A memory-based classification approach to marker-based EBMT

We describe a novel approach to examplebased machine translation that makes use of marker-based chunks, in which the decoder is a memory-based classifier. The classifier is trained to map trigrams of source-language chunks onto trigrams of target-language chunks; then, in a second decoding step, the predicted trigrams are rearranged according to their overlap. We present the first results of this method on a Dutch-to-English translation system using Europarl data. Sparseness of the class space causes the results to lag behind a baseline phrase-based SMT system. In a further comparison, we also apply the method to a word-aligned version of the same data, and report a smaller difference with a word-based SMT system. We explore the scaling abilities of the memory-based approach, and observe linear scaling behavior in training and classification speed and memory costs, and loglinear BLEU improvements in the amount of training examples.