Sociolinguistically Informed Natural Language Processing: Automating Irony Detection.

摘要

Irony detection is an important problem in computational sociolinguistic tasks, such as automatic community detection on the web. But the ironic/sincere distinction has proven to be a particularly difficult classification problem. Existing Machine Learning (ML) and Natural Language Processing (NLP) approaches, which tend to rely on simple statistical models built on top of word counts, are not very good at it. We hypothesize that this is because, in contrast to most text classification problems, word counts and syntactic features alone do not constitute an adequate representation for verbal irony detection. Indeed, sociolinguistic theories of verbal irony imply that a model of the speaker is a necessary condition for irony detection, at least in certain cases.