Large Movie Review Dataset v1.0

Dataset Overview

Dataset Name

Large Movie Review Dataset v1.0

Dataset Purpose

The dataset is used as a benchmark for sentiment classification, containing movie reviews with binary sentiment polarity labels.

Dataset Content

Dataset Structure

• Core Dataset: Contains 50,000 reviews, split into 25,000 training and 25,000 test samples with balanced labels (25 k positive and 25 k negative).
• Unsupervised Learning Dataset: An additional 50,000 unlabeled documents.

Dataset Characteristics

• A maximum of 30 reviews per movie to avoid inter‑review correlation.
• No overlap of movies between training and test sets, preventing performance gains from memorizing movie‑specific terminology and label associations.
• In the labeled training/test sets, negative reviews have scores ≤ 4/10, positive reviews have scores ≥ 7/10.
• In the unsupervised set, reviews of all scores are included, with equal numbers of reviews scoring above 5 and at most 5.

File Structure

Organization

• Training and Test Sets: Correspond to [train/] and [test/] directories.
• Label Categories: Each directory contains [pos/] and [neg/] subdirectories for positive and negative reviews.
• File Naming: Review text files follow the [[id]_[rating].txt] convention, where [id] is a unique identifier and [rating] is a 1‑10 star rating.
• URL Files: Contain IMDb URLs for each review, formatted as [urls_[pos, neg, unsup].txt].
• Feature Files: Provide tokenized bag‑of‑words features stored as .feat files in LIBSVM format.
• Vocabulary: The [imdb.vocab] file stores text tokens corresponding to feature indices.
• Expected Rating File: [imdbEr.txt] contains the expected rating for each token in [imdb.vocab].

Citation Information

• When using this dataset, please cite the ACL 2011 conference paper that introduced the dataset and provided baseline classification results for comparison.