ARCNN: A Semantic Enhanced Relation Detection Model for Knowledge Base Question Answering

Abstract

Relation detection plays an important role in knowledge base question answering (KBQA), and it is critical for the final performance of KBQA systems. The previous works mainly focused on enriching the information representations of questions and relations, and neglected the interaction information of questions and relations and different tokens within the relation. In this paper, we propose a semantic enhanced relation detection model called ARCNN, which is carefully designed by combining BiGRU, multi-scale semantic extracted CNN, and different attention mechanisms in a seamless way. Moreover, we combine four levels of relation abstractions to ensure the integrity of relation information and hence to enrich the relation representation. The experimental results on two benchmarks show that our ARCNN model achieves new state-of-the-art accuracies of 96.42% for SimpleQuestions and 90.4% for WebQuestions. Moreover, it helps our KBQA system to yield the accuracy of 81.5% and the F1 score of 72.0% on two benchmarks, respectively.

Tasks

Reproductions