University of Edinburgh

School of Informatics

An exploration of human emotion perception from short texts

4th Year Project Report

Artificial Intelligence with Psychology

Abstract: In order to make artificial agents more human it is important that the agents can detect and convey some level of emotional presence. But how well do humans themselves perceive emotion? Working independently, 56 judges each rated 10 short and 10 long text extracts. These texts were taken from the first 200 words of selected weblogs (online diaries) and were then divided into 50 and 150 word sections. The main finding was that it is easier to rate longer texts due to the increase in contextual information. The secondary finding is that the level of contextual similarity that emotions have differs from emotion to emotion and that Acceptance is much less distintive (and appears more neutral) than all of the other emotions. Subjects gave low intensity scores to the long texts which had been selected to convey Acceptance but much higher intensity for shorter Acceptance texts. Comparing this with statistics from the BNC, we suggest that in situations where there is less context, emotional stereotypy cannot be used as effectively as in longer texts where there is more context. Anger and Anticipation are also hard to distinguish from the rest of the emotions, but to a lesser extent, and so they are slightly easier to rate than Acceptance, but still less so than the other emotions.

Download the full dissertation