Predicting Next Dialogue Action in Emotionally Loaded Conversation

This paper reports on creating a neural network model for prediction of the next action in a dialogue considering conversation history, i.e. entities, context variables and emotion indicators marking emotionally loaded user utterances. Several experiments were performed to see how the information about emotions affects the accuracy of the model. For the purposes of these experiments, a dataset containing 206 dialogs in Latvian in the transport inquiry domain was created containing both neutral and emotionally loaded utterances. To see if the proposed next dialogue action prediction model architecture is suitable for other languages, the original Latvian utterances were translated into English and a separate model was trained with English data. Some experiments were performed training the model with the data in one language and testing with the data in another language as well as training a single model using data in both languages. Experiments were performed with several fastText and Transformer pre-trained embedding models. Models for both languages Latvian and English achieved 0.91 accuracy on 10-fold cross-validation.