A Transfer Learning Based Boosting Model for Emotion Analysis

Emotion Analysis determines the emotion of a text. Supervised Machine learning algorithms are effective for Emotion Analysis, but they need a lot of labelled data. It is a labor-intensive process and often needs instructions of experts to annotate data. In this paper, we propose a transfer learning approach for emotion analysis based on Adaboost(EATAdaBoost) by adapting the knowledge learned from labelled source data to the target domain which has none or few labelled data. We try to establish connections between source instances and target domain. Word2vec semantic similarities between source instances and common non-domain-specific emotional words which occur frequently in both domains are used as a bridge. If the similarity is bigger than a threshold, we think the source instance is useful for learning target task. In addition, we conduct extensive experiments and the results show that our algorithm is superior to baselines.