Diverse and Effective Question-Answer Pair Generation from Children Storybooks

Jun. 11, 2023
Sugyeong Eo et al.

1. QAG task (QA pair Generation)
- have raised interest in educational field
- however, existing methods generate biased explicit questions
- that means the diversity of QA types remains a challenge

2. Proposed QAG Framework
- QFS-based answer generator + iterative QA generator + relevancy-aware ranker

1) QFS-based answer generator
- QFS = Query-focused summarization
- aims to obtain diverse answer candidates
- summarization -> answer generation model(AGM)

2) Iterative QA generator
- designed to increase variety of question types
- question generation model(QGM) -> generative question-answering model(QAM)

3) Relevancy-aware ranker
- produces relevant scroses for each QA pair
- trained with in-context negative sampels
- determines the top-N outputs among the generated candidates

* there is a issue of duplication exists where the top-ranked pairs constitute similar forms
-> to alleviate this, they measure the Rough-L score between the selecting pairs and the previously selected pairs
-> through this, they down-scale the scores of pairs which exhibit high overlap with previously selected pairs
-> this allows the diversity of QA types

* AGM, QGM, QAM -> pretrained BART-large, Fairseq
* ranking model -> RoBERTa-base, huggingface

3. Performance: SOTA