A Home Service-Oriented Question Answering System with High Accuracy and Stability

Abstract

With the development of deep learning, neural network-based (NN-based) methods have been applied in question answering (QA) widely and achieved significant progress. Although an NN-based QA system can obtain better performance and save manual efforts, the system is likely to suffer attacks from the external perturbation, due to its character of being black boxes. Limited in the area of home service, we present an innovative method for constructing an NN-based QA system. In our method, the accuracy can be further increased, and the stability can be enhanced in the meantime. Inspired by observing the process of performing home services, the tool information (tool names and tool sequences) is integrated with question terms, as a way of extending the question representation. The conception of attribution (word importance) is introduced to gauge the word importance since NN-based models can be easily affected by the uninformative question terms. In order to optimize the model parameters effectively, the reinforcement learning is employed and both factors on accuracy and stability are regarded as rules in designing rewards. A few state-of-the-art methods are adopted to evaluate the effectiveness of the proposed method. The experimental results demonstrate that the model ability to produce effective answers in QA can be further improved with our method, and the model stability on perturbations can be enhanced with our method.