Reading to Translate or Translating to Read? Modeling Translators' Eye Movements with Multilingual Pre-Trained Models.

Abstract

Translation and post-editing both integrate reading into bilingual text production, yet it remains unclear which computational predictors from multilingual pre-trained models best account for translators' reading patterns across task types and translation directions. We recruited twenty-six Chinese L1 translators who completed en→zh and zh→en translation and post-editing tasks, yielding 104 eye-tracking sessions. Dependent measures were source reading time (TrtS), target reading time (TrtT), and target production duration (Dur). Predictors were derived from two model architectures, a decoder-only language model (LM) and an encoder-decoder neural machine translation (NMT) model, and they included monolingual surprisal, translation surprisal with source context, and attention features computed from models' internal weights. Analyses showed that LM surprisal provided the strongest account of target reading, while source reading was most strongly predicted by encoder self-attention with LM surprisal, a robust secondary predictor, and target production duration drew on both LM and NMT translation surprisal. Direction effects were broader than task effects, especially on target measures. These findings suggest that although translation reading is bilingual in task structure, cumulative reading time is best explained by monolingual LM surprisal, whereas production duration additionally reflects NMT translation surprisal and revision behavior.