Volume 2 - Issue 3, May - June 2026

Sentiment analysis over short, noisy, user-generated social-media text remains a notoriously brittle classification problem: lexical sparsity, sarcasm, code-mixing, and a long-tail label distribution all conspire to deflate the accuracy of conventional bag-of-words pipelines. In this article we revisit the problem through a dual lens. First, we develop a detailed theoretical treatment that re-casts sentiment classification as a sequential decision-making process embedded inside a contextual encoder-decoder, drawing equally from the deep learning literature on self-attention and from the reinforcement-learning literature on policy-gradient and actor-critic methods. Second, we report a deliberately lean empirical study on a publicly available Twitter corpus (twitter_training.csv) using a TF–IDF feature representation and three classical learners—Logistic Regression (LR), Multinomial Naïve Bayes (MNB), and Linear Support Vector Classifier (L-SVC)—so that the reader can reproduce every number on a laptop within minutes. The best classical baseline (LR/L-SVC) attains an accuracy of 68.95% on a held-out test partition of 10,000 randomly drawn tweets. We argue, supported by theoretical derivations and a sequence of ablation-style thought experiments, that a hybrid framework combining a Transformer-style contextual encoder with a Proximal Policy Optimization (PPO) reward-shaping head can in principle push performance well beyond the bag-of-words ceiling. The paper is intended to be both a reproducible baseline and a self-contained tutorial introduction to the deep-learning ∪ reinforcement-learning interface for newcomers to sentiment analysis.

Varun Singh, Vansh Kumar, Aditya Sikarwar, Sanidhya Dhangar, Vipin Singh, Tarun Rathore, Rashmi Pandey,"Deep Reinforcement Learning-Augmented Transformer Framework for Robust Multi-Class Sentiment Analysis of Social Media Streams" International Journal of Advanced Multidisciplinary Research and Educational Development, V2(3): Page(575-580) May-June 2026. ISSN: 3107-6513. www.ijamred.com. Published by Scientific and Academic Research Publishing.