The value of parsing as feature generation for gene mention recognition

We measured the extent to which information surrounding a base noun phrase reflects the presence of a gene name, and evaluated seven different parsers in their ability to provide information for that purpose. Using the GENETAG corpus as a gold standard, we performed machine learning to recognize from its context when a base noun phrase contained a gene name. Starting with the best lexical features, we assessed the gain of adding dependency or dependency-like relations from a full sentence parse. Features derived from parsers improved performance in this partial gene mention recognition task by a small but statistically significant amount. There were virtually no differences between parsers in these experiments.