Research Of Dialogue Act Recognition Algorithm Based On Deep Learning

Dialogue act describes the pragmatic function of the utterance and can indirectly reflect the intention of the speaker,it is essential for understanding human utterance.Dialogue act recognition,as an important part of dialogue system,machine translation,speech recognition and other applications,has attracted widespread attention of researchers in recent years and has achieved significant achievements.However,there are still many difficulties and challenges in the field of dialogue act recognition research,such as:irregular daily expression,chaotic sentence structures,and complex dialogue composition structures make semantic inference difficult and deep learning models difficult to train.And most of the existing models focus only on the utterance,ignore the statistical information contained in the dialogue act label,which makes it difficult to improve the recognition accuracy.In order to overcome the above-mentioned shortcomings,this paper proposes two types of dialogue act recognition algorithms based on deep learning:(1)Aiming at the problem of arbitrarily used words in the utterance and confusion of sentence structure,a dialogue act recognition algorithm based on semantic reasoning is proposed.The algorithm calculates accurate utterance semantic expression by introducing a cyclic semantic reasoning process.This process uses the multiple attention mechanism and semantic update mechanism to infer the utterance semantic information by combining the utterance itself and the previous context information.(2)Aiming at the problems of complex dialogue composition structure and lack of label statistical information in the existing methods,a dialogue act recognition algorithm is proposed which fuses the label information.The algorithm first uses hierarchical project method,in which DA labels are embedded in three levels(word,utterance,and DA label),and expressed as a vector.Then these three levels of vectors are added to different places on the network to find keywords,provide the above semantic information and guide the model classification through pragmatic function matching.We compare the model proposed in this paper with existing methods,and the experimental results show that both the semantic inference process and the DA label project method proposed in this paper have effectively improved the recognition results.our model has reached the highest accuracy under the same conditions of using information,and get the same performance as the latest model without using the following context information.