NEWSLETTER
Automatic translation (MT) is experiencing a paradigm shift, with systems based on large -tuning language models (LLM) that become increasingly competitive with traditional models of decoder of encoders trained specifically for translation tasks. However, LLM -based systems have an increased risk of generating hallucinations, which can severely undermine user trust and safety. Most of the previous research on hallucination mitigation focus on traditional MT models, with solutions that involve post-hoc mitigation, detect hallucinated translations and translate them again. While it is effective, this approach introduces additional complexity in the implementation of additional tools in production and also increases latency. To address these limitations, we propose a method that intrinsically learns to mitigate hallucinations during the model training phase. Specifically, we present a data creation frame to generate preferences data sets in hallucination. The LLM adjustment in these preferences data sets reduces the hallucination rate by an average of 96% in five language pairs, while preserving the general quality of the translation. In a zero shooting configuration, our approach reduces hallucinations by 89% in an average in three invisible target languages.
