UWB at WASSA-2024 Shared Task 2: Cross-lingual Emotion Detection

2025-08-12

AI Summary

This work addresses cross-lingual sentiment detection, performing six-way sentiment classification and simultaneous trigger word identification (via binary detection and numerical span localization) on tweets in five languages. Methodologically, we propose a joint framework integrating quantized large language models (LLMs) with multilingual Transformers: leveraging multilingual foundation models—including Orca-2, XLM-R, and mT5—we apply LoRA for parameter-efficient fine-tuning and 4-bit quantization for model compression; further, we incorporate machine translation to enhance cross-lingual transfer and design a trigger-word switching strategy to improve localization accuracy. Evaluated on SemEval-2024 Task 10, our approach ranks first in numerical trigger span detection, third in binary trigger detection, and seventh in sentiment classification—achieving state-of-the-art overall performance. Our core contributions are a lightweight, efficient paradigm for cross-lingual LLM adaptation and a structured optimization strategy explicitly designed for trigger word identification.

Technology Category

• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Constraint Satisfaction and Optimization: Satisfiability Modulo Theories

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

This paper presents our system built for the WASSA-2024 Cross-lingual Emotion Detection Shared Task. The task consists of two subtasks: first, to assess an emotion label from six possible classes for a given tweet in one of five languages, and second, to predict words triggering the detected emotions in binary and numerical formats. Our proposed approach revolves around fine-tuning quantized large language models, specifically Orca~2, with low-rank adapters (LoRA) and multilingual Transformer-based models, such as XLM-R and mT5. We enhance performance through machine translation for both subtasks and trigger word switching for the second subtask. The system achieves excellent performance, ranking 1st in numerical trigger words detection, 3rd in binary trigger words detection, and 7th in emotion detection.

Problem

• Detect emotion labels in multilingual tweets
• Identify trigger words for detected emotions
• Improve performance using quantized LLMs and translation

Innovation

• Fine-tuning quantized Orca~2 with LoRA
• Using multilingual XLM-R and mT5 models
• Enhancing performance via machine translation