基于宽度神经网络的FAQ问答模型

he FAQ question and answer data generally has problems such as short length, lack of context, and lack of grammatical structure information. Deep neural networks often causes optimization difficulties and over-fitting. Shallow features such as local features are especially important for FAQ questions and answers. In order to combine different shallow features, this paper proposes a FAQ question-and-answer model based on breadth neural network. This model splices the hidden layer of convolutional neural network based on word granularity and the hidden layer of recurrent neural network based on word granularity. Learning phrase features, local features, etc., can also learn timing features. In the real FAQ customer service data of a company, the model compares the current strong classification model and achieves the best results. Experiments show that using shallow-width neural networks to combine different shallow features can enhance the robustness of the model.