Søgaard, A., Ruder, S., & Vulić, I. (2018). On the limitations of unsupervised bilingual dictionary induction. Proceedings of ACL 2018, 778–788.
Read in full in NOTE-1337.
Why this is here
A5’s counterweight to LIT-1818: structural alignment recovers correspondence only where the two structures are alike, so a good fit is not established identity. It bears on CLAIM-135 and on the identification threat in CASE-185. One correction to that description: the paper tests Conneau et al.’s adversarial method, not GW, so it is a counterweight to LIT-1818 by shared premise, not by direct test. It adds that the likeness of two structures depends on how they were built: two spaces of the same language pair trained by different algorithms do not align at all (THEORY-1145).