DeepLearning HW3 — Spoken-SQuAD Extractive QA

A minimal, reproducible BERT baseline for extractive question answering over ASR-transcribed passages (Spoken-SQuAD). Trains a span-prediction head on top of bert-base-uncased, logs EM/F1/WER, and saves figures/tables for the report.

Project Structure

• Bert.py — main training + evaluation script

• run_outputs/ — generated artifacts after running

  • figures/ — plots (training loss, EM/F1, WER, length histograms)
  • tables/ — history.json, preds_epoch_*.json (sample preds vs. gold)
  • base_model_wer.txt — WER per epoch (plain text)

• data/ — dataset folder

Dataset

Spoken-SQuAD JSON files:

• spoken_train-v1.1.json
• spoken_test-v1.1.json

Environment

Python 3.10+, CUDA (optional), and:

pip install torch transformers tqdm matplotlib
# if using evaluate/jiwer variants:
pip install evaluate jiwer

## Train and Evaluation
python3 Bert.py \
  --train_json /data/spoken_train-v1.1.json \
  --valid_json /data/spoken_test-v1.1.json