Multitask deep learning for sentiment analysis with sarcasm detection in bilingual code-mixed social media content

Abstract

Sentiment analysis in social media often hindered by sarcasm, which can reverse text meaning, and bilingual code-mixing, which adds complexity in non-English primary context. Existing approaches extract separate features for each language and translate them into a single language, resulting in the loss of contextual meaning and omission of crucial features. This paper proposes a multitask learning model for sentiment analysis with sarcasm detection tailored to bilingual code-mixed social media content. A hybrid feature engineering technique is integrated into a multitask deep learning architecture designed to capture the nuances of sentiment and sarcasm while addressing the complexities of processing bilingual code-mixed content. The hybrid technique combines domain-knowledge-based natural language processing (NLP) with a deep learning-based embedding approach. It includes rule-based preprocessing, normalization, spellchecking, feature extraction and selection, and feature representation. The engineered features are integrated into a multitask deep learning network using bidirectional long short-term memory (Bi-LSTM) combined with gated recurrent units (GRU). Using a public dataset that contains bilingual code-mixed social media content related to public security, our proposed model achieved a higher F1score compared to two baseline models that employ single task and multitask approaches.