Research-level NLP system that summarizes Tamil sports news while preserving key named entities.
- 0
- technologies
- 0
- shipped features
- 0
- decisions logged
5 sections below
A Tamil sports news summarization system combining Named Entity Recognition with abstractive deep learning. The model preserves critical entities — players, scores, locations — that generic summarizers routinely drop, reducing information loss by roughly 30% against the baseline. Built as my postgraduate main project at Pondicherry University and deployed as a live Hugging Face Space.
- Combined NER with abstractive deep learning to preserve players, scores and locations in generated summaries
- Reduced information loss by ~30% compared to the baseline summarizer
- Solved low-resource language challenges: Tamil-specific tokenization, normalization and preprocessing
- Trained and evaluated on a curated Tamil sports news corpus
- Live Hugging Face Space demo
The section above says what this does. This one says why it is built the way it is — including the parts that did not come free.
- Fine-tune with NER wired into trainingoverUse a general-purpose summarisation model as-isPost-process generated summaries to reinsert missing entitiesBecause
Reinserting entities after the fact treats the symptom, not the cause — the model has already learned that dropping names is an acceptable shortcut. Training entity preservation directly into it is what produced the ~30% reduction in information loss over the baseline.
What it costEvery training run needs entity-tagged data, not just raw text-summary pairs — the curated Tamil sports corpus was more expensive to build than for a plain summarisation baseline.
- Parameter-efficient fine-tuning (PEFT)overFull fine-tuning of mT5Train a smaller model from scratch for Tamil onlyBecause
mT5 already carries broad multilingual and linguistic knowledge that a from-scratch model would need a much larger Tamil corpus to relearn. PEFT keeps that pretrained knowledge intact and trains only a small set of added parameters, which fit the compute available for a postgraduate project.
What it costPEFT trades some ceiling on task-specific performance for that efficiency — full fine-tuning would likely edge it out given unlimited compute, which was never the constraint being optimised for.
- Tamil-specific normalisation and preprocessingoverUse an off-the-shelf multilingual tokenizer unchangedTransliterate Tamil into a Latin-script intermediate formBecause
A multilingual tokenizer treats Tamil as one of a hundred languages it was never tuned for. Writing normalisation rules for Tamil's actual morphology addressed the low-resource-language problem the project set out to solve, instead of working around it.
What it costThat preprocessing is Tamil-specific and doesn't transfer — extending this to Telugu or Kannada would mean redoing this step, not reusing it.
// documentation is hidden by default to keep the page light
// slide deck is hidden by default to keep the page light
Seen enough?
There are more where this came from — or skip the browsing and just say hello.