TIAO-Eliasse/DeepSemantics-SemEval

fine-grained multi-label classification of polarization manifestations (Vilification, Dehumanization, Extreme Language, Lack of Empathy, Invalidation, Stereotype) in English and Hausa social media texts. # DeepSemantics at SemEval-2026 Task 9: Label-Wise Optimization with Adaptive Focal Loss for Polarization Manifestation Identification > **Team:** DeepSemantics — African Institute for Mathematical Sciences (AIMS), South Africa > **Competition:** SemEval-2026 Task 9 — Detecting Multilingual Online Polarization (POLAR) > **Subtask:** Subtask 3 — Manifestation Identification (English & Hausa) --- ## 📄 Abstract This repository contains the code for our system submitted to **SemEval-2026 Task 9, Subtask 3: Polarization Manifestation Identification**. We address fine-grained multi-label classification of polarization manifestations (Vilification, Dehumanization, Extreme Language, Lack of Empathy, Invalidation, Stereotype) in English and Hausa social media texts. Our approach combines: - **Transformer encoders** (RoBERTa-base for English, Afro-XLM-R-small for Hausa) - **One-vs-Rest (OvR) framework** for label-wise modeling - **Adaptive Focal Loss** (English) and **Weighted Binary Cross-Entropy** (Hausa) - **Controlled oversampling** with Easy Data Augmentation (EDA) - **Label-wise threshold optimization** via stratified K-fold validation **Official test results:** | Language | Macro-F1 | Leaderboard Rank | |----------|----------|-----------------| | English | 0.464 | 14th | | Hausa | 0.192 | 5th | --- ## 📂 Repository Structure ``` ├── substack_3_final.ipynb # Full pipeline: preprocessing, training, evaluation, error analysis └── README.md ``` The notebook contains the entire pipeline end-to-end: - Data loading and exploratory analysis - Controlled oversampling + Easy Data Augmentation (EDA) - One-vs-Rest training with Adaptive Focal Loss (EN) / Weighted BCE (HA) - Label-wise threshold optimization via stratified K-fold - Evaluation and error analysis --- ## 📊 Data Data comes from the **POLAR SemEval-2026 shared task**. Download the dataset from the official repository: 🔗 **Polar-SemEval/data-public** Once downloaded, updat …