Knowledge base question answering based on regularization and feature fusion

Abstract

Knowledge base based question and answer method has achieved remarkable success by deep learning. However, the method research process remains challenging due to the difficulty of capturing the global information of the question, the inconsistency of the outputs in training and prediction, and the low recognition accuracy of the current topic entity detection. Here, we propose an improved model TRBAM based on BAMnet to efficiently and effectively solve above problems. TRBAM is improved by Transformer and R-dropout. In the problem feature extraction layer, we use BiLSTM and Transformer to extract features from the problem respectively, and fuses the two extracted problem features to obtain a new problem representation, so that the model can more fully capture semantic information in the problem; especially, we improve the generalizability by R-dropout. We improve the subject entity detection model entnet by R-dropout to improve the accuracy in recognizing the best subject entities. The experimental results show that the improved model TRBAM has some performance improvement compared with the BAMnet model, the improved entnet model has effectively improved the recognition accuracy of the best subject entities, and the overall method of this paper has some advantages compared with other methods.