Predicting Retrieval Success Based on Information Use for Writing Tasks

摘要

This paper asks to what extent querying, clicking, and text editing behavior can predict the usefulness of the search results retrieved during essay writing. To render the usefulness of a search result directly observable for the first time in this context, we cast the writing task as "essay writing with text reuse," where text reuse serves as usefulness indicator. Based on 150 essays written by 12 writers using a search engine to find sources for reuse, while their querying, clicking, reuse, and text editing activities were recorded, we build linear regression models for the two indicators (1) number of words reused from clicked search results, and (2) number of times text is pasted, covering 69% (90%) of the variation. The three best predictors from both models cover 91-95% of the explained variation. By demonstrating that straightforward models can predict retrieval success, our study constitutes a first step towards incorporating usefulness signals in retrieval personalization for general writing tasks.