Word similarity score as augmented feature in sarcasm detection using deep learning

Abstract

Sarcasm detection is an important task in natural language processing (NLP). Sarcasm flips the polarity of a sentence and will affect the accuracy of sentiment analysis task. Recent researches incorporate machine learning and deep learning methods to detect sarcasm. Sarcasm can be detected by the occurrence of context disparity. This feature can be detected by observing the similarity score of each word in the sentence. Word embedding vector is used to calculate word similarity score. In this work, the word similarity score is incorporated as an augmented feature in the deep learning model. Three augmenting schemes in deep learning models are observed. Results show that in general, a word similarity score boosts the performance of the classifier. The accuracy of 85.625% with F-Measure of 84.884% was achieved at its best.