MT Error Detection for Cross-Lingual Question Answering

Abstract

We present a novel algorithm for detecting errors in MT, specifically focusing on content words that are deleted during MT. We evaluate it in the context of cross-lingual question answering (CLQA), where we try to correct the detected errors by using a better (but slower) MT system to retranslate a limited number of sentences at query time. Using a query-dependent ranking heuristic enabled the system to direct scarce MT resources towards retranslating the sentences that were most likely to benefit CLQA. The error detection algorithm