Data: IMDB movie reviews [provided by Kaggle]
Including 25,000 labeled review, 50,000 unlabeled review, 25,000 test set.
-
tf-idf feature model [bag of words]
Random Forest, accuracy: 0.94609
Naive Bayes + SGDClassifer, accuracy: 0.967
-
Word2Vec model [5000features_40minwords_10context]
Random Forest, accuracy: 0.93
After calculate the word embedding matrix, can go further by using K-Means so as to divide words into different categories (e.g. animal, divide, weather, etc). Represent each review with combination of categories, then train ML model.
-
Doc2Vec model [500features_1minwords_10context]
Logistic Regression, accuracy: 0.93169
Comparing with Word2Vec, the only differences is adding different specific document label to sentences in different reviews. Every paragraph is mapped to a unique vector, this paragraph token keeps the topic of the paragraph. At prediction, perform an inference step to compute the paragraph vector as the final feature.
