Search results are usually ranked lists of documents relevant to query terms. In this paper, the entity search results are bundled with those beyond the query term, by constructing the entity network where extracted entities and pairwise entity relations are augmented with metadata. The authors consider six bundling algorithms: (1) by super-categories of the query entity; (2) by super-categories of the result entities; (3) by the entity’s subtopics or specializations; (4) by sentiment; (5) by categories followed by sentiments; and (6) by sentiment and categories.

To test whether the bundled query results help the user, three large-scale user studies using Amazon Mechanical Turk were conducted. First, the six bundled query results were compared with the basic ranked list results; the top three bundling algorithms that are significantly better than the baseline include bundling by query categories, by result categories, and by specialization. Second, these three results were compared with each other to identify the best bundling algorithm: bundling by query entity categories. In a third experiment, the users were asked to perform a task using the best bundling result sets versus the ranked list result set. After the search, a user satisfaction survey was conducted, in addition to measuring the number of clicks and time it took to complete the task. The authors conclude that the explorative entity search is effective and promotes engagement.

The study results are interesting, but the bundled queries are heavily dependent on the entity network model. Unfortunately, the entity network construction is not detailed. It relies on manually curated parts, such as the three categories, which is not scalable. The typical precision and recall analyses of these bundled queries are also missing. The engagement-related survey questions turn out to be insignificant. Consequently, the research question on whether the bundling approach supports exploration and engagement needs further investigation.