Single-Entity-Single-Relation Question Answering with Minimal Annotation

Abstract

We present a novel bag-of-words based approach that automatically constructs a semantic parsing based question answering (QA) system tailored to single-entity-single-relation questions. Given a large community QA pair corpus and a knowledge base, our approach uses knowledge base entries to supervise relation extraction from the corpus, reduces noise in the extracted data via unsupervised clustering, and learns to identify each relation’s question patterns. We implement the approach on a large Chinese corpus with little annotation, which we believe is one of the first of its kind. Experiments show that our implementation manages to answer questions in test cases independent of the corpus with relatively high accuracy and to avoid answering questions beyond its scope, achieving a high accuracy on answered questions.