Extractive Question Answering

An extractive QA system built on the SQuAD dataset. Instead of generating an answer, the model learns to locate the span in the context that answers the question.

Why extractive

Extractive QA forces a clean definition of success: the answer already exists in the text, so the model's only job is to find its boundaries. That makes training, evaluation, and debugging far more interpretable than open generation.

Pipeline

1. Preprocess the SQuAD examples into question-context-answer triples.
2. Tokenize and align answer spans to token positions.
3. Train a span-prediction model (start/end indices).
4. Evaluate with exact-match and F1 against gold spans.
5. Infer on new context + question pairs.

What it taught me

NLP systems are easier to trust when you can point at the evidence. Building from the dataset level through to inference made the whole stack — tokenization, representation, evaluation — concrete rather than abstract.