You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Translating training data into many languages has emerged as a practicalsolution for improving cross-lingual transfer. For tasks that involvespan-level annotations, such as information extraction or question answering,an additional label projection step is required to map annotated spans onto thetranslated texts. Recently, a few efforts have utilized a simplemark-then-translate method to jointly perform translation and projection byinserting special markers around the labeled spans in the original sentence.However, as far as we are aware, no empirical analysis has been conducted onhow this approach compares to traditional annotation projection based on wordalignment. In this paper, we present an extensive empirical study across 57languages and three tasks (QA, NER, and Event Extraction) to evaluate theeffectiveness and limitations of both methods, filling an important gap in theliterature. Experimental results show that our optimized version ofmark-then-translate, which we call EasyProject, is easily applied to manylanguages and works surprisingly well, outperforming the more complex wordalignment-based methods. We analyze several key factors that affect theend-task performance, and show EasyProject works well because it can accuratelypreserve label span boundaries after translation. We will publicly release allour code and data.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)
The text was updated successfully, but these errors were encountered: