Research map / Label-efficient and robust learning / Recognition and long-tail

Boosting Short Text Classification with Multi-Source Information Exploration and Dual-Level Contrastive Learning

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Multilingual language representation pretraining · 3 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 MI-DELIGHT, a graph-based short text classifier that fuses statistical, linguistic, and factual graphs with a hierarchical instance-level then cluster-level contrastive auxiliary tasks, trained on few labeled examples.

Problem

Short text classification suffers from semantic sparsity and very few labeled samples; prior auxiliary-task and contrastive methods run tasks in parallel and use only instance-level contrast.

Approach

Builds word, POS, and entity (NELL/TAGME) graphs for multi-source information, then chains instance-level contrastive learning, cluster-level contrastive learning with pseudo-cluster labels from nearest-neighbor connected components, and classification in a hierarchical rather than parallel structure.

Key results

• MI-DELIGHT reaches 75.11% ACC on Twitter, 66.49% on MR, 87.90% on Snippets, 48.56% on Ohsumed, and 69.72% on TagMyNews, exceeding the best non-LLM baseline (GIFT/SHINE) on all five datasets.
• On Snippets, Ohsumed and TagMyNews it outperforms the fine-tuned LLMs reported (e.g., Llama3-8B 61.68% ACC on Snippets, Llama2-7B 67.31% on TagMyNews), while Llama models score higher on Twitter and MR.
• Ablations show removing the word graph causes the largest drop (Twitter ACC 75.11 to 62.60), and the hierarchical design beats the parallel variant on most datasets.

Measured effects reported in the paper

Surprising findings

• Fine-tuned LLMs such as Llama3-8B (89.50% ACC on Twitter) and Llama2-7B outperform MI-DELIGHT on Twitter and MR, which the authors attribute partly to possible test-data exposure during pretraining.
• The 'parallel' variant beats the proposed hierarchical model on Ohsumed in F1 and on TagMyNews in some metrics, and the context-based augmentation achieves the best Twitter result (75.66% ACC), so the default WordNet setting is not uniformly best.

Limitations

• LLM baselines were only lightly fine-tuned with GPU reduction techniques due to compute constraints, so the LLM comparison is limited.
• Performance is weaker on Twitter and MR relative to fine-tuned LLMs, and Ohsumed absolute scores stay low (around 48% ACC).
• Reliance on TAGME entity linking to NELL and WordNet synonym augmentation may limit domain transfer; the paper does not analyze this.

Experimental setup

Datasets: Twitter MR Snippets Ohsumed TagMyNews NELL

Benchmarks: Twitter MR Snippets Ohsumed TagMyNews

Metrics: accuracy macro-F1

Baselines: TF-IDF+SVM PTE CNN LSTM BERT TextGCN HGAT SHINE GIFT GPT-3.5 Llama2-7B Llama3-8B

Builds on: GCN TextGCN SHINE GIFT TransE

Closest papers by meaning

• Improved Graph Contrastive Learning for Short Text Classification (AAAI 2024 · similarity 0.92)
• A Simple Graph Contrastive Learning Framework for Short Text Classification (AAAI 2025 · similarity 0.91)
• Contrastive Learning with Simplicial Convolutional Networks for Short-Text Classification (ICML 2025 · similarity 0.89)
• Multiscale Positive-Unlabeled Detection of AI-Generated Texts (ICLR 2024 · similarity 0.83)
• TGLsta: Low-resource Textual Graph Learning with Semantic and Topological Awareness via LLMs (AAAI 2025 · similarity 0.83)
• DeTeCtive: Detecting AI-generated Text via Multi-Level Contrastive Learning (NeurIPS 2024 · similarity 0.83)