From: Comeau, Don (NIH/NLM/NCBI) [E]
Sent: Tuesday, March 04, 2008 10:13 AM
To: NLM/NCBI List ncbi-seminar
Subject: CBB seminar, Tuesday 3/4/2008, 11 am, B2 Library

CBB seminar, Tuesday 3/4/2008, 11 am, B2 Library

Don Comeau

What Are You Looking For?  PubMed Queries and Evaluating Retrieval with
Machine Learning

Examining sample PubMed queries shows that they usually contain known
phrases rather than merely series of words.  To determine whether this
observation could be used to improve retrieval quality, we turned to
machine learning rather than human experts.  Bayesian machine learning
can clearly distinguish between documents with a particular phrase in
contrast to documents simply containing the words of the phrase, as
retrieved by a Boolean AND of the query terms.  It also showed that
documents containing the query terms in a single sentence are more like
those that contain the exact phrase, than documents that only contain
the query terms.