We present a simple and effective way to perform out-of-domain statistical parsing by drastically reducing lexical data sparseness in a PCFG-LA architecture. We replace terminal symbols with unsupervised word clusters acquired from a large newspaper corpus augmented with biomedical target- domain data. The resulting clusters are effective in bridging the lexical gap between source-domain and target-domain vocabularies. Our experiments combine known self-training techniques with unsupervised word clustering and produce promising results, achieving an error reduction of 21% on a new evaluation set for biomedical text with manual bracketing annotations.