Evaluation of Distributed Text Representations for Document Classification: A Comparative Study of Word2Vec, Doc2Vec, and Phrase Regression

This paper presents a comprehensive comparison of various text representation methods, including Word2Vec, Doc2Vec, Phrase Regression, and Multinomial Inverse Regression (MNIR), applied to the task of classifying Yelp reviews based on their star ratings. I explore several approaches, including Word2Vec inversion and aggregation, Doc2Vec with distributed memory (DM) and distributed bag-of-words (DBOW) specifications, and phrase-based logistic regression. Through an evaluation of out-of-sample misclassification rates, I find that simple phrase regression consistently performs well across most tasks, though Word2Vec inversion yields more accurate and interpretable document rankings in terms of fitted probabilities. The analysis highlights the trade-offs between discriminative and generative classifiers and suggests that while phrase regression is optimized for performance, distributed representations like Word2Vec and Doc2Vec could benefit from further optimization, particularly for large-scale unlabeled corpora. This work underscores the value of Word2Vec inversion as an interpretable and scalable alternative for text classification tasks, while also noting areas where more sophisticated document representation methods may improve with enhanced tuning and model refinement.