3.6.2: Pairs of Terms

Many systems used pairs of terms for indexing. The idea is that such pairs should be more specific then single terms. For example SMART allowed in TREC 3 term pairs for indexing if they occurred in more than 25 documents. In the pseudo relevance feedback step the 10 most frequent term pairs were used in the same way as the 500 most frequent single terms.

The "INQUERY" system extracted "phrases" of two or three words from the documents of a large sub collection and added the most similar phrases to the query.

The use of term pairs seems to have only little or even no positive effect on the results. In TREC 4 several systems reduced the number of pairs used in their queries.


© 1998 / HTML-Version 17. 11. 1998: R. Ferber