AylinNaebzadeh/Text-Based-Emotion-Detection-SemEval-2025

Official implementation for our paper: "GinGer at SemEval-2025 Task 11: Leveraging Fine-Tuned Transformer Models and LoRA for Sentiment Analysis in Low-Resource Languages" , Proceedings of SemEval-2025, ACL 2025 # Emotion Detection in Low-Resource Languages with Fine-Tuned Transformers and LoRA ## Overview This repository contains the code for our participation in **SemEval-2025 Task 11: Bridging the Gap in Text-Based Emotion Detection**, covering: - **Track A** — Multi-label emotion detection (joy, sadness, fear, anger, surprise, disgust) - **Track B** — Emotion intensity prediction (ordinal scale 0–3) We achieved **9th place** in the Arabic Algerian track (ARQ) among 40 competing teams. Our approach combines **transformer-based models** with **Parameter-Efficient Fine-Tuning (PEFT)** via **LoRA (Low-Rank Adaptation)** to handle the challenges of low-resource language settings. --- ## Languages | Track | Languages | |-------|-----------| | Track A | Afrikaans (AFR), Arabic Algerian (ARQ), Hindi (HIN), Swedish (SWE) | | Track B | Russian (RUS), Romanian (RON) | --- ## Models Used **Track A** - XLM-RoBERTa-Base - T-XLM-RoBERTa - BERT-Multilingual - DiziBERT-Sent. (Arabic Algerian) - BERT-Base-Swedish-Cased-Sent. **Track B** - XLM-RoBERTa - T-XLM-RoBERTa - BERT-Multilingual --- ## Results ### Track A (Micro F1) | Language | Best Model | Micro F1 | |----------|-----------|----------| | Afrikaans | T-XLM-RoBERTa | 0.54 | | Hindi | XLM-RoBERTa-Base | 0.84 | | Arabic Algerian | DiziBERT-Sent. | 0.58 | | Swedish | BERT-Base-Swedish-Cased-Sent. | 0.72 | ### Track B (Pearson Correlation) | Language | Best Model | Pearson Corr. | |----------|-----------|---------------| | Russian | XLM-RoBERTa | 0.83 | | Romanian | XLM-RoBERTa | 0.57 | --- ## Methodology ### Track A - Fine-tuned pretrained transformer models for multi-label classification - Sigmoid-based output with threshold tuning to maximize F1 score - Initial threshold: 0.3, then tuned on the dev set ### Track B - LoRA applied to transformer layers for parameter-efficient adaptation - **Loss function:** Mean Squared Error (MSE) - **Post-processing:** Floor clipping to enforce predictions in [0, 3] - **Eval …