SF-ANN: leveraging structural features with an attention neural network for candidate fact ranking

Candidate ranking is the process of selecting the candidate with the best matching probability to the question after generating candidates in the knowledge base question answering (KBQA) task. It is a representative problem in mining matching relationships between candidates and questions. Previous research works always model questions and candidate representations separately, ignoring their impact on each other. The text information is too short to capture rich features in the KBQA task. Therefore, our work presents an attention neural network (ANN) fused with structural features (SF-ANN) to rank candidate facts jointly. First, two types of attention mechanisms are used to capture the correlation between the question and the candidate fact: a mutual-attention mechanism that captures the correspondence between the sentence components of a question and each part of a candidate and an intra-attention mechanism that captures the self-dependency of the concatenation between a question and a candidate fact. Second, an ANN is designed for fusing these two types of attention mechanisms to deeply couple interactive information of the input. Finally, knowledge base structural features are introduced to supplement the semantic information to increase the richness of the information. Three mutual attention mechanisms are applied for fusing them into the ANN, resulting in higher information gain. The experimental results on the SimpleQuestions (SimpleQ) benchmark demonstrate that the proposed model achieves a higher ranking accuracy (82.9%) than the state-of-the-art models. Moreover, the ablation study on SimpleQ and WebQuestionsSP (WebQSP) shows that leveraged features and the propoesd ANN both contribute to performance improvement.