Korea University (graduate project) · Apr 2020 – Jul 2020
Open-Domain Korean Question Answering
01Background & Goals
- We built a system that answers open-domain questions using search engines.
- It collects documents from Naver and Daum, splits them into paragraphs and finds the answer span with a BERT-based language model.
02Contributions
Retrieval and reading comprehension
- Collected candidate documents from search engines and split them into paragraphs
- Fine-tuned BERT, KoBERT and HanBERT for span extraction and chose the final model through quantitative and qualitative evaluation
Web demo
- Built and ran a Flask web demo
03Tech Stack
- Framework / Platform
- PyTorch, Flask, Transformers
- Methodology
- Machine reading comprehension, Span extraction
04Results
- Ran a web demo; code released