Knowledge Base Question Answering with Attentive Pooling for Question Representation

Abstract

This paper presents a neural network model for a knowledge base (KB)-based single-relation question answering (SR-QA). This model is composed of two main modules, i.e., entity linking and relation detection. In each module, an embedding vector is computed from the input question sentence to calculate its similarity scores with entity candidates or relation candidates. This paper focuses on attention-based question representation in SR-QA. In the entity linking module, two attentive pooling methods, inner-sentence attention and structure attention, are employed to derive question embeddings, and their performances are compared in experiments. In the relation detection module, a new attentive pooling structure, named multilevel target attention (MLTA), is proposed to utilize the multilevel descriptions of relations. In this structure, the attention weights for aggregating the hidden states of question sentences are calculated using relation candidates as queries at the relation level, word level, and character level. Then, the similarity scores for relation detection are computed by matching questions to relation candidates at all three levels. The experimental results show that our proposed model achieves a state-of-the-art accuracy of 82.29% on the simple questions dataset. Furthermore, the results of ablation tests demonstrate the effectiveness of our proposed MLTA method for question representation.