Preprocessing Methods for Word Alignment

This paper compares four preprocessing approaches for word alignment: 1) sentence removal approach, 2) good points approach, 3) sentence duplication approach, and 4) removal of doubtful alignments approach. Two are statistically motivated and the other two are heuristics. We focus on the ability of a word aligner of IBM Model 4 that it should often face with troubles when handling paraphrase, multi-words and non-literal translation. We assume that IBM Model 4 works 90% correct, while only around 5% wrong.