For human and automatic text annotation of emotions, it is as- sumed that affect can be traced in language on (combinations of) individual words, text fragments, or other linguistic pat- terns, which can be identified and labelled correctly. For exam- ple, many sentiment analysis systems consider isolated words affectively meaningful units, whose proportions in a given text reveal its overall affective meaning. However, whether these words and their combinations as identified either by humans or algorithms also match the actual feelings of the authors remains unclear. Potential discrepancies between affect expression and perception in text have received surprisingly little scholarly at- tention, although a number of studies has already identified dis- parities between self- and other-annotation in affect detection for speech and audio-visual data. Therefore, we ask whether a similar difference shows in annotations of emotions in text.

The influence of emotion on (the early stages of) speech production processes, notably content selection has received little scholarly attention. Goudbeek & Krahmer (2012) found evidence for alignment at the conceptual level: speakers may start using a dispreferred attribute over a preferred attribute in their referring expressions when they are primed by a pre- recorded female voice in a preceding interaction. The current study aimed to assess the role of emotion (using amusement and disgust) in alignment, while simultaneously replicating this finding in a more naturalistic setting involving two human participants in naturalistic dialogue. Our results replicate the findings by Goudbeek & Krahmer (2012), generalizing their findings to a much more naturalistic setting. In addition, we found that amused, but not disgusted speakers tend to use the preferred attribute more to describe objects to their conversational partner.