Research On Answer Rejection Task For Customer Service Robot

The customer service robots in the limited domain often encounter scenes that need to reject to recognize during the service process,which refers to the situation that the robot cannot answer correctly due to the limitations of the knowledge base.Compared with giving an irrelevant answer,the customer service robot can gain more trust in users by rejecting to recognize and then transferring to human services.The early traditional solution is based on rules,such as sensitive word recognition,which requires a lot of expert knowledge and has no transferability.In recent years,with the extensive application of deep learning,neural networks are mostly used as encoders in natural language processing tasks to extract text features and then input into machine learning algorithms.The related research work of the scene of rejection recognition can be divided into three parts: classification with rejection,outlier detection and open set text recognition.These algorithms are also based on the two-stage framework.The model cannot affect the representation of text features by optimizing the task-related objective function.To solve the above problems,this paper proposes a recurrent neural network based on maximum margin squared hinge loss function to realize the rejection recognition ability of customer service robots.First,according to the characteristics of the short text of the user's questions,a bi-directional GRU model with attention mechanism supervised by classification tasks is proposed to encode the sentence vector,which achieves 95%classification accuracy on the closed set,indicating that the encoder model can well construct text data.Then we will achieve a deep rejection model based on this structure.Since there are new classes in the test set that are unseen in the training process,the normalization of softmax activation function will bring statistical deviation which causes the probability of unseen classes to be 0.Therefore,sigmoid function is used in the output layer and implement the "one-vs-rest" strategy in the multi-classification problem.Inspired by the objective function of the one-class support vector machine for deep outlier detection algorithm,the maximum margin squared hinge loss from SVM is used to replace the cross entropy in the recurrent neural network,and the text feature extraction is realized based on the maximum margin motivation,and the convergence speed of the model is also greatly improved.Finally,according to the 3? principle of normal distribution,the probability threshold of rejection is set.Compared with the baseline method,this end-to-end rejection model shows superior performance in the scene with fewer seen classes.The performance is increased by 2.24 times.