Direct relation detection for knowledge-based question answering

This study addresses the problem of relation detection for answering single-relation factoid questions over knowledge bases (KBs). In this kind of questions, the answer is obtained from a single KB fact in the form of subject-predicate-object. Conventional fact extraction methods have two steps: entity linking and relation detection, in which the output of the entity linking is used by the relation detection step to first find candidate relations, and then choose the best relation from candidate relations. Such methods have difficulties with the relation detection if there is an error or ambiguity in the entity linking step. This paper explores the relation detection task without the entity-linking step utilizing the hierarchical structure of relations and an out-of-box POS tagger. As relations are of different levels of abstraction, the proposed solution uses multiple classifiers in pipeline, each of which uses separate BiGRU neural networks fed with questions embedded with one-hot encoding at the character level. Besides, to increase the accuracy of the proposed model and to avoid the need for large amounts of training data, after each word of the question, its POS tag is inserted before feeding the network. The experimental results show that the accuracy of the proposed solution for the direct relation detection is 89.5%. In addition, the proposed solution can be used for the indirect relation detection whose accuracy is 96.3%, which is higher than state-of-the-art relation detection techniques. Finally, the positive effects of using POS tags have been examined.