FLAICOL/FLAICOL

Hate speech detection in low-resource, code- mixed languages is a challenging task as people often switch between scripts and languages in a single post. Code-Mixed scripts can take the form of explicit slurs, subtle insults, or frag- mented abuse, and is often hidden by spelling variants and Romanized script. # FLAICOL: Flip-Point-Led Augmentation for Imbalanced Code-Mixed Offensive Language Detection This repository contains the implementation of **FLAICOL**, a flip-point-led augmentation framework for low-resource, code-mixed offensive language detection, as described in the accompanying anonymous ACL submission. FLAICOL addresses severe class imbalance in code-mixed hate speech datasets by generating minimal boundary-crossing examples in embedding space and using them to augment the minority class. --- ## 1. Overview Hate speech detection in code-mixed Indian languages (Tamil–English, Malayalam–English, Kannada–English) presents three key challenges: * Script mixing (native + Romanized) * Orthographic variation * Severe minority-class imbalance FLAICOL introduces a flip-point-based augmentation framework that: 1. Fine-tunes a Transformer classifier. 2. Identifies near-boundary examples. 3. Performs embedding-space search under a bias-homotopy continuation strategy. 4. Maps continuous embedding perturbations back to discrete tokens. 5. Validates flips under the original classifier. 6. Augments minority-class samples. 7. Retrains the classifier and evaluates performance gains. --- ## 2. Method Summary FLAICOL adapts the flip-point and homotopy framework introduced by Yousefzadeh & O’Leary to pretrained Transformer classifiers. For an input predicted as class *i*: * A target class *j* is selected (typically highest competing logit). * A minimal bias perturbation is solved via constrained optimization. * Bias is interpolated back toward original parameters. * At each step, embedding-space search identifies small semantic perturbations. * The resulting embedding is mapped to discrete tokens via nearest-neighbor lookup. * Only validated boundary-crossing flips are retained. The final augmented examples are appended to the training set. --- ## 3. Models Used Two encoder backbones were evaluated: * **XLM-RoBERTa-base** * **MuRIL-base** Both were fine …