Reproduce “The Multilingual Amazon Reviews Corpus” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Phillip Keung, Yichao Lu, György Szarvas, Noah A. Smith
Researchers at Amazon, the University of Washington, and the Allen Institute for AI introduce the Multilingual Amazon Reviews Corpus (MARC), a large-scale, publicly available dataset comprising 300,000 customer reviews for each of six languages, designed for multilingual text classification. The corpus features balanced five-star ratings and standardized data splits, with multilingual BERT baselines demonstrating consistent performance improvements when incorporating review titles and product categories.
Key insights
- MARC addresses a critical data gap, providing an extensive, high-quality, and reproducible resource that enables robust benchmarking and accelerates research in multilingual NLP.
- Mean Absolute Error (MAE) is a more appropriate evaluation metric than classification accuracy for ordinal tasks like star rating prediction, as it accounts for the magnitude of prediction error.
- Incorporating additional metadata, specifically review titles and product categories, consistently improves model performance across all languages and classification tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.