When conducting a search, a user typically reformulates the query after observing the results to maximize the likelihood of satisfying her information needs. An interesting research question is how to model the query term evolution and measure its effectiveness for the search. The authors of this paper propose an elegant and effective way of answering this question.

Each time the user issues a query, an “impression” is generated, which refers to the entire search data related to a query such as the ranked list of documents and the click-throughs. Elements of an impression include snippets (and their titles), clicks, dwell time, and documents. The authors model the query reformulation by defining term actions and measuring their effectiveness using various similarity measures for elements of impressions. The three term actions include retention, removal, and addition. The two similarity measures applied are cosine similarity and Jaccard similarity. They are used to measure the similarity between query at time n and query at time n + 1. The authors found that, on average, 63 percent of query term changes can be explained by retention and removal, and the remaining 37 percent are due to term addition. The added terms come from term sources, any elements of an impression such as snippets and clicked documents. The authors find high similarity between the query terms and the terms from the term sources, indicating the importance of the snippets and the clicked documents in query evolution.

The method proposed by the authors is very effective in modeling and measuring query reformulation in a search process. The method is intuitive, easy to understand, yet very innovative. The flow and the logic of the paper are very smooth, making it an enjoyable and enlightening read.