Deep Learning-based Multi-Class Sentiment Classification from Social Media Comments using LSTM Architecture

Abstract

This paper presents a lightweight sentiment classification model based on Long Short-Term Memory (LSTM) networks, developed as a foundational text-analysis component for future multimodal emotion recognition systems. The study uses a multi-platform English social media dataset comprising 526 comments collected from Twitter, Instagram, and Facebook. A synonym-based sentiment harmonization strategy was applied to consolidate semantically overlapping labels into five unified sentiment classes, improving label consistency in noisy user-generated data. Text preprocessing included lowercasing, punctuation removal, tokenization, and sequence padding before model training. Using an 80:20 train–test split, the proposed model achieved approximately 92% training accuracy and 88% test accuracy, with corresponding loss values of 0.25 and 0.35. The results indicate stable learning behaviour, with a macro F1-score of approximately 0.87–0.89 across sentiment classes. Rather than emphasizing architectural complexity, this study highlights the role of label harmonization and consistent preprocessing in enabling effective multi-class sentiment classification on small, real-world datasets. The proposed framework provides a reproducible and computationally efficient baseline suitable for integration into broader multimodal emotion recognition pipelines.