Multi-Source Domain Adaptation for Emotion Classification Using Bi-LSTM and Broad Learning

Existing domain adaptation methods for TEC task have tended to explore single-source domain, rather than on multi-source domain adaption. However, in fact, there is limited information and volume from single-source domain, which may affect emotion classification efficiency. Thus, to improve the performance of domain adaptation, we present a novel multi-source domain adaptation approach for TEC task, by combining BL and DL in this chapter. Specifically, to better capture the contextual features, we first use BERT and Bi-LSTM to extract DIF from each source domain to the same target domain. Then, to more effectively conduct the multi-label classification task, we adopt BL to train multiple classifiers on the basis of DIF. In addition, we design a co-training model to boost these classifiers. Finally, we conduct several experiments on four data sets by comparing with the baseline methods. The experimental results show that our proposed method can outperform the baseline methods for the textual emotion classification.