ParaCLEAN: Improving Translation Quality through Systematic Parallel Data Cleaning

Abstract

Parallel corpora often contain significant noise, particularly in low-resource settings where both collected and synthetic data are combined. We present ParaCLEAN, a modular pipeline for cleaning parallel data that integrates embeddings-based filtering, language identification, deduplication, and normalisation. Experiments on Catalan to Japanese translation demonstrate that ParaCLEAN improves data quality and downstream MT performance. Ablation studies highlight the contribution of each step. ParaCLEAN is lightweight, reproducible, and extensible for diverse language pairs.

Resources

Details