We deal with the task of generating a query that retrieves a given set of documents. In its abstract form, this can be seen as a "compression" of the document set to a short query. But the task also has a real-world application: cluster labeling (e.g., for faceted search). Our solution to cluster labeling is the usage of queries that approximately retrieve a cluster's documents. To be generalizable, our approach does not require access to a search index but only a public interface like an API. This way, our approach can also be implemented at client side. In an experimental evaluation, a basic version of our approach using a simple retrieval model is on par with standard cluster labeling techniques. A further user study reveals that queries as labels are often preferred when they are not too long.
展开▼
