text classification method based on multi-expert fusion gating chart neural network comparative learning

Abstract

The invention discloses a text classification method based on multi-expert fusion gating graph neural network comparative learning, which comprises the following steps: preprocessing an original text, generating a standardized corpus, and extracting words, part-of-speech and entity features; constructing a first heterogeneous graph based on word features, constructing a second heterogeneous graph based on part-of-speech features, and constructing a third heterogeneous graph based on entity similarity; dynamically weighting and fusing the first heterogeneous graph, the second heterogeneous graph and the third heterogeneous graph through an expert gating fusion module to generate a fused graph structure; performing graph convolution coding on the fused graph structure to generate a node representation vector; executing double-layer comparative learning based on the node representation vector: optimizing single sample representation consistency by implementing instance-level comparative learning, and synchronously implementing cluster-level comparative learning to optimize intra-class center aggregation; and inputting the representation vector subjected to comparative learning optimization into a classifier, and outputting a text category label. According to the method, the comprehensive performance of the short text classification model in the aspects of semantic expression, structural modeling and cross-sample discrimination can be effectively improved.

Description

Technical Field

[0001] This invention belongs to the fields of natural language processing, graph neural networks and contrastive learning, and particularly relates to a text classification method based on contrastive learning of multi-expert fusion gated graph neural networks. Background Technology

[0002] With the rapid development of social media platforms, news aggregation services, and online commenting systems, short text data is experiencing explosive growth in various applications, placing higher demands on efficient and accurate text classification technologies. Text classification tasks aim to categorize texts into predefined categories based on their content and are widely used in scenarios such as public opinion monitoring, product recommendation, spam detection, and sentiment analysis.

[0003] Traditional text classification methods mostly rely on bag-of-words models, TF-IDF features, and shallow classifiers (such as SVM and logistic regression). While they perform reasonably well with sufficient data, their expressive power and classification effectiveness are significantly limited when dealing with short texts that are semantically complex and have sparse context. To overcome these problems, researchers have gradually introduced deep learning models, using CNN, RNN, and Transformer architectures to model text and improve semantic capture capabilities.

[0004] In recent years, with the rise of pretrained language models (PLMs) such as BERT, RoBERTa, and DeBERTa, text classification performance has significantly improved. PLMs capture general language patterns on large-scale corpora and demonstrate strong semantic transfer capabilities in downstream tasks. However, relying solely on Transformers for sequence modeling still has certain limitations, especially in scenarios with short texts that are semantically sparse, have weak entity connections, and implicit grammatical structures, where their representation space is difficult to align accurately.

[0005] To enhance the model's ability to capture structured semantics in text, Graph Neural Networks (GNNs) have been widely introduced into short text classification. By constructing heterogeneous graph structures based on word co-occurrence, syntactic dependencies, or semantic similarity, GNNs can achieve efficient information propagation between nodes, thereby improving their ability to model long-distance dependencies and local contexts. However, current graph structure modeling often faces two key problems: first, the graph structure origin is singular, failing to comprehensively cover multi-granularity semantic cues; second, static graph structures struggle to adapt to dynamic semantic changes in different textual contexts, resulting in insufficient generalization ability of graph representations.

[0006] Meanwhile, contrastive learning has garnered widespread attention due to its superior performance in unsupervised representation learning. In text classification, introducing contrastive learning mechanisms can effectively improve the discriminative power of the representation space and enhance the model's ability to identify class boundaries. However, most existing methods only construct positive and negative samples at the instance level, ignoring the semantic distinguishing value of the graph structure itself at the structural level, and failing to consider how to coordinate the complementarity between different graphs, thus limiting the overall performance improvement.

[0007] In summary, current text classification methods have shortcomings in terms of structural modeling granularity, graph information fusion methods, and contrastive learning strategies. To better utilize the complementarity between various semantic graph structures and enhance the model's ability to represent semantic relationships, it is urgent to design a text classification method based on contrastive learning using a multi-expert fusion gated graph neural network to address key issues such as graph information redundancy, representation degradation, and weak classification discrimination. Summary of the Invention

[0008] To address the aforementioned technical issues, this invention proposes a text classification method based on multi-expert fusion gated graph neural network contrastive learning, which can effectively improve the overall performance of short text classification models in terms of semantic representation, structural modeling, and cross-sample discrimination.

[0009] To achieve the above objectives, this invention provides a text classification method based on multi-expert fusion gated graph neural network contrastive learning, comprising:

[0010] The original text is preprocessed to generate a standardized corpus and extract word, part-of-speech, and entity features;

[0011] A first heterogeneous graph is constructed based on the word features, a second heterogeneous graph is constructed based on the part-of-speech features, and a third heterogeneous graph is constructed based on entity similarity.

[0012] The first heterogeneous graph, the second heterogeneous graph, and the third heterogeneous graph are dynamically weighted and fused using an expert-gated fusion module to generate a fused graph structure.

[0013] Graph convolution encoding is performed on the fused graph structure to generate node representation vectors;

[0014] Based on the node representation vector, perform two-layer contrastive learning: implement instance-level contrastive learning to optimize the consistency of single-sample representation, and simultaneously implement cluster-level contrastive learning to optimize intra-class center aggregation;

[0015] The representation vector optimized through contrastive learning is input into the classifier, which outputs text category labels.

[0016] Optionally, the preprocessing steps include:

[0017] The original corpus was uniformly encoded and illegal characters were removed.

[0018] Use a part-of-speech tagging tool to generate a sequence of part-of-speech tags;

[0019] Identify named entities using entity linking tools;

[0020] Pre-trained word embeddings are used to initialize word features, entity embeddings are generated through training triple vectors, and part-of-speech features are encoded using one-hot encoding.

[0021] Construct a text-word TF-IDF pooling matrix and fill in placeholders for empty terms in the part-of-speech and entity channels.

[0022] Optionally, the construction process of the first heterogeneous graph includes:

[0023] Calculate the co-occurrence probability of word pairs using a sliding window;

[0024] Valid word-edge relationships are filtered based on point mutual information thresholds;

[0025] Store the filtering results as a sparse adjacency matrix.

[0026] Optionally, the construction process of the second heterogeneous graph includes:

[0027] Generate part-of-speech co-occurrence relationships based on part-of-speech tag sequences;

[0028] POS-POS connection edges are constructed by calculating point mutual information;

[0029] Store the connection relationships as a sparse matrix.

[0030] Optionally, the construction process of the third heterogeneous graph includes:

[0031] Calculate the cosine similarity between entities;

[0032] Entity pairs with a similarity greater than 0.2 are retained for connection;

[0033] Limit each entity to a maximum of 100 similar neighbor nodes.

[0034] Optionally, the operation of the expert gating fusion module includes:

[0035] Receive the node representation vectors of the first heterogeneous graph, the second heterogeneous graph, and the third heterogeneous graph;

[0036] Three sets of dynamic weight coefficients are generated using a gating network;

[0037] Weighted linear fusion is performed on the three sets of node representations according to the aforementioned weighting coefficients.

[0038] Optionally, the two-layer contrastive learning process includes:

[0039] Instance-level contrastive learning: Generate positive sample pairs through text augmentation to narrow the representation distance of similar samples;

[0040] Cluster-level comparative learning: Aggregate the sample center representation within a class to increase the distance between different clusters.

[0041] Optionally, the classification process includes:

[0042] Input the fused graph representation vector into the multilayer perceptron;

[0043] The class probability distribution is calculated using the Softmax function;

[0044] Supervised training is performed using label-smooth cross-entropy loss.

[0045] Technical effects of the invention: The invention discloses a text classification method based on multi-expert fusion gated graph neural network contrastive learning, which can effectively improve the comprehensive performance of short text classification models in semantic expression, structural modeling and cross-sample discrimination. Attached Figure Description

[0046] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0047] Figure 1 This is a flowchart illustrating the text classification method based on multi-expert fusion gated graph neural network contrastive learning according to an embodiment of the present invention.

[0048] Figure 2 This is an example of a short text multi-classification task in an embodiment of the present invention;

[0049] Figure 3 This is a diagram illustrating the overall structural framework of the text classification method based on multi-expert fusion gated graph neural network contrastive learning, as described in an embodiment of the present invention.

[0050] Figure 4 This is a schematic diagram of the expert gating fusion module (MoE) according to an embodiment of the present invention. Detailed Implementation

[0051] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0052] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0053] like Figure 1As shown, this embodiment provides a text classification method based on multi-expert fusion gated graph neural network contrastive learning, including:

[0054] A. Multi-source expert graph construction stage: Semantic rule features at multiple levels are extracted from the input text, including part-of-speech co-occurrence, TF-IDF similarity, entity type attribution, etc., to construct multiple heterogeneous graph structures. Each graph structure generates an adjacency matrix by weighting the domain semantic relevance between node pairs, and then the graph fusion module performs weighted combination to obtain the global graph structure. This graph is used to model the integration of multi-source prior knowledge in structural representation.

[0055] B. Multi-graph feature extraction stage: Feature encoding is performed on the original dependency graph and the expert fusion graph respectively. Multi-layer graph convolutional neural network (GCN) and gating mechanism are introduced to perceive syntactic path and expert path information respectively, capture the effective dependency relationship between aspect words and their semantic context, and generate graph embedding representations for different semantic levels.

[0056] C. Two-layer contrastive learning stage: The objective function is to jointly optimize instance-level and cluster-level contrastive learning, respectively constraining the consistency of single-sample augmented representations and the compactness of intra-class semantic centers. Positive and negative sample pairs are generated by introducing sentence augmentation and graph structure perturbation, and the contrastive loss function guides the model to learn a discriminative and robust semantic representation space.

[0057] D. Classification and Joint Optimization Stage: The sentence representation output by the graph neural network is fed into a fully connected layer and supervised training is performed using label smoothing cross-entropy. Finally, the loss function is a weighted combination of classification loss and contrastive loss to form a unified optimization objective, thereby achieving the fusion modeling of structural knowledge and semantic knowledge.

[0058] In the graph construction phase, this embodiment constructs a word co-occurrence graph G. (w) Part-of-speech co-occurrence graph G (p) and entity graph G (e) These heterogeneous graph structures extract relational information from different dimensions of the original text:

[0059] G is constructed based on the PMI co-occurrence statistics extracted using a sliding window. (w) :

[0060]

[0061] The part-of-speech graph and entity graph are generated by calculating sparse adjacency matrices using part-of-speech tag sequences and entity similarity.

[0062] Obtain the adjacency matrix for each type of graph structure. Node feature matrix

[0063] For each view v∈{w, p, e}, the corresponding graph structure is input into a graph convolutional encoder (GCN) to extract its structural semantic features:

[0064]

[0065] in, To add a self-loop adjacency matrix, W is the degree matrix. (v) These are learnable weights.

[0066] To dynamically adjust the importance of different views, an expert gating mechanism (Mixture of Experts) is introduced, embedding the three views into the input gating fusion module:

[0067]

[0068] The final fusion is represented as:

[0069]

[0070] To further enhance the model's discriminative ability and robustness, this method introduces a two-stage contrastive learning approach:

[0071] 1. Use instance-level contrastive learning (ICL) to extract fine-grained contrastive information:

[0072]

[0073] Where z i and z j Let represent the output embedding of the i-th text and its enhanced text, respectively, and τ represent the temperature parameter.

[0074] 2. Use Cluster-Level Contrastive Learning (CCL) to explore their similarity from a clustering perspective and leverage their high-level feature clustering information to group similar instances together:

[0075]

[0076] The fusion representation H fused Connect to an MLP classifier to output text category probabilities:

[0077] y pred =Softmax(W c ·H fused +b c );

[0078] Cross-entropy is used as the final classification loss:

[0079]

[0080] The final model can automatically determine the semantic category of different text contents and select the most suitable one from multiple predefined categories, such as... Figure 2 As shown.

[0081] The information of the benchmark model used in the test data of this invention is as follows:

[0082] TF-IDF+SVM: A traditional text classification method that represents text as TF-IDF vectors and uses a Support Vector Machine (SVM) for classification. This method is simple in structure and computationally efficient, but it has significant shortcomings in capturing deep semantic structures, especially in short text classification tasks.

[0083] PTE (Predictive Text Embedding) is a method that constructs three local co-occurrence graphs between words and documents, learning text representations through matrix factorization. This method integrates unsupervised corpus information and supervised label information, improving the embedding's expressive power, but it is still limited by the static graph structure and lacks contextual modeling capabilities.

[0084] CNN: A text classification method based on convolutional neural networks. It extracts n-gram-level text features through local convolutional operations, making it suitable for capturing short-range dependency information. It has a simple structure and high training efficiency, but it has limitations in handling long-range dependencies and global semantic modeling. LSTM: Long Short-Term Memory network is a variant of recurrent neural network. It excels at handling temporal information and can model word order relationships and long-term dependencies, but it is insufficient in modeling graph structure information in text.

[0085] BERT: A pre-trained language model with a bidirectional Transformer structure, including two methods: BERT-avg (averaging all word vectors) and BERT-cls (extracting sentence vectors at the [CLS] position), widely used in various text classification tasks. While possessing powerful language modeling capabilities, it still suffers from insufficient perceptual ability in structural modeling.

[0086] TextGCN is a text classification method based on Graph Convolutional Networks (GCNs). It models word co-occurrence relationships in text by constructing a heterogeneous graph from words and documents. This model transforms text classification into a graph node classification task, performing multi-layer graph convolutional feature propagation between nodes to capture global co-occurrence information among words. However, TextGCN does not consider semantic weights or node heterogeneity during graph construction, making it susceptible to interference from low-frequency noise words and lacking the ability to model semantic diversity.

[0087] TLGNN: Combines word order and dependency structure information to construct a text graph, and introduces a label-guided graph neural network mechanism to guide the model to focus on semantic regions related to the labels.

[0088] HyperGAT introduces higher-order hyperedge connections into graph attention networks to construct a hypergraph structure, thereby enhancing semantic modeling capabilities. This method captures complex interactions between multiple words and demonstrates strong performance in text classification.

[0089] DADGNN: Proposes a dynamic attention graph neural network that combines multi-layer graph building blocks with attention mechanisms to dynamically learn contextual semantic relationships and effectively model complex syntactic structures.

[0090] TextING models text as a dependency graph corresponding to word sequences and uses a graph convolutional network to encode word nodes. Its main advantage lies in introducing syntactic dependency structures to improve the model's ability to understand inter-word dependencies. However, this model still uses a single graph structure for modeling and does not dynamically adjust or enhance the structural features in the graph, making it difficult to fully express the interactive information between heterogeneous semantics.

[0091] STCKA, HGAT, STGCN, and NCHGAT are deep learning models proposed in recent years for short text features. STCKA combines external knowledge to enhance contextual modeling; HGAT integrates heterogeneous graph structures to model the interaction between words and tags; STGCN integrates syntactic graphs and sentence structure information; and NCHGAT utilizes nested structures to perceive multi-granular features of context.

[0092] This invention effectively improves the overall performance of short text classification models in terms of semantic representation, structural modeling, and cross-sample discrimination. This method is particularly suitable for classification tasks with limited context, severe semantic ambiguity, or sparse domain labels, and demonstrates significantly better performance than traditional methods on various publicly available short text datasets (such as MR, Snippets, Twitter, etc.).

[0093] This embodiment uses ACC and F1 scoring metrics to systematically evaluate the model's performance on five publicly available benchmark datasets (including Snippets, Ohsumed, TagMyNews, MR, and Twitter). Experimental results show that the model proposed in this invention outperforms traditional machine learning models (such as TF-IDF+SVM, PTE), deep learning models (such as CNN, LSTM, BERT), various graph neural network models (such as TextGCN, TextING, HyperGAT, TLGNN, DADGNN), and short text modeling methods (such as STCKA, STGCN, HGAT, etc.) on most datasets, demonstrating stronger generalization ability and robustness.

[0094] Example 1

[0095] like Figure 3 As shown, the overall model structure of this invention consists of the following key modules: a text input processing module, a multi-source graph structure construction module, an expert fusion module, a graph neural network encoding module, and a two-layer contrastive learning optimization module.

[0096] First, after initializing and preprocessing the text input using word vectors, multiple semantic heterogeneous graph structures are constructed using various perspectives (such as word co-occurrence, dependency syntax, TF-IDF, etc.). Each perspective captures different structural relationships to comprehensively model word interactions and semantic dependencies in the text.

[0097] Subsequently, the model incorporates graph experts and uses an expert-gated fusion module (MoE) to dynamically fuse graph representations. This module receives multi-path graph representations as input and automatically selects the most suitable structural representation for subsequent modeling based on a gating weight mechanism, fully leveraging the complementary characteristics of different semantic graphs.

[0098] Next, the fused graph structure is fed into a graph neural network (such as GCN) for feature propagation and node representation learning, forming a structure-aware embedding representation of short text instances.

[0099] Finally, by designing a dual-layer contrastive learning objective, the model performs two types of contrastive optimization tasks during the training phase: intra-class alignment and inter-class separation, to improve semantic consistency in the embedding space while enhancing the clarity of classification boundaries. This mechanism can alleviate feature ambiguity caused by contextual sparsity in short texts and improve the model's generalization ability.

[0100] Example 2

[0101] Figure 4 This paper presents a schematic diagram of a "multi-expert sparse gating fusion network structure" proposed in this invention. This structure is an expert collaboration mechanism built on a multi-graph perspective, aiming to achieve effective fusion and dynamic control of multiple types of heterogeneous graph structures in the text representation learning process, thereby enhancing the model's expressive and generalization capabilities in complex structural scenarios.

[0102] In this structure, the input sample x i After preprocessing, the input is fed into a gating network, which outputs multiple expert fusion control vectors based on the semantic distribution of the input, graph structure context, or predefined rules, including word co-occurrence graph control vectors. Syntactic graph control vector and entity semantic graph control vector These control signals are used for dynamic scheduling and feature fusion among subsequent expert graph modules.

[0103] Specifically, expert modules E1 to E r Corresponding word co-occurrence graph represents the sub-network; Module E r+1 ~E h Representation of syntactic dependency graph subnetworks; Module E h+1 ~E n A network of expert networks that corresponds to entity or named entity structures. Each submodule accepts its own constructed graph structure G(x) and text input x. i Structure-sensitive representation vectors are extracted based on graph neural networks (such as GCN or GAT).

[0104] Subsequently, this invention designs a group-weighted fusion strategy:

[0105] Group 1: Word co-occurrence graph experts E1-E r The output is weighted Weighted summation is used to form a local word adjacency representation E(x). w ;

[0106] Group 2: Dependency Graph Expert E r+1 ~E h The output is based on Gated convergence is used to form a syntactic dependent expression E(x). p ;

[0107] Group 3: Entity Expert Module E h+1 ~E n The output is through The entities are merged to form a solid graph representation E(x). e .

[0108] In the final fusion stage, the system can use weighted linear combination, splicing or attention mechanism to unify and integrate the expert graph representations in the above three directions to generate the final text structure representation E(x) with global semantic awareness, which serves as the input for downstream classification, clustering or contrastive learning modules.

[0109] As shown in Table 1, this invention underwent systematic experimental evaluation on five standard short text classification datasets: Twitter, MR, Snippets, Ohsumed, and TagMyNews. These datasets cover a variety of real-world short text scenarios, including social media texts, movie reviews, search engine summaries, medical literature, and news headlines. They are highly representative and feature diverse data structures, effectively validating the effectiveness and generalization ability of the method in various short text tasks. Table 1 lists the basic statistical information for each dataset.

[0110] Table 1

[0111] Dataset #docs #tokens #entities #classes Snippets 12340 14.5 4.4(94％) 8 Ohsumed 7400 6.8 3.1(96％) 23 TagMyNews 32549 5.1 1.9(86％) 7 MR 10662 7.6 1.8(76％) 2 Twitter 10000 3.5 1.1(63％) 2

[0112] Tables 2 and 3 show the comparison results of the classification accuracy and macro-average F1 score of the method of the present invention on various datasets. This embodiment compares the method with four mainstream benchmark models, including: traditional models: TF-IDF+SVM and PTE; deep learning models: CNN, LSTM, BERT-avg and BERT-cls; graph neural network models: TLGNN, HyperGAT, TextING, DADGNN, and TextGCN; and deep short text models: STCKA, HGAT, SIGCN and NC-HGAT.

[0113] Table 2

[0114]

[0115] Table 3

[0116]

[0117] This method aims to address the challenges of semantic sparsity, insufficient structural information, and difficulties in heterogeneous graph fusion in short text classification. First, it constructs expert graph structures from multiple semantic perspectives for the original text, including word co-occurrence graphs, dependency syntax graphs, and semantic similarity graphs, to mine structural information at different levels. Then, an expert-gated fusion module is designed to adaptively learn the combined weights of each expert graph based on the input text content, generating a semantically enhanced fusion graph structure. In the feature modeling stage, the model uses a graph neural encoder to extract node structural features and combines this with a text encoder to obtain contextual semantic representations. Furthermore, a two-layer contrastive learning mechanism is introduced: structural contrast is used to capture the similarities and differences between multiple graph structures, and semantic contrast is used to optimize sentence-label alignment, thereby improving the model's discriminative and generalization abilities.

[0118] This invention conducted systematic experiments on five publicly available short text classification benchmark datasets (Snippets, Ohsumed, TagMyNews, MR, and Twitter), covering typical task scenarios such as news, multi-domain comments, and social media text. The comparison targets included various mainstream baselines such as traditional machine learning models (TF-IDF+SVM, PTE), deep text models (CNN, LSTM, BERT), graph neural network models (TextGCN, TextING, DADGNN, HyperGAT, TLGNN), and deep short text modeling methods (STCKA, HGAT, STGCN, SHINE, NCHGAT, GIFT). Experimental results show that the model proposed in this invention comprehensively outperforms existing methods on the Macro-F1 metric, especially exhibiting stronger robustness and stability in scenarios with blurred inter-class boundaries, weak contextual dependencies, or imbalanced label counts. The model has a flexible and adaptable structure, can be extended to integrate various pre-trained language models, and is suitable for various application scenarios such as social text analysis, news clustering, and financial text recognition, possessing significant engineering application value and theoretical generalization significance.

[0119] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A text classification method based on multi-expert fusion gated graph neural network contrastive learning, characterized in that, include: The original text is preprocessed to generate a standardized corpus and extract word, part-of-speech, and entity features; A first heterogeneous graph is constructed based on the word features, a second heterogeneous graph is constructed based on the part-of-speech features, and a third heterogeneous graph is constructed based on entity similarity. The first heterogeneous graph, the second heterogeneous graph, and the third heterogeneous graph are dynamically weighted and fused using an expert-gated fusion module to generate a fused graph structure. Graph convolution encoding is performed on the fused graph structure to generate node representation vectors; Based on the node representation vector, perform two-layer contrastive learning: implement instance-level contrastive learning to optimize the consistency of single-sample representation, and simultaneously implement cluster-level contrastive learning to optimize intra-class center aggregation; The representation vector optimized through contrastive learning is input into the classifier, which outputs text category labels.

2. The text classification method based on multi-expert fusion gated graph neural network contrastive learning as described in claim 1, characterized in that, The preprocessing process includes: The original corpus was uniformly encoded and illegal characters were removed. Use a part-of-speech tagging tool to generate a sequence of part-of-speech tags; Identify named entities using entity linking tools; Pre-trained word embeddings are used to initialize word features, entity embeddings are generated through training triple vectors, and part-of-speech features are encoded using one-hot encoding. Construct a text-word TF-IDF pooling matrix and fill in placeholders for empty terms in the part-of-speech and entity channels.

3. The text classification method based on multi-expert fusion gated graph neural network contrastive learning as described in claim 1, characterized in that, The construction process of the first heterogeneous graph includes: Calculate the co-occurrence probability of word pairs using a sliding window; Valid word-edge relationships are filtered based on point mutual information thresholds; Store the filtering results as a sparse adjacency matrix.

4. The text classification method based on multi-expert fusion gated graph neural network contrastive learning as described in claim 1, characterized in that, The construction process of the second heterogeneous graph includes: Generate part-of-speech co-occurrence relationships based on part-of-speech tag sequences; POS-POS connection edges are constructed by calculating point mutual information; Store the connection relationships as a sparse matrix.

5. The text classification method based on multi-expert fusion gated graph neural network contrastive learning as described in claim 1, characterized in that, The construction process of the third heterogeneous graph includes: Calculate the cosine similarity between entities; Entity pairs with a similarity greater than 0.2 are retained for connection; Limit each entity to a maximum of 100 similar neighbor nodes.

6. The text classification method based on multi-expert fusion gated graph neural network contrastive learning as described in claim 1, characterized in that, The operation of the expert gating fusion module includes: Receive the node representation vectors of the first heterogeneous graph, the second heterogeneous graph, and the third heterogeneous graph; Three sets of dynamic weight coefficients are generated using a gating network; Weighted linear fusion is performed on the three sets of node representations according to the aforementioned weighting coefficients.

7. The text classification method based on multi-expert fusion gated graph neural network contrastive learning as described in claim 1, characterized in that, The two-layer contrastive learning process includes: Instance-level contrastive learning: Generate positive sample pairs through text augmentation to narrow the representation distance of similar samples; Cluster-level comparative learning: Aggregate the sample center representation within a class to increase the distance between different clusters.

8. The text classification method based on multi-expert fusion gated graph neural network contrastive learning as described in claim 1, characterized in that, The classification process includes: Input the fused graph representation vector into the multilayer perceptron; The class probability distribution is calculated using the Softmax function; Supervised training is performed using label-smooth cross-entropy loss.

Images