Analysis of an Advanced Sentiment Classification Model Using Sentiment140 Dataset with Semantic Text Normalization and Comparative Machine Learning

Abstract

Sentiment analysis of social media text is still an important but difficult problem in Natural Language Processing (NLP) because of the highly informal and noisy text nature. This paper proposes the design and evaluation of an advanced sentiment classification system designed for the Sentiment140 dataset which contains 1.6 million labeled tweets. The work introduces a complex pipeline for semantic text normalization including slang expansion, negation handling, URL/mention removal, and domain-specific tokenization before the feature extraction step. Three classical machine learning classifiers (Logistic Regression (LR), Support Vector Machine (SVM), Naive Bayes (NB)) and two ensemble methods (Random Forest (RF), Gradient Boosting (GB)) are considered in the experimental part. The BoW representation, TF-IDF weighting of unigrams and bigrams, and combination of word-level n-grams are used as the features in the experiments. The results show that the SVM classifier using TF-IDF bigram features produces the best classification accuracy of 87.63%, surpassing the baseline NB classifier by 9.4 percentage points. In general, the presented normalization pipeline consistently leads to better results with respect to all classifiers, with the average improvement being 4.2% in terms of F1-score. This research also includes an ablation study that focuses on the importance of each of the normalization steps and proves the advantage of using linear kernel SVM for solving problems related to high-dimensional sparse feature sets.