Syntactic and Semantic Tree Structure in HTML

All browsers parse HTML into a Document Object Model, or DOM. This represents the syntactic tree of an HTML document.

The RDF standard is used to group semantic meaning into hierarchical form.

There tends to be some overlap between the syntactic DOM tree of a web site and the semantic tree of the information it presents. For instance, the DOM of Google's search results page is (with some nodes omitted for clarity:

HTML
 \- BODY
     \- DIV
         |- P
	 |  |- A
	 |  |   \- [ResultURL]
	 |  \- FONT
	 |      |- [ResultSummary]
	 |      |- SPAN
	 |      |   \- A
	 |      |      \- [Category]
	 |      |- [ResultURL]
	 |      |- A
	 |      |  \- [CacheURL]
	 |      \- A
	 |         \- [SimilarPagesURL]
         |- P
	 |  |- A
	 |  |   \- [ResultURL]
	 |  \- FONT
	 |      |- [ResultSummary]
	 |      |- SPAN
	 |      |   \- A
	 |      |      \- [Category]
	 |      |- [ResultURL]
	 |      |- A
	 |      |  \- [CacheURL]
	 .     \- A
	 .        \- [SimilarPagesURL]
	 .

A semantic ontology for describing Google's results might look like:

[SearchResult]
  |- [ResultURL]
  |- [Summary]
  |- [CacheURL]
  |- [Category]
  \- [SimilarPagesURL]

There are some similarities between the syntactic and semantic trees. For instance, one might say that the entire <p> tag represents the [Result] class in the ontology. Each time a similar syntactic tree, rooted at <p>, is found, it is an instance of the [Result] ontology.

Existing Work

For a summary of non-hierarchical wrapper induction approaches, see my other notes here and here.

Stalker, by Muslea, et. al., also takes a hierarchical approach to wrapper induction by developing a semantic tree structure for each page. However, they seem to ignore the hierarchical structure of HTML itself, instead proceeding to extract information in a traditional, linear fashion.

Approach

• Labeling (and later identifying) semantic data within the syntactic HTML DOM involves three parts
  1. Identifying the predecessor tags (starting at the root) which lead up to the semantic subtree.
  2. Identifying the context of the semantic subtree (its sibling nodes at its root)
  3. Identifying the subtree itself.
• All of these items may have variations between intances. For instance:
  1. the Google [SearchResult] class may appear as a child of the nodes HTML-BODY-DIV or HTML-BODY-DIV-BLOCKQUOTE.
  2. [need contextual example]
  3. the [Category] class may or may not appear, depending on the result.
  The representation must have enough flexibility to accomodate these variations.
• Use both syntactic structures (predecessors, context, and subtree) to identify other instances of this semantic data.

Interface

• User loads the page to be tagged in browser.
• User highlights an example of the text to be semantically labeled. For instance, highlights a single Google search result.
• System finds other similar items in page using the selection's underlying syntactic structure, and highlights those as well.
• User may add items that the system missed, or remove those that are mislabeled. System updates selections as appropriate.
• When selections are correct, user inputs a semantic label for them and their syntactic structure is stored in RDF along with the semantic label.
• Upon returning to this page or a similar one, the semantic data for that page may be retrieved using the previously defined syntactic expression.

Unresolved Issues

• Internet Explorer does not tie selection directly to DOM, making it difficult to label full subtrees of the DOM. Possibly use Mozilla (which already has "View Selection Source" context menu which highlights source that user has selected).

Current Work / Next Steps

• Have already implemented test browser using IE component. Allows highlighting of single, leaf-level elements of DOM with semantic labelling. Able to extrapolate other occurrences and add variants. Successful at highlighting Google results, slashdot.org stories. Unsuccessful at some IMDB.com semantic data, such as actors and directors.
• Need to expand the ability to highlight full subtrees of the DOM and label them semantically.
• Need to incorporate context of subtree into representation.
• Need way of removing erroneous inductions from representation.

References

Berners-Lee, T., Hendler, J., Lassila, O. The Semantic Web. Scientific American, May 2001. http://www.scientificamerican.com/article.cfm?articleID=00048144-10D2-1C70-84A9809EC588EF21&catID=2

Miller, R., Myers, B. Lightweight Structured Text Processing. In Proceedings of USENIX 1999 Annual Technical Conference, June 1999, Monterey, CA. http://www-2.cs.cmu.edu/~rcm/papers/usenix99/

Muslea, I., Minton, S., and Knoblock, C. 1999. A hierarchical approach to wrapper induction. In Proceedings of the Third International Conference on Autonomous Agents (Agents'99), Seattle, WA. http://citeseer.nj.nec.com/muslea99hierarchical.html

Shih, L., Karger, D. Learning Classes Correlated to a Hierarchy. MIT AI Lab Technical Note.