ariful-TheJedi/bangla-wsd

# 📖 “Bengali Word Sense Disambiguation Using Supervised Approach (BERT)” WSD is the task of determining the correct meaning of a polysemous word based on its **context**. Bengali, spoken by over **250 million people**, is morphologically rich and highly context-dependent. This makes WSD a particularly challenging but essential problem for applications like **machine translation, chatbots, search engines, and information retrieval**. --- ## 🚀 Project Highlights - **Curated Dataset** - 122 polysemous Bengali words - 440 distinct senses - ~22,600 manually annotated sentences (JSON format) - **Models Used** - `csebuetnlp/BanglaBERT` - `sagorsarker/bangla-bert-base` - **Pipeline** - Preprocessing with normalization, tokenization, and special markers `[TGT] ... [/TGT]` to highlight target words - Stratified dataset split: 80% training, 10% validation, 10% testing - Fine-tuned transformer-based BERT models for classification - Evaluation using **Accuracy, Precision, Recall, F1-score, Confusion Matrices** - **Performance** - `BanglaBERT`: **88.60% accuracy** - Precision: 88.93% | Recall: 89.82% | F1-score: 89.38% - `sagorsarker/bangla-bert-base`: **83.12% accuracy** - BanglaBERT achieved stronger generalization across distinct categories, with some challenges in nuanced/low-resource senses --- ## 📊 Example Input & Output ### Example 1 **Input Sentence:** সে [TGT]চাল[/TGT] বাজার থেকে কিনে এনেছে। **Model Prediction:** - Correct Sense: *Rice (food grain)* ✅ --- ### Example 2 **Input Sentence:** সে খুব দ্রুত গাড়ি [TGT]চাল[/TGT]াচ্ছে। **Model Prediction:** - Correct Sense: *Driving (to operate a vehicle)* ✅ ---