A new approach to detecting neural machine translation hallucinations uses attention misalignment between forward and backward translation models. The method, described in a Towards Data Science article, leverages existing NMT infrastructure to provide interpretable uncertainty signals at the token level without requiring expensive retraining or massive computational overhead.

Traditional uncertainty estimation methods like entropy calculations or semantic clustering have significant limitations. They either fail to explain the nature of uncertainty or require computationally expensive processes like generating multiple outputs or fine-tuning massive models like XLM-R. The proposed method instead compares cross-attention maps from forward and backward translation models, identifying cases where the backward model cannot find a reverse translation path for hallucinated tokens.

Experiments with language pairs including French→English and Chinese→English demonstrate the approach's effectiveness across different language families. The method captures alignment quality even when patterns are less visually obvious, such as with typologically distant languages. By training a lightweight classifier on attention alignment features while keeping the main NMT model frozen, the technique provides a practical solution for identifying unreliable translations without the computational cost of existing quality estimation approaches.