Using Features Inspired by Linguistics and Psychology to Improve Automatic Detection of Subtle Information in Text
The goal of natural language understanding is to extract relevant information from natural language, including more subtle information such as intentions, emotions, and identity. In this particular area of information extraction, there's been a historical divide between approaches relying on sophisticated computational tools and those relying on theoretical constructs from psychology and linguistics. Pearl tries to bridge this divide, and discusses recent findings in automatic deception detection and sentiment analysis that leverage this kind of bridging approach. In particular, she incorporates features inspired by linguistics and psychology into symbolic computational approaches. In each case, she finds significant improvement for "hard cases" where previous state-of-the-art approaches have struggled.