This paper studies the parameter sensitivity of term frequency normalization in the probabilistic model for information retrieval. High parameter sensitivity indicates that a slight change of the parameter value may considerably affect the retrieval performance. Therefore, a weighting model with a high parameter sensitivity is not robust enough to provide a consistent retrieval performance across different collections and queries. We suggest that the parameter sensitivity is due to the fact that the query term weights are not adequate enough to allow informative query terms to differ from non-informative ones. We show that query term reweighing, which is part of the relevance feedback process, can be successfully used to reduce the parameter sensitivity. Experiments on five Text REtrieval Conference (TREC) collections show that the parameter sensitivity does remarkably decrease when query terms are reweighed.
The authors examine the issue of parameter sensitivity in probabilistic information‑retrieval models such as BM25 and PL2. High sensitivity means that small changes in normalisation parameters can cause large variations in retrieval performance, making the models less robust across collections and query types.
They propose using query‑term reweighing, a component of blind relevance feedback, to mitigate this problem. Experiments on five TREC collections demonstrate that reweighing reduces both Entropy and Spread measures of sensitivity and improves cross‑collection parameter stability.
It refers to the situation where a slight change in a normalisation parameter causes a considerable change in retrieval performance, indicating the model is not robust.