Amazon product reviews in four categories: books, dvd, electronics, and kitchen & housewares.
1000 positive, 1000 negative and various unlabeled reviews per category.
Data is available here.

├── sorted_data_acl/
    ├── books/
    │   ├── negative.review
    │   ├── positive.review
    │   ├── unlabeled.review
    ├── dvd/
    │   ├── negative.review
    │   ├── positive.review
    │   ├── unlabeled.review
    ├── electronics/
    │   ├── negative.review
    │   ├── positive.review
    │   ├── unlabeled.review
    |── kitchen_&_houswares/
        ├── negative.review
        ├── positive.review
        ├── unlabeled.review

This will create a reviews_forEmbedding.txt file in each category folder. The file will contain all reviews (positive, negative and unlabeled) of that categories with one sentence of a review per line. The sentences do not contain any special characters or any punctuation.

This will merge all the above files into one file and store them in the sorted_data_acl/all/ folder.

This will create word embeddings for each category (including all) of the reviews using GloVe. In particular, this creates the following 4 files in each category folder:

• reviews.vocab: word count per category in the format word -> count
• reviews.cooccur: cooccurance matrix of words
• reviews.cooccur.shuf: sorted cooccurence matrix
• reviews.vectors.txt: word embeddings per category in the format word -> vector

This will create Python dictionaries in the format word -> vector from the files reviews.vectors.txt and store it in the files reviews.vectors.pkl.

This will create a reviews_positive.txt, ratings_positive.txt, reviews_negative.txt and ratings_revative.txt files in each category folder. The files will contain the respective reviews and ratings with one review/rating per line.

This will merge the preprocessed reviews from all four categories into the all/ folder.

This will transfrom the text reviews into embedded reviews by converting each word into a vector using the dictionaries from previous steps. The resulting matrices will be stores in reviews_positive.npy and reviews_negative.npy.

This will train a neural network to classify the sentiments in each category.