Research map / Vision-language and multimodal / Multimodal LLMs
M3UCD: A Multi-task Multimodal Metaphor Understanding Challenge Dataset for LLMs
Tianlong Zheng, Yating Yang, Rui Dong, Bo Ma, Lei Wang, Xi Zhou, Siru Miao, Turghun Osman
Read the paperCheck your idea against 137,217 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.
Introduces M3UCD, a multi-task multimodal metaphor understanding challenge dataset for evaluating LLMs, addressing superficial annotations, limited LLM evaluation, and fragmented task frameworks.
Problem
Multimodal metaphor understanding is important for machine cognition, but existing research has superficial dataset annotations, insufficient systematic evaluation of LLMs, and fragmented task frameworks.
Approach
Presents a large, fine-grained multi-task multimodal metaphor dataset with a unified task framework to systematically evaluate large language models on metaphor understanding.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with MCLF instead of single-task LoRA | MLLMs (8 models avg) · M3UCD test | Emotion Analysis macro-F1: +5.41 | Improved |
| train with MCLF instead of single-task LoRA | MLLMs (8 models avg) · M3UCD test | Intention Recognition macro-F1: +4.93 | Improved |
| train with MCLF instead of single-task LoRA | MLLMs (8 models avg) · M3UCD test | Offensiveness Detection macro-F1: +2.55 | Improved |
| train with MCLF instead of single-task LoRA | MLLMs (8 models avg) · M3UCD test | Sarcasm Detection macro-F1: +0.25 | No clear effect |
| train with MCLF instead of single-task LoRA | MLLMs (8 models avg) · M3UCD test | Metaphor Detection macro-F1: +0.65 | No clear effect |
| train with MCLF instead of single-task LoRA | MLLMs (8 models avg) · M3UCD test | Source Domain Detection macro-F1: +7.85 | Improved |
| train with MCLF instead of single-task LoRA | MLLMs (8 models avg) · M3UCD test | Target Domain Detection macro-F1: +6.92 | Improved |
| train with MCLF instead of single-task LoRA | MLLMs (8 models avg) · M3UCD test | Metaphor Classification macro-F1: +8.72 | Improved |
| train with MCLF instead of single-task LoRA | MLLMs (8 models avg) · M3UCD test | Tenor Extraction F1: +2.47 | Improved |
| train with MCLF instead of single-task LoRA | MLLMs (8 models avg) · M3UCD test | Vehicle Extraction F1: +2.22 | Improved |
| train with MCLF instead of single-task LoRA | MLLMs (8 models avg) · M3UCD test | Dual Extraction F1: +2.05 | Improved |
| add LoRA fine-tuning to zero-shot model | Qwen2-VL-7B-Instruct · M3UCD test | Dual Extraction F1: 63.57 -> 68.62 | Improved |
Experimental setup
Benchmarks: M3UCD
Closest papers by meaning
- CAMEL: Capturing Metaphorical Alignment with Context Disentangling for Multimodal Emotion Recognition (AAAI 2024 · similarity 0.87)
- LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.86)
- Can Large Language Models Help Multimodal Language Analysis? MMLA: A Comprehensive Benchmark (NeurIPS 2025 · similarity 0.86)
- MCIF: Multimodal Crosslingual Instruction-Following Benchmark from Scientific Talks (ICLR 2026 · similarity 0.85)
- MMIE: Massive Multimodal Interleaved Comprehension Benchmark for Large Vision-Language Models (ICLR 2025 · similarity 0.85)
- MME-Unify: A Comprehensive Benchmark for Unified Multimodal Understanding and Generation Models (ICLR 2026 · similarity 0.85)
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.
