example, if a year-founded string is in the first row, there is nothing to prevent the string macchiatone from appearing beneath it. Any useful application making use of Web data must also be able to address uncertain data design and quality.

Vast number of topics. A tradi-
tional database typically focuses on
a particular domain (such as prod-
ucts or proteins) and therefore can
be modeled in a coherent schema. On
the Web, data covers everything, and
is also one of its appeals. The breadth
and cultural variations of data on the
Web make it inconceivable that any

figure 1. typical use of the table tag to describe relational data that has structure never explicitly declared by the author, including metadata consisting of several typed and labeled columns, but that is obvious to human observers. the navigation bars at the top of the page are also implemented through the table tag but do not contain relational-style data.

figure 2. Results of a keyword query search for “city population,” returning a relevance-ranked list of databases. the top result contains a row for each of the most populous 125 cities and columns for “city/urban Area,” “country,” “Population,” and “rank” (by population among all the cities in the world). the system automatically generated the image at right, showing the result of clicking on the “Paris” row. the title (“city mayors…”) links to the page where the original HtmL table is located.

manual effort would be able to create a clean model of all of it.

Before addressing the challenges associated with accessing structured data on the Web, it is important to ask what users might do with such data. Our work is inspired by the following example benefits:

Improve Web search. Structured Web data can help improve Web search in a number of ways; for example, Deep Web databases are not generally available to search engines, and, by surfacing this data, a Deep Web exploration tool can expand the scope and quality of the Web-search index. Moreover, the layout structure can be used as a relevance signal to the search ranker; for example, an HTML table-embedded database with a column calories and a row latte, should be ranked fairly high in response to the user query latte calories. Traditionally, search engines use the proximity of terms on a page as a signal of relatedness; in this case, the two terms are highly related, even though they may be distant from each other on the page.

Enable question answering. A long-standing goal for Web search is to return answers in the form of facts; for example, in the latte calories query, rather than return a URL a search engine might return an actual numerical value extracted from the HTML table. Web search engines return actual answers for very specific query domains (such as weather and flight conditions), but doing so in a domain-independent way is a much greater challenge.

Enable data integration from multiple Web sources. With all the data sets available on the Web, the idea of combining and integrating them in ad hoc ways is immensely appealing. In a traditional database setting, this task is called data integration; on the Web, combining two disparate data sets is often called a “mashup.” While a traditional database administrator might integrate two employee databases with great precision and at great cost, most combinations of Web data should be akin to Web search— relatively imprecise and inexpensive; for example, a user might combine the set of coffeehouses with a database of WiFi hotspots, where speed