TNO-UT at TREC-9: How Different are Web Documents?

Although at first sight, the web track might seem a copy of the ad hoc track, we discovered that some small adjustments had to be made to our systems to run the web evaluation. As we expected, the basic language model based IR model worked effectively on this data. Blind feedback methods however, seem less effective on web data. We also experimented with rescoring the documents based on several algorithms that exploit link information. These methods yielded no positive result.