Learning Word Vectors for Sentiment Analysis: A Python Reproduction

A Python reproduction of the Maas et al. (2011) paper on learning word vectors for sentiment analysis using IMDb reviews. The post walks through building a vocabulary from 75,000 reviews, implementing a probabilistic semantic model with MAP estimation, adding a supervised sentiment objective using star ratings, and evaluating learned representations with a linear SVM classifier. Four document representations are compared: Bag of Words baseline, semantic-only word vectors, full semantic+sentiment vectors, and a combined dense+sparse representation. Results closely match the original paper, demonstrating how unlabeled data can learn semantic structure while labeled ratings inject sentiment polarity into the same vector space.