You are part of a machine learning team at a gaming company with millions of global customers. Each game features a real-time chat function that supports over 20 languages, with messages translated using the Cloud Translation API. The task is to build an ML system to moderate these chat messages instantly and ensure uniform performance across all languages, without altering the current serving infrastructure. Your initial model was trained with an in-house word2vec for embedding chat messages translated by the Cloud Translation API, but it showed significant performance variation across different languages. What should be your next step to improve the model's performance consistency across all languages?