Localization of Difficult-to-Translate Phrases

Proceedings of the Second Workshop on Statistical Machine Translation, pages 248–255, Prague, June 2007. c⃝2007 Association for Computational Linguistics Localization of Difficult-to-Translate Phrases Behrang Mohit1 and Rebecca Hwa1,2 Intelligent Systems Program1 Department of Computer Science2 University of Pittsburgh Pittsburgh, PA 15260 U.S.A. {behrang, hwa}@cs.pitt.edu Abstract This paper studies the impact that difficult-to- translate source-language phrases might have on the machine translation process. We formu- late the notion of difficulty as a measurable quantity; we show that a classifier can be trained to predict whether a phrase might be difficult to translate; and we develop a frame- work that makes use of the classifier and ex- ternal resources (such as human translators) to improve the overall translation quality. Through experimental work, we verify that by isolating difficult-to-translate phrases and processing them as special cases, their nega- tive impact on the translation of the rest of the sentences can be reduced. 1 Introduction For translators, not all source sentences are created equal. Some are straight-forward enough to be automatically translated by a machine, while others may stump even professional human translators. Similarly, within a single sentence there may be some phrases that are more difficult to translate than others. The focus of this paper is on identify- ing Difficult-to-Translate Phrases (DTPs) within a source sentence and determining their impact on the translation process. We investigate three ques- tions: (1) how should we formalize the notion of difficulty as a measurable quantity over an appro- priately defined phrasal unit? (2) To what level of accuracy can we automatically identify DTPs? (3) To what extent do DTPs affect an MT system's performance on other (not-as-difficult) parts of the sentence? Conversely, would knowing the correct translation for the DTPs improve the system’s translation for the rest of the sentence? In this work, we model difficulty as a meas- urement with respect to a particular MT system. We further assume that the degree of difficulty of a phrase is directly correlated with the quality of the translation produced by the MT system, which can be approximated using an automatic evaluation metric, such as BLEU (Papineni et al., 2002). Us- ing this formulation of difficulty, we build a framework that augments an off-the-shelf phrase- based MT system with a DTP classifier that we developed. We explore the three questions in a set of experiments, using the framework as a testbed. In the first experiment, we verify that our pro- posed difficulty measurement is sensible. The sec- ond experiment evaluates the classifier's accuracy in predicting whether a source phrase is a DTP. For that, we train a binary SVM classifier via a series of lexical and system dependent features. The third is an oracle study in which the DTPs are perfectly identified and human translations are ob- tained. These human-translated phrases are then used to constrain the MT system as it translates the rest of the sentence. We evaluate the translation quality of the entire sentence and also the parts that are not translated by humans. Finally, the …

Content truncated to 500 words