Using key phrases as new queries in building relevance judgements automatically

Abstract

We describe a new technique for building a relevance judgment list (qrels) for TREC test collections with no human intervention. For each TREC topic, a set of new queries is automatically generated from key phrases extract-ed from the top k documents retrieved from 12 different Terrier weighting models when the initial TREC topic is submitted. We assign a score to each key phrase based on its similarity to the original TREC topic. The key phrases with the highest scores become the new queries for a second search, this time using the Terrier BM25 weighting model. The union of the documents retrieved forms the automatically-build set of qrels.