Look! Who's Talking?

Abstract

Automatic classification of personality from language depends upon large quantities of relevant training data, which raises two potential problems. First, collecting personality information from the author or speaker can be invasive and expensive, especially in sensitive contexts. Second, issues of context or genre can reduce the usefulness of available training resources for broader personality classification. One approach to dealing with the first issue is to use external judges rather than the text's author. In this paper, we test the extent to which these personality perceptions are useful for training a classifier between different linguistic genres. Following disappointing cross-training results, we explore the projection of personality through specific linguistic factors. We find that while some differences are between the genres overall, some indicate that indeed personality is evidenced differently across situations. It is clear that care is needed leveraging resources from different domains for computational personality recognition.