Research Of Multi-turn Dialogic Discourse Coherence

Coherence is an important attribute of dialogue,which enables dialogue to be integrated into a comprehensible whole.The assessment of dialogue coherence is a key indicator of dialogue quality evaluation.At present,the researches on text coherence mainly focus on monologue discourse such as news articles,while there are few studies on coherence modeling of dialogue.Therefore,this paper studies dialogue coherence modeling based on the traditional machine learning method and the deep learning method,and combines the semantic information contained in the dialogue text and the intention information represented by dialogue act labels to construct dialogue coherence modelsEntity-grid based coherence model is the most popular coherence modeling method.From the perspective of linguistics,the text coherence is modeled according to the entity distribution pattern between adj acent sentences in the text.Therefore,the first work of this paper is to improve the classical entity grid model by introducing dialogue-specific intention information and using the traditional machine learning method to model dialogue coherence.Firstly,we construct the dialogue act entigy grid containing the dialogue intention,and the dialogical intention transitions of entities between utterances are modeled.Then the dialogue coherence model is constructed by combining supervised machine learning algorithm.The experimental results on publicly available multi-turn dialogue dataset verify that the intention information has a certain guiding effect on dialogue coherence modelingThe above dialogue entity grid coherence model based on the traditional machine learning method requires manual feature extraction and has poor generalization performance.Moreover,it can not capture the long range entity transitions between utterances.Therefore,the second work of this paper proposes a deep neural network based dialogue entity grid coherence model.It integrates the entity grid representation with dialogue act and convolution neural network to capture the long range intention transitions of entities between utterances and model dialogue coherence.The experimental results show the effectiveness of the neural network model combined with dialogue intention to model dialogue coherence.The first and the second work of this paper is based on the entity grid for dialogue coherence modeling,which requires entity extraction.However,while sentences in dialogues are usually short,informal and colloquial,extracting entities is relatively difficult and may lead to error propagation.Therefore,in the third work of this paper,instead of extracting entities,we propose a novel dialogue act enhanced hierarchical model for dialogue coherence modeling.Through hierarchical encoder,we can directly model dialogue text to obtain dialogical semantic representation.Then the intention information of dialogue is fused at the utterance layer and dialogue layer respectively.The experimental results verify the effectiveness and robustness of our proposed model and the corresponding paper has been accepted by International Joint Conference on Neural Network 2019(IJCNN 2019).This paper studies several different methods of dialogue coherence modeling.Based on the different forms of dialogue(entity grid and text),we use a variety of natural language processing research methods(traditional machine learning and deep learning),and utilize different kinds of infromation of dialogue(semantic and intention)to construct dialogue coherence models.The experimental results verify the effectiveness and robustness of the proposed dialogue coherence models.