N-gram measures of translation quality, such as BLEU and the related NIST metric, are becoming increasingly important in machine translation, yet their behaviors are not fully understood. In this paper we examine the performance of these metrics on professional human translations into German of two literary genres, the Bible and Tom Sawyer. The most surprising result is that some machine translations outscore some professional human translations. In addition, it can be difficult to distinguish some other human translations from machine translations with only two reference translations; with four reference translations it is much easier. Our results lead us to conclude that much care must be taken in using n-gram measures in formal evaluations of machine translation quality, though they are still valuable as part of the iterative development cycle.

Current alternatives for language modeling are statistical techniques based on large amounts of training data, and hand-crafted context-free or finite-state grammars that are difficult to build and maintain. One way to address the problems of the grammar-based approach is to compile recognition grammars from grammars written in a more expressive formalism. While theoretically straight-forward, the compilation process can exceed memory and time bounds, and might not always result in accurate and efficient speech recognition. We will describe and evaluate two approaches to this compilation problem. We will also describe and evaluate additional techniques to reduce the structural ambiguity of the language model.