Multi-domain Emotion Detection using Transfer Learning

Anonymous

Abstract: The task of emotion detection in text, particularly in informal and spontaneous messaging, such as email, posts, or tweets, varies in its scope and depth depending upon the requirements of the end application as well as the domain of use. The most popular emotion categories reported in research include the Ekman’s or Plutchik’s emotion models, but often the application domain requires a more specialized emotion categorization, for which there are insufficient annotated datasets available for training. The task is additionally complicated by social and cultural factors that make certain words and expressions emotionally charged in one context but entirely neutral in another. In this paper, we present a generalized approach of transfer learning for emotion detection that can be adapted to any domain and any set of classification labels. We show the performance improvements that could be achieved by fine-tuning our approach with limited annotated data from the target domain. This approach demonstrates good performance in predicting emotion categories previously unseen to the model, including domains different than those on which the model was originally trained. Furthermore, the system output can be easily adapted by end users to detect additional emotion categories. Lastly, we present an evaluation of this method on the publicly available SemEval 2018 Task 1e-c dataset and also a new annotated dataset consisting of tweets related to the French elections in 2017.

Paper Type: long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining