Open Internet by MindsNet

Addressing Metric-Model Affinity and Training Data Bias in Machine Translation

The evaluation of machine translation models revealed issues with metric-model affinity and training data bias, particularly in handling language codes and character scripts. TranslateGemma's performance was impacted by its training data being skewed toward Simplified Chinese, leading to incorrect outputs for Traditional Chinese. This highlights the need for more robust and diverse training data.

Computing & Technology, Computer Science, Machine Learning