Translating Negation: Induction, Search And Model Errors
Federico Fancellu, Bonnie Webber · 2015
Statistical Machine Translation systems show considerably worse performance in translating negative sentences than positive ones (Fancellu and Webber, 2014;Wetzel and Bond, 2012).Various techniques have addressed the problem of translating negation, but their underlying assumptions have never been validated by a proper error analysis.A related paper (Fancellu and Webber, 2015) reports on a manual error analysis of the kinds of errors involved in translating negation.The present paper presents ongoing work to discover their causes by considering which, if any, are induction, search or model errors.We show that standard oracle decoding techniques provide little help due to the locality of negation scope and their reliance on a single reference.We are working to address these weaknesses using a chart analysis based on oracle hypotheses, guided by the negation elements contained in a source span and by how these elements are expected to be translated at each decoding step.Preliminary results show chart analysis is able to give a more in-depth analysis of the above errors and better explains the results of the manual analysis.