dissertation_gec

MSc Computational Linguistics dissertation: Error-Type-Aware, CEFR-Adaptive Feedback for L2 Writing.

Stack

• Python 3.11+
• PyTorch, Transformers, PEFT/LoRA
• ERRANT, spaCy, language-tool-python
• pandas, scikit-learn, matplotlib
• Streamlit (demo UI)

Datasets

• W&I+LOCNESS v2.1 (BEA-2019)
• JFLEG

Conventions

• Random seed: 42 everywhere (numpy, torch, random, transformers).
• All structured outputs written as JSONL.
• Academic text uses Harvard referencing style.
• Code is modular: one file per concern.
• Log everything for reproducibility (config, seeds, versions, metrics).

Project goals

1. Ingest and normalise W&I+LOCNESS and JFLEG into unified JSONL.
2. Type errors with ERRANT and aggregate by CEFR level.
3. Fine-tune a correction model with LoRA adapters.
4. Adapt feedback to learner CEFR level (scaffolded vs. concise).
5. Evaluate with ERRANT P/R/F0.5 and GLEU (JFLEG).
6. Ship a Streamlit demo.

Reproducibility

• Pin dependencies in pyproject.toml / requirements.txt.
• Log run configs and git commit hash per experiment.
• Deterministic seeding across all libraries.