Most state-of-the-art statistical machine translation systems use log-linear models, which are defined in terms of hypothesis features and weights for those features. It is standard to tune the feature weights in order to maximize a translation quality metric, using held-out test sentences and their corresponding reference translations. However, obtaining reference translations is expensive. In this paper, we introduce a new full-sentence paraphrase technique, based on English-to-English decoding with an MT system, and we demonstrate that the resulting paraphrases can be used to drastically reduce the number of human reference translations needed for parameter tuning, without a significant decrease in translation quality. 1

Most state-of-the-art statistical machine translation systems use log-linear models, which are defined in terms of hypothesis features and weights for those features. It is standard to tune the feature weights in order to maximize a translation quality metric, using heldout test sentences and their corresponding reference translations. However, obtaining reference translations is expensive. In our earlier work (Madnani et al., 2007), we introduced a new full-sentence paraphrase technique, based on English-to-English decoding with an MT system, and demonstrated that the resulting paraphrases can be used to cut the number of human reference translations needed in half. In this paper, we take the idea a step further, asking how far it is possible to get with just a single good reference translation for each item in the development set. Our analysis suggests that it is necessary to invest in four or more human translations in order to significantly improve on a single translation augmented by monolingual paraphrases. 1

In this paper, we discuss how we apply automatically generated semantic knowledge to benefit statistical machine translation (SMT). Currently, almost all statistical machine translation systems rely heavily on memorizing translations of phrases. Some systems attempt to go further and generalize these learned phrase translations into templates using empirically derived information about word alignments and a small amount of syntactic information, if at all. There are several issues in a SMT pipeline that could be addressed by the application of semantic knowledge, if such knowledge were easily available. One such issue, an important one, is that of reference sparsity. The fundamental problem that translation systems have to face is that there is no such thing as the correct translation for any sentence. In fact, any given source sentence can often be translated into the target language in many valid ways. Since there can be many “correct answers, ” almost all models employed by SMT systems require, in addition to a large bitext, a held-out development set comprised of multiple high-quality, human-authored reference translations in the target language in order to tune their parameters relative to a translation quality metric. 1 There are several reasons that this requirement is not an easy one to satisfy. First, with a few exceptions—notably NIST’s annual MT evaluations—most new MT research data sets are provided with only a single reference translation. Second, obtaining multiple reference translations in rapid development, low-density source language scenarios (e.g. (Oard, 2003)) is