Open-Domain Question-Answering using NLG based query expansion and ranking method

Abstract

Open-Domain Question-Answering is the combined task of machine reading comprehension and information retrieval. Unlike general Question-Answering, Open-Domain Question-Answering lacks context data that contains the answer of the question, so it requires retrieving context candidates. To solve this task, we propose 3 approaches. 1.Query expansion from reading comprehension dataset, 2.Normalize reading comprehension output via sigmoid function, 3. Ranking and merging score with a threshold. In our experiment with Japanese Question-Answering dataset without context, our approach improves exact match score over previous method.