Intent Recognition Of Short Text In Human-computer Dialogue Based On Deep Learning

Natural language understanding is a core component of a human-computer dialogue system,and intention recognition is one of the key technologies for realizing natural language understanding.With the continuous development of human-machine dialogue systems,more and more dialogue robot products come into people's lives,but humans often use short sentences and abbreviated words during human-machine dialogue.These short texts have the characteristics of short content,large amount of data,and irregular expression.These characteristics lead to the problems of many text noises,sparse features,polysemy,and inconsistent dialogue information before and after intent recognition.In order to solve the above problems,this paper first introduces the BERT model and the word2 vec model,and the fusion method of the two model vectors for the problem that the short text in the human-machine dialogue can not represent the polysemy and the lack of semantics in the vectorized representation.The short text vectorized representation method based on the joint model of BERT and word2 vec.Experiments show that the fused vectorized representation can greatly improve the classification performance of the classification model.Secondly,for the problem that the short text of human-machine dialogue contains multiple intents and sparse features,CNN and LSTM intent recognition models are described.The process of weight distribution of short text sequences by the multi-head attention mechanism is mainly explained.The multi-intent recognition method of mechanism and feature fusion,the comparison of experimental results shows that the recognition effect of this method is better.Finally,based on the BLSTM model,based on the BLSTM model,the two historical storage methods are described respectively by analyzing the two different multi-round dialogue situations,and the method of using the gate function to filter useful historical information is described.A short text intent recognition method based onBLSTM-based multi-round dialogue is proposed.Experimental comparison results show that the recognition effect of the BLSTM model with independent storage unit and gating function is superior to the simple BLSTM model and the BLSTM model with only storage unit.