Linguistic-based Patterns for Figurative Language Processing: The Case of Humor Recognition and Irony Detection

Figurative language represents one of the most difficult tasks regarding natural language processing. Unlike literal language, figurative language takes advantage of linguistic devices such as irony, humor, sarcasm, metaphor, analogy, and so on, in order to communicate indirect meanings which, usually, are not interpretable by simply decoding syntactic or semantic information. Rather, figurative language reflects patterns of thought within a communicative and social framework that turns quite challenging its linguistic representation, as well as its computational processing. In this Ph. D. thesis we address the issue of developing a linguisticbased framework for figurative language processing. In particular, our efforts are focused on creating some models capable of automatically detecting instances of two independent figurative devices in social media texts: humor and irony. Our main hypothesis relies on the fact that language reflects patterns of thought; i.e. to study language is to study patterns of conceptualization. Thus, by analyzing two specific domains of figurative language, we aim to provide arguments concerning how people mentally conceive humor and irony, and how they verbalize each device in social media platforms. In this context, we focus on showing how fine-grained knowledge, which relies on shallow and deep linguistic layers, can be translated into valuable patterns to automatically identify figurative uses of language. Contrary to most researches that deal with figurative language, we do not support our arguments on prototypical examples neither of humor nor of irony. Rather, we try to find patterns in texts such as blogs, web comments, tweets, etc., whose intrinsic characteristics are quite different to the characteristics described in the specialized literature. Apart from providing a linguistic inventory for detecting humor and irony at textual level, in this investigation we stress out the importance of considering user-generated tags in order to automatically build resources for figurative language processing, such as ad hoc corpora in which human annotation is not necessary. Finally, each model is evaluated in terms of its relevance to properly identify instances of humor and irony, respectively. To this end, several experiments are carried out taking into consideration different data sets and applicability scenarios. Our findings point out that figurative language processing (especially humor and irony) can provide fine-grained knowledge in tasks as diverse as sentiment analysis, opinion mining, information retrieval, or trend discovery.