The Importance of Focused Evaluations: a Case Study of TREC and DUC

Donna K. Harman

NIST


Abstract

Evaluation has always been an important part of scientific research, and in information retrieval, this evaluation has mostly been done using test collections. In 1992 a new test collection was built at the National Institute of Standards and Technology (NIST), and a focused evaluation (the Text REtrieval Conference or TREC) was started to use this collection. Results from nearly 10 years of this focused evaluation show significant technology transfer across systems, leading to major improvements in system performance. Focused evaluations also create the ability to target specific problems in language technology, such as retrieval across languages, and design tasks for evaluation such that issues can be studied concurrently by multiple groups. This talk will discuss some of the tasks that have been examined in TREC and what was learned during those tasks. Additionally a new focused evaluation, the Document Understanding Conference (DUC) which will examine text summarization, will be presented.