While the evaluation task of microblog emotion is a multi-label classification problem,the traditional text representing methods,which are usually based on vector space model,fail to provide more effective semantic features.Word embedding technology is based on deep learning,which can well capture the syntax and semantic relations between words,and build sentence representing effectively according to semantic compositionality.A multi-label emotion classification system was proposed.First,word embedding for Chinese words was learned from a large scale of unlabeled Chinese microblog text dataset.Second,the Convolution Neural Network (CNN) model was exploited to train a supervised multi-emotion classifier.Third,the learned CNN model was used to composite the feature vector for sentences from microblog.At last,these sentence vectors were treated as semantic features to train the multi-label classifier,which was used to finish the multi-label emotion classification for microblog.Based on the open dataset from microblog emotion evaluation task of NLPCC (Natural Language Processing and Chinese Computing) conference in 2013,the best performance of the proposed system achieved 19.16% and 17.75% improvement in the loose metric and the strict metric,respectively,comparing to the best performance of all the evaluation results.The state-of-art performance,which was achieved by the method of exploiting Recursive Neural Tensor Network model to composite the sentence vector,was also outperformed by the proposed system up to 3.66% and 2.89% on the two metrics.Several multi-label classifiers were employed to compare different feature representing methods,and the sentence vectors based CNN feature space were showed to have the most discriminative emotion semantic.The pattern recognition in the semantic composition procedure was showed by analyzing the training iteration of CNN model.
