Sentiment Classification with Convolutional Neural Networks: An Experimental Study on a Large-Scale Chinese Conversation Corpus

Abstract

Previous research on sentiment classification by machine learning algorithms has shown that they usually work well with large-scale dataset. However, most open datasets for Chinese sentiment classification are quite small. In this paper we build a large-scale annotated Chinese sentiment dataset by filtering a vast amount of human-computer conversations. We conduct thorough experiments by using Convolutional Neural Networks (CNNs) and other classical machine learning methods. The experimental evaluation is made on a human-annotated dataset and COAE2014 task. The extensive experiments demonstrate that Chinese sentiment classification task can benefit from our dataset and CNNs model can achieve better performance than classical machine learning approaches.