Multilingual Sarcasm Detection for Enhancing Sentiment Analysis using Deep Learning Algorithms

Abstract

Recent years have seen a notable rise in online opinion-sharing, underscoring the demand for automated sentiment analysis tools. Addressing sarcasm in text is crucial, as it can significantly influence the effectiveness of sentiment analysis models. This research explores how sentiment analysis (SA) and sarcasm detection (SD) intersect, highlighting challenges in identifying how sarcasm influences sentiment polarity. Sarcasm, a type of irony, poses computational difficulties due to the lack of nonverbal cues in written texts. Users often express opinions in their preferred languages, underscoring the need for sentiment analysis tools that can adeptly handle sentiment and sarcasm across diverse languages. We propose the incorporation of sarcasm features into the architecture of sentiment analysis models, employing classifiers and embeddings, including BILSTM or LSTM alongside word embedding techniques such as Word2vec, FastText, Glove, and Bert. We conducted experiments using the ArSarcasm-v2 Dataset for Arabic, the IMDB Movie dataset and IsarcasmEval dataset for English, and the SentiMixArEn dataset for code-mixed language scenarios. The results demonstrated consistent accuracy enhancements ranging from 2% to over 10%, highlighting the positive impact of incorporating sarcasmrelated information. Additionally, the Bi-LSTM model with GloVe embeddings achieved higher accuracy across all scenarios compared to other methods.