Conversational transfer learning for emotion recognition

Recognizing emotions in conversations is a challenging task due to the presence of contextual dependencies governed by self- and inter-personal influences. Recent approaches have focused on modeling these primarily via supervised learning. However, purely strategies demand large amounts annotated data, which lacking most available corpora this task. To tackle challenge, we look at transfer learning as viable alternative. Given amount conversational investigate whether generative models can be leveraged affective knowledge for detecting context. We propose an approach, TL-ERC, where pre-train hierarchical dialogue model multi-turn (source) then its parameters emotion classifier (target). In addition popular practice using pre-trained sentence encoders, our approach also incorporates recurrent that inter-sentential context across whole conversation. Based idea, perform several experiments multiple datasets find improvement performance robustness against limited training data. TL-ERC achieves better validation performances significantly fewer epochs. Overall, infer acquired from generators indeed help recognize conversations.