Detecting Translation Hallucinations with Attention Misalignment

A method is proposed for detecting hallucinations in neural machine translation by leveraging attention misalignment. This technique uses both a forward (source-to-target) and a backward (target-to-source) translation model to get interpretable uncertainty signals. After generating a translation, the backward model uses teacher forcing to create a cross-attention map, which is then compared to the forward model's map. Discrepancies and low reciprocal attention scores between these two maps indicate potential errors or hallucinations at the token level, offering a computationally efficient alternative to other quality estimation methods.