A Deviation based Ensemble Algorithm for Sarcasm Detection in Online Comments

Sarcasm refers to the use of irony to mock or convey contempt and involves the use of words that mean the opposite of what someone truly intends to convey. Online forums which enable users to express sarcasm as a sentiment tend to induce misunderstandings between different parties and obscure the users' true intentions. This leads to ambiguity being one of the prime challenges in detecting sarcasm. Another challenge in sarcasm detection is the rapidly growing size of language vocabularies with the addition of new slang words every day. Additionally, usage of emojis in online text can greatly influence the polarity of a sentence by inducing a sarcastic tone. These setbacks make sarcasm a particularly demanding sentiment to determine. In this paper, the statistical significance of various deep learning models for the purpose of detecting sarcasm in online comments containing emojis is explored. For the task of binary classification, GRU achieves an accuracy score of 73.44% with an F1-score of 73.96%. The proposed ensemble-based approach yields an accuracy score of 74.41% for the combination of LSTM and GRU, which is comparable to the accuracy achieved with conventional ensemble techniques such as max-voting and averaging. Twenty-six different hybrid combinations of deep learning models were explored and the most optimal performing ones were identified. CNN and Global Average Pooling 1D are two other architectures that were explored.