Research On Dialog-Act Recognition Machine Learning Algorithm

As the Dialogue Acts can describe the intention of the speaker in a certain extent, dialogue acts recognition has been the basic requirements of many tasks, such as the designing of the human-machine dialogue system, discourse understanding and translation. However, the conventional classifiers get poor performance when applied to this task due to the multiple modal features, the diversity and the heavily imbalanced labels.The related work of dialogue acts recognition at home and abroad is introduced after the dialogue theory expounded. Then the research of dialogue acts' modeling and recognition is developed from two aspects: recognition method and feature selection. For a Chinese dialogue acts recognition task in the CASIA-CASSIL corpus, 10 kinds of features are extracted after analysis and summary of the 7920 samples in it. Different kinds of features have different dimensions and stay in different feature space, it seems unreasonable to concatenate these features into a long vector and then deliver this feature vector to a classifier. To address this issue, we investigate into the multiple kernel learning method as well as the traditional but effective method SVM. This method uses the kernel trick to map the original features into a unified Hilbert space and then combine them. To show the effectiveness, we compare it with several baseline methods. Results of the experiments demonstrate that the multiple kernel learning method performs much better than all the baseline methods.In view of the multimodal features and imbalanced distribution, this paper puts forward the adapted multiple kernel SVM method in terms of the dialog acts recognition algorithm. Compared with other algorithms, it can capture the multi-modal features better and make full use of the information. It can improve the classification accuracy eventually.