... lexicalize the model. Sentences starttheir life in short form, s, are ranked by a source language model, p(s), and then probabilistically ex-panded to form the long sentence, p(l |s ). Duringdecoding, ... of them for such compres-sions/expansions. We make the simplifying assump-tion that all such edits also retain the core mean-ing of the sentence, and are therefore valid trainingdata for our ... Volume), pages 137–140,Columbus, Ohio, USA, June 2008.c2008 Association for Computational LinguisticsMining Wikipedia Revision Histories for Improving Sentence CompressionElif Yamangil Rani NelkenSchool...