Research map / Label-efficient and robust learning / Domain adaptation and generalization
Uncovering and Mitigating the Hidden Chasm: A Study on the Text-Text Domain Gap in Euphemism Identification
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.
Identifies a text-text domain gap between self-supervised training data and test data in euphemism identification and proposes FA-Net (in-domain contrastive, hard cross-domain contrastive, and entropy minimization) that boosts BERT and beats LLMs on Drug, Weapon, and Sexuality datasets.
Problem
Euphemism identification maps a euphemism to its target keyword; self-supervised training data built by masking target keywords differs in distribution from test data with masked euphemisms, causing a text-text domain gap and poor test accuracy.
Approach
Explains the text-text domain gap via data distribution and the cone effect (borrowed from the modality gap literature), then bridges it with FA-Net combining instance-wise in-domain contrastive learning, hard cross-domain contrastive alignment of nearest source-target positives, and entropy minimization on unlabeled target data, with no additional parameters.
Key results
- Bert+FA reaches Acc@1 of 0.35 on Drug, 0.47 on Weapon, and 0.50 on Sexuality, versus 0.20/0.33/0.32 for SelfEDI and 0.24/0.38/0.38 for Bert.
- Bert+FA beats GPT-3.5 (0.33/0.17/0.42 Acc@1) and other LLMs (StableLM, mPLUG-Owl, Llama 2) on top-1 accuracy while costing about 0.00003$ per sentence versus about 0.00035$ for GPT-3.5.
- Ablation shows IC+HC combined gives the largest gain among component pairs, and the full IC+HC+EM model is best on all three datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FA-Net to Bert | Bert · Drug | Acc@1: 0.24 -> 0.35 | Helped |
| Add FA-Net to Bert | Bert · Weapon | Acc@1: 0.38 -> 0.47 | Helped |
| Add FA-Net to Bert | Bert · Sexuality | Acc@1: 0.38 -> 0.50 | Helped |
| Add EM-net (ablation) | Bert · Drug | Acc@1: 0.24 -> 0.27 | Helped |
| Add EM-net (ablation) | Bert · Weapon | Acc@1: 0.38 -> 0.38 | No clear effect |
| Add EM-net (ablation) | Bert · Sexuality | Acc@1: 0.38 -> 0.40 | Helped |
| Add IC-net (ablation) | Bert · Drug | Acc@1: 0.24 -> 0.27 | Helped |
| Add IC-net (ablation) | Bert · Weapon | Acc@1: 0.38 -> 0.39 | Helped |
| Add IC-net (ablation) | Bert · Sexuality | Acc@1: 0.38 -> 0.41 | Helped |
| Add HC-net (ablation) | Bert · Drug | Acc@1: 0.24 -> 0.29 | Helped |
| Add HC-net (ablation) | Bert · Weapon | Acc@1: 0.38 -> 0.42 | Helped |
| Add HC-net (ablation) | Bert · Sexuality | Acc@1: 0.38 -> 0.43 | Helped |
Surprising findings
- Average cosine similarity between embeddings of training and test text is very high (0.966 and 0.965 for BERT), indicating a narrow cone, which the authors use as evidence of the text-text domain gap.
- Entropy minimization performs poorly when applied from the first epoch and needs a later start epoch to be beneficial.
Limitations
- Ground truth euphemism lists are incomplete and error-prone due to rapidly evolving language, as the authors note.
- Hard cross-domain positives depend on a similarity threshold gamma that needs tuning (best 0.6), and entropy minimization needs a tuned start epoch.
Experimental setup
Datasets: Drug Weapon Sexuality
Benchmarks: Drug Weapon Sexuality
Metrics: Acc@1 Acc@2 Acc@3 A-distance cost per sentence
Baselines: Word2Vec SelfEDI LSTM CNN Bert StableLM mPLUG Llama2 GPT-3.5
Builds on: SelfEDI Liang et al. 2022 modality gap instance-wise contrastive learning entropy minimization
Closest papers by meaning
- A Multi-Task Benchmark for Abusive Language Detection in Low-Resource Settings (NeurIPS 2025 · similarity 0.84)
- ToxicTextCLIP: Text-Based Poisoning and Backdoor Attacks on CLIP Pre-training (NeurIPS 2025 · similarity 0.81)
- Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text (NeurIPS 2025 · similarity 0.81)
- TRACE: Textual Relevance Augmentation and Contextual Encoding for Multimodal Hate Detection (AAAI 2026 · similarity 0.81)
- Hateful Meme Detection through Context-Sensitive Prompting and Fine-Grained Labeling (Student Abstract) (AAAI 2025 · similarity 0.81)
- Semantic Representation Attack against Aligned Large Language Models (NeurIPS 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.