Cross Corpus Emotion Classification Using Survey Data

Abstract

Although semantic analysis and machine learning are becoming well established parts of Natural Language Processing (NLP), extraction of discrete emotions from text remains an under-developed area. Even less frequently do we see application of these technologies to open-ended survey questions in fields such as political science, psychology, public policy and sociology. In these domains, the need for more fined-grained emotion analysis of text responses has become apparent, particularly for assessing nuanced responses of the population to unexpected high impact events or incidents. Doing such assessments in real time is even more difficult. We report preliminary results on an ambitious attempt to perform a cross-corpus emotion classification that applies data gathered in one survey to text collected at a different time from different sources. This research is one step in a broader agenda to create new NLP methods to code large-scale text data from surveys and social media to improve studies of emotion contagion through social media networks. Our report is based on a medium-scale experiment from a survey conducted in the Fall of 2016 during a crisis event. Preliminary evidence suggests that with careful calibration of survey instruments, and proper understanding of natural language expressions (encoded as machine learning features), a transfer of classification code should be possible for some strongly expressed and potentially actionable emotions, like anger.