Research map / Vision-language and multimodal / Human and embodied modeling
Mitigating Idiom Inconsistency: A Multi-Semantic Contrastive Learning Method for Chinese Idiom Reading Comprehension
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 MSCLM, a BERT-based Chinese idiom cloze model with metaphor contrastive learning (prompt-based definition positives) and a multi-semantic cross-attention module, and introduces the CIDT definition-matching test set.
Problem
Chinese idiom reading comprehension (cloze) is hard because idioms' metaphorical meanings diverge from literal ones (metaphorical inconsistency) and the same idiom takes different meanings across contexts (contextual inconsistency).
Approach
Uses a prompt-based definition template as a metaphorical static representation to form positive pairs with the contextual [MASK] representation under SimCSE-style contrastive loss, plus a convolution-and-cross-attention module that models relations among candidate idioms conditioned on context.
Key results
- MSCLM (full) reaches 96.8% test accuracy on ChID and 98.8% on the Sim test set, versus 95.7% for PRIEM on Test.
- On CCT, MSCLM achieves 96.4% accuracy, versus 92.0% for BTSM.
- GPT-3.5 scores only 39.8% on ChID test, far below BERT-based models, though it reaches 84.8% on CIDT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the multi-semantic cross-attention module to MSCLM-C (ablation) | MSCLM · ChID Test | accuracy: 81.8 -> 96.8 | Helped |
| Add the metaphor contrastive learning module to MSCLM-M (ablation) | MSCLM · ChID Test | accuracy: 86.8 -> 96.8 | Helped |
| Add the multi-semantic cross-attention module to MSCLM-C (ablation) | MSCLM · ChID Sim | accuracy: 75.3 -> 98.8 | Helped |
| Add the metaphor contrastive learning module to MSCLM-M (ablation) | MSCLM · ChID Sim | accuracy: 91.4 -> 98.8 | Helped |
| Use MSCLM instead of BTSM | MSCLM · ChID-Competition Test | accuracy: 92.0 -> 96.5 | Helped |
| Use MSCLM instead of BTSM | MSCLM · CCT Test | accuracy: 93.7 -> 96.4 | Helped |
| Use MSCLM instead of GPT-3.5 | MSCLM · CIDT | accuracy: 84.8 -> 98.8 | Helped |
Surprising findings
- GPT-3.5 performs far worse than fine-tuned BERT-based models on ChID and CCT, despite being a large LLM.
Limitations
- Evaluated only on Chinese idiom cloze datasets; generalization to other metaphorical language is untested.
- Plans to combine with LLMs are future work, so LLM integration is not explored.
Experimental setup
Datasets: ChID CCT
Benchmarks: ChID ChID-Competition CCT CIDT
Metrics: accuracy
Baselines: LM AR SAR BERT-WWM SKER BTSM CM BERT-IDM PRIEM GPT-3.5
Builds on: BERT SimCSE PRIEM BERT-WWM
Closest papers by meaning
- McHirc: A Multimodal Benchmark for Chinese Idiom Reading Comprehension (AAAI 2025 · similarity 0.92)
- Translate Meanings, Not Just Words: IdiomKB’s Role in Optimizing Idiomatic Translation with Language Models (AAAI 2024 · similarity 0.83)
- MCIF: Multimodal Crosslingual Instruction-Following Benchmark from Scientific Talks (ICLR 2026 · similarity 0.82)
- Thought-Path Contrastive Learning via Premise-Oriented Data Augmentation for Logical Reading Comprehension (AAAI 2025 · similarity 0.82)
- InterpretARA: Enhancing Hybrid Automatic Readability Assessment with Linguistic Feature Interpreter and Contrastive Learning (AAAI 2024 · similarity 0.80)
- MMIE: Massive Multimodal Interleaved Comprehension Benchmark for Large Vision-Language Models (ICLR 2025 · similarity 0.80)
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.