Sarcasm Detection in Telugu Text: A Deep Learning Framework Leveraging Hybrid Approach

Abstract

Sarcasm, a form of speech that conveys contempt or disdain, positions a significant challenge in Natural Language Processing (NLP). Automatically detecting sarcasm is difficult, due to its subtle and context-dependent nature, often relying on nuanced cues such as tone, irony, and hyperbole. Existing approaches, relying on either machine learning (ML) or rule-based methods have shown limited success in accurately identifying sarcasm. To tackle this challenge, a novel hybrid method, named Bidirectional Long Short-Term Memory + Gated Recurrent Unit (BiLSTM+GRU), BiLSTM with attention and BiLSTM+ Recurrent Neural Network (BiLSTM+RNN) combines the strengths of both models. Moreover, the methodology leverages Bidirectional Encoder Representations from Transformers (BERT)-based tokenization, TeluguBERT used for context-aware word embedding. Term Frequency- Inverse Document Frequency (TF-IDF), to capture the complexities of sarcasm. After that, the Kernel linear discriminant analysis (Kernel LDA) for dimensionality reduction is used, and feed the resulting features into a Deep learning architecture for Sarcasm Detection (SD). This study paves the way for creating more sophisticated and reliable SD tools, which can be leveraged in various domains, including emotional intelligence analysis, public opinion analysis, and online content moderation, ultimately enhancing the accuracy of human-computer interactions and social media intelligence.