Research map / Graph learning / Graph self-supervised learning

A Simple Graph Contrastive Learning Framework for Short Text Classification

AAAI 2025 · Graph learning › Graph self-supervised learning · approach: Contrastive learning with negatives and views · 6 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SimSTC, a graph contrastive learning framework for short text classification that builds word, POS, and entity component graphs and contrasts their resulting text embeddings directly, without data augmentation.

Problem

Short text classification suffers from semantic sparsity and label scarcity; existing graph+contrastive STC models rely on explicit data augmentation that corrupts semantics and only contrast two views.

Approach

Build three text-related component graphs (word, POS, entity) with GCNs; the per-view text embeddings form natural contrastive views, so multi-view contrastive learning is applied across all three views without any data augmentation, with a theoretical mutual-information lower bound justification.

Key results

• SimSTC reaches 73.51% ACC on Twitter, 66.52% on MR, 85.14% on Snippets and 84.07% on StackOverflow, best among non-LLM baselines (GIFT is second-best).
• On Snippets and StackOverflow SimSTC outperforms fine-tuned GPT-3.5, Bloom-7.1B, Llama2-7B and Llama3-8B; LLMs are better on Twitter and MR (e.g., Llama3-8B 87.65% on Twitter).
• Ablation shows each added contrastive view improves accuracy over the no-contrast baseline; using all three views gives the best results, and SimSTC trains about 1.5-4.7% of the parameters of the 7-8B LLMs.

Measured effects reported in the paper

Surprising findings

• Fine-tuned LLMs (Llama3-8B, Llama2-7B, Bloom-7.1B) substantially outperform the proposed model on Twitter and MR binary sentiment tasks.

Limitations

• LLM baselines are fine-tuned with LoRA and 4-bit quantization, and the GPT-3.5 comparison relies on a closed API, so the comparison may not be fully controlled.
• Mutual-information theorem is loosely stated (N described as number of datasets, though it is number of texts), and the bound depends on sufficient optimization of the contrastive loss.
• Method relies on external NELL knowledge base and POS/entity tagging pipelines, and does not clearly beat LLMs on binary sentiment datasets.

Experimental setup

Datasets: Twitter MR Snippets StackOverflow NELL knowledge base

Benchmarks: Twitter MR Snippets StackOverflow

Metrics: accuracy macro-F1

Baselines: CNN LSTM BERT RoBERTa TextING DADGNN TextGCN SHINE GIFT GPT-3.5 Bloom-7.1B Llama2-7B

Builds on: GCN TextGCN contrastive learning NELL knowledge base

Closest papers by meaning

• Improved Graph Contrastive Learning for Short Text Classification (AAAI 2024 · similarity 0.94)
• Contrastive Learning with Simplicial Convolutional Networks for Short-Text Classification (ICML 2025 · similarity 0.92)
• Boosting Short Text Classification with Multi-Source Information Exploration and Dual-Level Contrastive Learning (AAAI 2025 · similarity 0.91)
• ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.86)
• Perfect Alignment May be Poisonous to Graph Contrastive Learning (ICML 2024 · similarity 0.85)
• Weakly-Supervised Contrastive Learning for Imprecise Class Labels (ICML 2025 · similarity 0.85)