ParaCycle

Reinforcement learning framework for reference-free machine translation using bidirectional paraphrase consensus

Master’s Thesis - Johns Hopkins University

ParaCycle is a novel reinforcement learning framework for low-resource machine translation that eliminates dependency on parallel corpora. By using bidirectional paraphrase consensus, the system achieves improved translation quality across multiple language pairs.

Key Contributions

  • Semantic Consistency Rewards: Designed customized RL objectives for unsupervised quality estimation
  • Paraphrase Consensus: Leveraged bidirectional paraphrasing to establish translation quality without reference translations
  • Low-Resource MT: Validated on four English↔X language pairs using FLORES-200 benchmark
  • RL Formulation: Formulated translation quality optimization as a reinforcement learning problem

Technical Details

Framework: PyTorch, Hugging Face Transformers, Reinforcement Learning (SFT, DPO, GRPO)

Evaluation: FLORES-200 benchmark, multiple language pairs

Advisor: Prof. Philipp Koehn, Center for Language and Speech Processing (CLSP), JHU

Impact

This research addresses the critical challenge of machine translation quality estimation in low-resource scenarios where parallel corpora are unavailable or limited. The paraphrase consensus approach provides a reference-free alternative that maintains translation quality while reducing data dependency.

Status: Master’s thesis (Expected completion: December 2025)

References