Context-Sensitive Retrieval for Example-Based Translation.pdf

Context-Sensitive Retrieval for Example-Based Translation Ralf D. Brown Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213 ralf+@ Abstract Example-Based Machine Translation (EBMT) sys- tems have typically operated on individual sen- tences without taking into account prior context. By adding a simple reweighting of retrieved fragments of training examples on the basis of whether the previous translation retrieved any fragments from examples within a small window of the current in- stance, translation performance is improved. A fur- ther improvement is seen by performing a similar reweighting when another fragment of the current input sentence was retrieved from the same training example. Together, a simple, straightforward imple- mentation of these two factors results in an improve- ment on the order of 1.0–1.6% in the BLEU metric across multiple data sets in multiple languages. 1 Introduction While context has long been recognized as an im- portant factor in translating texts, it tends to be given lower priority in machine translation system devel- opment than improving the quality of isolated trans- lations. Quality can only be improved so far, how- ever, when operating strictly on isolated sentences, and thus further improvements must eventually be sought by taking other sentences into account when performing a translation. EBMT systems typically treat both training data and the input to be translated as bags of unrelated sentences, though in practice, consecutive sentences are in fact related. Rather than consisting of random sentences, the training data consists of a set of co- herent documents, and the input to be translated is one or more documents. In particular, retrieval is done without regard to the results of the prior sen- tence’s translation, and thus differing word senses receive equal weighting. In contrast, by considering whether the previous sentence that was translated used adjacent sentences in the training corpus, th