Analyzing Various Machine Learning Algorithms for Opinion Extraction from Web Text Using AI Across Multiple Datasets

Abstract

Opinion extraction from web text is essential for understanding public attitudes in e-commerce, news, and social media, yet it remains challenging due to noisy language, short informal messages, and inconsistent sentiment labels. This study proposes a unified AI-driven pipeline for three-class sentiment classification (positive, neutral, negative) across multiple web-text domains. The workflow performs label normalization, missing-value removal, de-duplication, and text cleaning (URL/mention removal, hashtag normalization, and whitespace standardization). Cleaned text is represented using TF–IDF with unigram and bigram features and evaluated using twelve classic machine learning classifiers, with a focus on LinearSVM and Calibrated LinearSVM for robust discrimination and probability-based analysis. Experiments are conducted on three datasets: product reviews, Times of India headlines, and English political tweets. Performance is assessed using accuracy, precision, recall, F1-score, confusion matrices, and OvR ROC/precision–recall curves. On the Times of India dataset, LinearSVM achieves the best accuracy of 0.894, while Calibrated LinearSVM attains a comparable accuracy of 0.893, demonstrating strong and consistent performance for headline sentiment classification. The results indicate that TF–IDF combined with linear margin-based models provides an effective and scalable baseline for multi-domain opinion extraction.