Improved Neural Relation Detection for Knowledge Base Question Answering

摘要

Relation detection is a core component for many NLP applications includingKnowledge Base Question Answering (KBQA). In this paper, we propose ahierarchical recurrent neural network enhanced by residual learning thatdetects KB relations given an input question. Our method uses deep residualbidirectional LSTMs to compare questions and relation names via differenthierarchies of abstraction. Additionally, we propose a simple KBQA system thatintegrates entity linking and our proposed relation detector to enable oneenhance another. Experimental results evidence that our approach achieves notonly outstanding relation detection performance, but more importantly, it helpsour KBQA system to achieve state-of-the-art accuracy for both single-relation(SimpleQuestions) and multi-relation (WebQSP) QA benchmarks.