OTTAWA: Optimal TransporT Adaptive Word Aligner for Hallucination and Omission Translation Errors Detection

Recently, there has been considerable attention on detecting hallucinationsand omissions in Machine Translation (MT) systems. The two dominant approachesto tackle this task involve analyzing the MT system's internal states orrelying on the output of external tools, such as sentence similarity or MTquality estimators. In this work, we introduce OTTAWA, a novel OptimalTransport (OT)-based word aligner specifically designed to enhance thedetection of hallucinations and omissions in MT systems. Our approachexplicitly models the missing alignments by introducing a "null" vector, forwhich we propose a novel one-side constrained OT setting to allow an adaptivenull alignment. Our approach yields competitive results compared tostate-of-the-art methods across 18 language pairs on the HalOmi benchmark. Inaddition, it shows promising features, such as the ability to distinguishbetween both error types and perform word-level detection without accessing theMT system's internal states.