Reproduce “BET: A Backtranslation Approach for Easy Data Augmentation in Transformer-based Paraphrase Identification Context” 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

Jean-Philippe Corbeil, Hadi Abdi Ghadivel

The authors introduce BET, which expands labeled paraphrase data by translating one sentence in a pair into another language and back, then retaining the label. They test it with Google Translate and transformer classifiers. On MRPC, RoBERTa with Vietnamese backtranslation raises F1 from 0.906 to 0.915; the result is limited to this setting and is not compared with other augmentation methods.

Key insights

  • Backtranslation can help, but the best pivot language varies by model and dataset; the all-language combination is not consistently better than a single language.
  • Some augmentation runs increase recall while reducing precision, indicating a trade-off between finding more paraphrases and producing more false positives.
  • The low-data results include large gains as well as failures, and should be treated cautiously because experiments use a single random seed and do not report uncertainty.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.