Machine Translationese: Effects of Algorithmic Bias on Linguistic Complexity in Machine Translation

Machine translation systems simplify language in ways that go beyond gender bias: The research found that algorithmic bias causes MT systems to lose lexical and morphological richness—meaning they use fewer unique words and simpler grammatical structures—across multiple translation approaches (phrase-based and neural MT) and language pairs (English-French and English-Spanish).

Bias amplification creates "machine translationese," an artificially impoverished form of language: By over-representing frequently occurring patterns in training data while dropping less common ones, MT systems produce translations that are linguistically flatter and less diverse than human translation, potentially degrading the quality and nuance of translated content over time.

The problem affects all current MT paradigms, not just newer neural approaches: Both traditional statistical machine translation and modern neural machine translation systems exhibited losses in linguistic richness, suggesting this is a systemic issue in how data-driven translation models learn rather than a problem unique to one technology.

This linguistic simplification could have real-world consequences beyond translation quality: When MT systems consistently produce simpler, less varied language, this could affect how information is communicated across languages in business, healthcare, legal, and educational contexts where precision and nuance matter.

Current MT training data may be the root cause: The research suggests that the datasets used to train these systems contain biased patterns that the algorithms learn to amplify, pointing to a need for more diverse and balanced training data as a potential solution.