2000 character limit reached
Debugging Neural Machine Translations (1808.02733v1)
Published 8 Aug 2018 in cs.CL
Abstract: In this paper, we describe a tool for debugging the output and attention weights of neural machine translation (NMT) systems and for improved estimations of confidence about the output based on the attention. The purpose of the tool is to help researchers and developers find weak and faulty example translations that their NMT systems produce without the need for reference translations. Our tool also includes an option to directly compare translation outputs from two different NMT engines or experiments. In addition, we present a demo website of our tool with examples of good and bad translations: http://attention.lielakeda.lv
Collections
Sign up for free to add this paper to one or more collections.
