Topic-Enriched Word Embeddings for Sarcasm Identification

Sarcasm is a type of nonliteral language, where people may express their negative sentiments with the use of words with positive literal meaning, and, conversely, negative meaning words may be utilized to indicate positive sentiment. User-generated text messages on social platforms may contain sarcasm. Sarcastic utterance may change the sentiment orientation of text documents from positive to negative, or vice versa. Hence, the predictive performance of sentiment classification schemes may be degraded if sarcasm cannot be properly handled. In this paper, we present a deep learning based approach to sarcasm identification. In this regard, the predictive performance of topic-enriched word embedding scheme has been compared to conventional word-embedding schemes (such as, word2vec, fastText and GloVe). In addition to word-embedding based feature sets, conventional lexical, pragmatic, implicit incongruity and explicit incongruity based feature sets are considered. In the experimental analysis, six subsets of Twitter messages have been taken into account, ranging from 5000 to 30.000. The experimental analysis indicate that topic-enriched word embedding schemes utilized in conjunction with conventional feature sets can yield promising results for sarcasm identification.