chandana-nuthangi/telugu-ner-thesis

Transfer learning for low-resource Telugu NER using DistilBERT-multilingual. Fine-tuned on Naamapadam (ACL 2023) with F1 0.78 on the manually annotated test set. Includes deployed Streamlit app for PDF entity extraction. # Telugu NER — Transfer Learning for a Low-Resource Language Fine-tuning DistilBERT-multilingual for Named Entity Recognition on Telugu, with a deployed Streamlit app for PDF entity extraction. **Result: F1 = 0.78** on Naamapadam test set — within 2 points of published SOTA (IndicNER), using a **2.7× smaller model** trained on **10% of the data**. --- ## The research problem Telugu has ~80 million speakers but remains **low-resource for NLP** — labelled datasets, pretraining corpora, and benchmarks are sparse. Existing SOTA Telugu NER models (IndicNER, MuRIL fine-tuned) achieve ~0.80 F1 but require 180–240M parameter models that are expensive to deploy at scale. This raises a research question about **model efficiency tradeoffs for low-resource NLP**: how much accuracy is lost when the model is heavily compressed via distillation, and is that loss acceptable for practical deployment? **The question this research answers:** Can a smaller, faster, distilled multilingual model deliver acceptable Telugu NER performance through transfer learning? ## Approach - **Model:** `distilbert-base-multilingual-cased` (66M parameters, 40% smaller than BERT) - **Dataset:** Naamapadam (Mhaske et al., ACL 2023) — largest publicly available Indic NER corpus, 507k Telugu sentences - **Method:** Fine-tuning for token classification with subword-to-word label alignment (3 epochs, learning rate 2e-5, batch size 32, fp16) - **Metric:** Entity-level F1 via seqeval on the 847-sentence manually annotated test set - **Baseline landscape:** IndicNER (ACL 2023 SOTA, IndicBERT-based), MuRIL, mBERT, XLM-R ## Results | Metric | Value | |---|---| | **F1 (test set)** | **0.7803** | | Precision | 0.7827 | | Recall | 0.7779 | | Accuracy | 0.9389 | | Model size | 66M params | | Training data used | 50k / 507k sentences (10%) | | Training time | 10 min 42 sec on Colab T4 | | Inference latency | ~50ms per sentence (CPU) | Trained on only 10% of Naamapadam Telugu, the fine-tuned DistilBERT lands …