The Recent ‘The Unreasonable Effectiveness of Data‘ Provides Important Hints

To even the most casual Web searcher, it must now be evident that Google is constantly introducing new structure into its search results. This past week three world-class computer scientists, all now research directors or scientists at Google, Alon Halevy, Peter Norvig and Fernando Pereira, published an opinion piece in the March/April 2009 issue of IEEE Intelligent Systems titled, ‘The Unreasonable Effectiveness of Data.’ It provides important framing and hints for what next may emerge in semantics from the Google search engine.

I had earlier covered Halevy and Google’s work on the deep Web. In this new piece, the authors describe the use of simple models working on very large amounts of data as means to trump fancier and more complicated algorithms.

Some of the research they cite is related to WebTables [1] and similar efforts to extract structure from Web-scale data. The authors describe the use of such systems to create ‘schemata’ of attributes related to various types of instance records — in essence, figuring out the structure of ABoxes [2], for leading instance types such as companies or automobiles [3].

These observations, which they call the semantic interpretation problem and contrast with the Semantic Web, they generalize as being amenable to a kind of simple, brute-force, Web-scale analysis: “Relying on overt statistics of words and word co-occurrences has the further advantage that we can estimate models in an amount of time proportional to available data and can often parallelize them easily. So, learning from the Web becomes naturally scalable.”

Google had earlier posted their 1 terabyte database of n-grams, and I tend to agree that such large-scale incidence mining can lead to tremendous insights and advantages. The authors also helpfully point out that certain scale thresholds occur for doing such analysis, such that researchers need not have access to indexes the scale of Google to do meaningful work or to make meaningful advances. (Good news for the rest of us!)

As the authors challenge:

1. Choose a representation that can use unsupervised learning on unlabeled data
2. Represent the data with a non-parametric model, and
3. Trust the important concepts will emerge from this analysis because human language has already evolved words for it.

My very strong suspicion is that we will see — and quickly — much more structured data for instance types (the ‘ABox’) rapidly emerge from Google in the coming weeks. They have the insights and approaches down, and clearly they have the data to drive the analysis! I also suspect many of these structured additions will just simply show up on the results listings to little fanfare.

The structured Web is growing all around us like stalagmites in a cave!

Topics Range from the Deep Web to Semantic Web in this Search Luminaries Series

I’m pleased to wrap up a multi-part interview with the Federated Search Blog as part of their ongoing ‘Search Luminaries’ series. Sol Lederman, editor of the blog, does a thorough and comprehensive job! Over the past month on every Friday, I have answered some 25 or so of his detailed questions.

Federated Search Blog was particularly interested in the deep Web, its discovery and size. Many of the early questions deal with those themes. However, by Part 4 things get a bit more current, with the topics shifting to the semantic Web, linked data and Zitgist.

Here are the links to the series:

• Preview (Oct. 17, 2008)
• Part I (Oct. 24)
• Part 2 (Oct. 31)
• Part 3 (Nov. 7)
• Part 4 (Nov. 14).

To give you a flavor of the interview, here is an example of one of the questions (and probably my favorite):

20. Tim Berners-Lee, credited with inventing the World Wide Web, has been talking about the importance and value of the Semantic Web for years yet common folks don't see much evidence of the Semantic Web gaining traction. Is there substance to the Semantic Web? What's happening with it now and what does its future look like?

Thanks again to Sol and Federated Search Blog for this opportunity.

Timelines, Semantics and Ontologies are Coming to the Fore

The past two weeks have seen an interesting emergence of new perspectives on the ‘deep Web‘. The deep Web, a term Thane Paulsen and I coined for my oft-quoted study from 2000, The Deep Web: Surfacing Hidden Value [1], is the phenomenon of database-backed content served from interactive Web search forms.

Because deep Web content is dynamic and produced only on request, it has been difficult for traditional search engines to index. It is also huge and of high quality (though likely not the 100x to 500x figure larger than the standard ‘surface’ Web that I used in that first study.)

Deep Web Timeline

This is the most recent of the three notable events over the past two weeks, and came out on Tuesday. Maureen Flynn-Burhoe of the oceanflynn @ Digg blog has produced a very informative and comprehensive timeline of deep Web and related developments from 1980 to the present (database-backed content and early Web precursors, of course, precede the Web itself and the term ‘deep Web’).

I have been directly involved in this field since 1994 and have not yet seen such a comprehensive treatment. She cites studies noting “hundreds of thousands” of deep Web sites and the faster growth of dynamic (database-served) as opposed to static (‘surface’) content on the Web.

As someone directly involved in estimating the size of the deep Web, I appreciate the analytic difficulties and take all of the estimates (my own older ones included!) with a grain of salt. Nonetheless, the deep Web is important, its content is huge, often of unique and high quality, and it deserves serious attention by Web scientists.

Great job, Maureen! I always appreciate thorough researchers. (BTW, I suspect you might also like the Timeline of Information History.)

Trends and Role in the Semantic Web

The next notable event was the publishing of Searching the Deep Web by Alex Wright in the Communications of the ACM (October 2008) [2]. Alex had first written about the deep Web for Salon magazine in 2004 and had given nice attention to my company at that time, BrightPlanet [3].

In this current update, Alex does an excellent job of characterizing current status and research in search techniques for the deep Web. I also liked the fact he used our fishing analogy of trawling for standard search crawlers versus direct angling in the deep Web (see our earlier figure at upper left).

As some may recall, Google has stepped up its activities in this area, an event I reported on a few months back. Those perspectives, and others from some other notable figures, are included in Alex’s piece as well.

My own contribution to the piece was to suggest that RDF and semantic Web approaches offered the next evolutionary stage in deep Web searching. Alex was able to take that theme and get some great perspectives on it. I also appreciate the accuracy of my quotes, which gives me confidence in the quality for the rest of the story.

Without a doubt there is high quality in the deep Web and bringing structure and semantic characterization to it through metadata is a task of some consequence.

For myself, I chose to move beyond the deep Web when its focus seemed stuck in a document-level perspective to retrieval and analysis. However, there is much to be learned from the techniques used to select and access deep Web content, which could be readily transferable to linked data.

Thanks, Alex, for making these prospects clearer! Maybe it is time to dust off some of my old stuff!

Getting Deeper into the Semantics

This emerging joining of deep Web and semantics is actually taking place through the efforts of a number of academic researchers. Recently and prominently has been James Geller from the New Jersey Institute of Technology and his colleagues Soon Ae Chun and Yoo Jung [4]. Their recently published paper, Toward the Semantic Deep Web, shows how ontologies and semantic Web constructs can be combined to more effectively extract information from the deep Web. They call this combination the ‘semantic deep Web.’

The authors posit that the structured roots of deep Web content lend themselves to better ontology learning from the Web. They also point to the usefulness of deep Web structure to annotations.

That such confluences are occurring between the semantic and deep “Webs” is a function of focused academic attention and the growing maturity of both perspectives. This year, for example, saw the inauguration of the first Workshop on Advances in Accessing Deep Web (ADW 2008). As part of the International Conference on Business Information Systems (BIS 2008), this meeting saw a lot of elbow rubbing with semantic Web and enterprise topics.

It might seem strange (indeed, sometimes it does to me ) to envision structured database content being served through a Web form and then converted via ontologies and other means to semantic Web formats. After all, why not go direct to the data?

And, of course, direct conversion is less lossy and more efficient.

But, one interesting point is that semantic Web techniques are increasingly working as a structure-extraction layer wrapping the standard Web. In that regard, starting with inherently structured source data — that is, the deep Web — can lead to higher quality inputs across the distributed, heterogeneous content of the Web.

Given the impossibility of everyone starting with the same premises and speaking the same languages and concepts, semantic Web mediation methods offer a way to overcome the Tower of Babel. And, when the starting content itself is inherently structured and (generally) of higher quality — that is, the deep Web — the logic of the combination becomes more obvious.

For More Information

Interested in learning more about the deep Web? I firstly recommend the resources posted at the bottom of Flynn-Burhoe’s timeline. And, for a very thorough treatment, I also recommend Denis Shestakov’s Ph.D. thesis from earlier this year [5]. It has a bibliography of some 115 references.

As late as 2002, no single search engine indexed the entire surface Web. There is much that has been written about that time, but emergence of Google (indeed others, it was a key battle at the time), worked to extend full search coverage to the Web, ending the need for so-called desktop metasearchers, then the only option for getting full Web search coverage.

Strangely, though full coverage of document indexing had been conquered for the Web, dynamic Web sites and database-backed sites fronted by search forms were also emerging. Estimates as of about 2001, made by myself and others, suggested such ‘deep Web‘ content was many, many times larger than the indexable document Web and was found in literally hundreds of thousands of sites.

Standard Web crawling is a different technique and technology than “probing” the contents of searchable databases, which require a query to be issued to a site’s search form. A company I founded, BrightPlanet, but many others such as Copernic or Intelliseek and others, many of which no longer exist, were formed with the specific aim to probe these thousands of valuable content sites.

From those company’s standpoints, mine at that time as well, there was always the threat that the major search engines would draw a bead on deep Web content and use their resources and clout to appropriate this market. Yahoo, for example, struck arrangements with some publishers of deep content to index their content directly, but that still fell short of the different technology that deep Web retrieval requires.

It was always a bit surprising that this rich storehouse of deep Web content was being neglected. In retrospect, perhaps it was understandable: there was still the standard Web document content to index and conquer.

Today, however, Google posted on one of its developer blog sites, Crawling through HTML forms, written by Jayant Madhavan and Alon Halevy, noted search and semantic Web researcher, announcing its new deep Web search:

In the past few months we have been exploring some HTML forms to try to discover new web pages and URLs that we otherwise couldn’t find and index for users who search on Google. Specifically, when we encounter a <FORM> element on a high-quality site, we might choose to do a small number of queries using the form. For text boxes, our computers automatically choose words from the site that has the form; for select menus, check boxes, and radio buttons on the form, we choose from among the values of the HTML. Having chosen the values for each input, we generate and then try to crawl URLs that correspond to a possible query a user may have made. If we ascertain that the web page resulting from our query is valid, interesting, and includes content not in our index, we may include it in our index much as we would include any other web page.

To be sure, there are differences and nuances to retrieval from the deep Web. What is described here is not truly directed nor comprehensive. But, the barrier has fallen. With time, and enough servers, the more inaccessible aspects of the deep Web will fall to the services of major engines such as Google.

And, this is a good thing for all consumers desiring full access to the Web of documents.

So, an era is coming to a close. And this, too, is appropriate. For we are also now transitioning into the complementary era of the Web of data.