Abstract
Emotion detection in natural language processing is a challenging task due tothe complexity of human emotions and linguistic diversity. While significantprogress has been made in high-resource languages, emotion detection inlow-resource languages remains underexplored. In this work, we addressmultilingual and cross-lingual emotion detection by leveraging adapter-basedfine-tuning with multilingual pre-trained language models. Adapters introduce asmall number of trainable parameters while keeping the pre-trained modelweights fixed, offering a parameter-efficient approach to adaptation. Weexperiment with different adapter tuning strategies, including task-onlyadapters, target-language-ready task adapters, and language-family-basedadapters. Our results show that target-language-ready task adapters achieve thebest overall performance, particularly for low-resource African languages withour team ranking 7th for Tigrinya, and 8th for Kinyarwanda in Track A. In TrackC, our system ranked 3rd for Amharic, and 4th for Oromo, Tigrinya, Kinyarwanda,Hausa, and Igbo. Our approach outperforms large language models in 11 languagesand matches their performance in four others, despite our models havingsignificantly fewer parameters. Furthermore, we find that adapter-based modelsretain cross-linguistic transfer capabilities while requiring fewercomputational resources compared to full fine-tuning for each language.