What do post-editors correct? A fine-grained analysis of SMT and NMT errors

The recent improvements in neural MT (NMT) have driven a shift from statistical (SMT) to NMT. However, assess the usefulness of models for post-editing (PE) and detailed insight output they produce, we need analyse most frequent errors how affect task. We present pilot study fine-grained analysis based on post-editors corrections an English Spanish medical text translated with SMT use MQM taxonomy compare two categorized classification produced. Even though results show great variation among post-editors’ corrections, this language combination fewer are corrected by NMT output. also produces accuracy that less critical.