Core Concepts
The study explores the alignment between online and offline evaluations in search clarification, revealing that engaging clarification questions can be accurately identified using both approaches.
Abstract
The study examines the effectiveness of clarification models by comparing online user engagement with offline evaluation labels. It highlights the importance of user engagement in interactive information retrieval. The research investigates various aspects of engagement, such as query length and uncertainty in online assessments. Results show that combining offline labels does not significantly improve model performance compared to individual labels.
Stats
Contrary to common belief, offline evaluations align with online evaluations in search clarification.
Engagement Level is constructed based on click-through rate of real user interactions with clarification panes.
The dataset used contains 1,034 query-clarification pairs for analysis.
LTR models incorporating offline labels do not outperform individual offline labels in ranking engaging clarification questions.
SVM-rank showed better performance among LTR models but was not significantly different from others.