Sentiment Analysis in Social Media Posts using Deep Learning

Abstract

This master’s thesis presents a rigorous empirical comparative study evaluating text sentiment classification capabilities of deep learning and Transformer-based architectures. Utilizing two benchmark datasets ,the SST-2 binary sentiment corpus [29]and the TweetEval sentiment dataset [8]the research designs, implements, and optimizes three distinct neural pipelines: CNN, LSTM, and fine-tuned DistilBERT. The empirical results reveal that finetuned DistilBERT achieves the highest performance in sentiment classification with a testing accuracy of 70.29% on TweetEval, while the LSTM network records a testing accuracy of 56.38%. CNN achieves competitive results with a testing accuracy of 57.95%. Furthermore, an independent cross-domain validation experiment tracking 1,500 ChatGPT generated sentences [15]reveals distinct architectural biases under distribution shifts. The study addresses critical technical challenges including severe class imbalance and hardware optimization constraints, and proposes future research directions toward integrating Large Language Models (LLMs) and extending these pipelines to Algerian Dialectal Arabic (Darja).