Transfer learning for emotion detection in conversational text: a hybrid deep learning approach with pre-trained embeddings

Understanding the emotions and sentiments from conversations has relevance in many application areas. Specifically, conversational agents, question-answering systems, or areas where natural language inference is used. Therefore, techniques to detect emotions from conversations have become the need of the moment. The convolutional network and recurrent networks have shown different capabilities in text representation. This work proposes a hybrid deep learning network based on the convolutional-recurrent network used to detect the emotions of people based on conversational text. A convolutional network has the ability to capture local patterns and relationships and is inherently shift-invariant. At the same time, the recurrent network captures long-range dependencies in sequential information. This work also utilises the power of transfer learning by employing pre-trained embeddings from Neural Network Language Model models. These pre-trained representations, generated from vast text corpora, encode rich semantic information about words. This study investigates a novel approach towards text-based emotion detection using pre-trained Neural Network Language Model embeddings with hybrid convolutional-recurrent architecture. The proposed hybrid experimental setup has been evaluated on the Empathetic Dialogues dataset and contrasted with the state-of-the-art works. A comparative analysis reveals that the proposed Convolutional Neural Network with a Bidirectional Gated Recurrent Unit hybrid approach with Neural Network Language Model embeddings achieves superior performance and accuracy.