AI project2025.06 – 2025.08Team project · Excellence Award
POSPEAK (KILL KONG) — Konglish → natural English correction AI (FLAN-T5 · Qwen + RAG)
A generator plus retrieval (RAG) that rewrites awkward or Konglish sentences into natural English
Problem
Konglish and stiff formal English slip past grammar checkers; learners need natural, context-aware rewrites
My role
Build the correction data and find a training strategy for T5-family models
What I did
- [Own] Built and cleaned word-level and context-level correction pairs
- [Own] Fine-tuned FLAN-T5 with LoRA and compared three strategies (word→context, context→word, joint) with BERTScore
- [Final service KILL KONG] Qwen-2.5 (1.5B) fine-tuning + a 630-entry Konglish RAG DB + per-user memory, combined with LangChain in a React Native · FastAPI app
Results
- Found catastrophic forgetting in sequential training (BERTScore F1 0.47 → 0.32) and switched to joint training
- Joint training reached validation BERTScore F1 ≈ 0.74 · team Excellence Award
Decisions & trade-offs
What I chose, and what I gave up for it
Sequential training (easy word-level first, then context) felt intuitive, but I compared three strategies under the same conditions — extra experiments, but it surfaced catastrophic forgetting and led to joint training.
Performance was logged right after each stage; looking only at the final score would have ended at 'it's just low'.
Looking back · open questions
What I would change, and what is still open
If sequential training is required, consider replay (mixing in earlier data) or regularization such as EWC; here the data were small enough that joint training was simplest and best.
POSCO academy — Excellence Award & certificate