Wavelet-Enhanced Euclidean-Hyperbolic Graph Convolutional Networks for Text Classification

ACL ARR 2025 May Submission977 Authors

Abstract: Graph convolutional networks (GCNs) have been successfully applied to text classification tasks. However, existing GCN-based methods fail to fully utilize the representational advantages of tree-like structures in hyperbolic space and struggle to capture hierarchical hypernym-hyponym relationships between words. Additionally, text graph construction heavily relies on structural information from fixed corpora. To address these limitations, this study proposes a wavelet-enhanced Euclidean-hyperbolic graph convolutional network (EHGCN) for text classification. The method establishes complementary semantic enhancement across multiple dimensions through frequency-domain analysis and Euclidean-hyperbolic cross-space topology restructuring. The frequency-domain perspective captures text fine-grained features via multi-scale semantic decoupling of word vectors, while the Euclidean-hyperbolic semantic topology constructs cross-space text structures and integrates heterogeneous features from cross-space graph convolution to achieve text representations combining local semantics with hierarchical dependencies. Experiments on five benchmark datasets (R8, R52, MR, Ohsumed, TREC) show that EHGCN achieves a 1.89\% average accuracy improvement over mainstream methods. Compared to task-specific LLM-based models (CAPR, COT), EHGCN demonstrates a 7.21\% average performance gain.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: text classification; graph neural network; wavelet; hyperbolic space

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 977