Title: Learning by asking: a word-level based active strategy for interactive-predictive neural machine translation In spite of the huge advancement in NMT, MT outputs are far from commercial usage without post-editing. Human in the loop in neural machine translation, aka. interactive neural machine translation, was proposed to reduce human post editing effort. Kreutzer, J. et al. 2017 proposes to improve neural structured prediction using weak feedback, e.g. rating, which is then combined in an interactive-predictive framework (Lam et al. 2018) for human effort reduction. However, weak feedback based on rating or sentence scores can be inconsistent (Kreutzer et al. 2018). In this colloquium, the author proposes to utilise basic editing operations, e.g., word substitution and word rejection, on critical locations suggested by the translation system for effort reduction and, at the same time, aiming at improving both model performance and the quality of edited translations. In addition, the proposed method is combined with two main types of interactive NMT systems: (1) full-sentence based and (2) left-to-right partial translation system.