This seems like it would screw up phrase queries? An alternate fix would be to add Lucene's stopword list to our parser and submit the appropriately modified phrase query.