Harnessing the Deep Web: Present and Future

Over the past few years, we have built a system that has exposed large volumes of Deep-Web content to Google.com users. The content that our system exposes contributes to more than 1000 search queries per-second and spans over 50 languages and hundreds of domains. The Deep Web has long been acknowledged to be a major source of structured data on the web, and hence accessing Deep-Web content has long been a problem of interest in the data management community. In this paper, we report on where we believe the Deep Web provides value and where it does not. We contrast two very different approaches to exposing Deep-Web content -- the surfacing approach that we used, and the virtual integration approach that has often been pursued in the data management literature. We emphasize where the values of each of the two approaches lie and caution against potential pitfalls. We outline important areas of future research and, in particular, emphasize the value that can be derived from analyzing large collections of potentially disparate structured data on the web.