Introduction
The World Wide Web serves as a vast repository of information, yet much of it remains trapped in HTML formats that lack explicit semantic meaning. Traditional search engines, relying primarily on keyword matching, often fail to extract specific, relevant data points from complex web pages. The primary challenge lies in document heterogeneity: a single page may contain highly structured tables, semi-structured lists, and unstructured prose.
The SHIRI-Annot system addresses this by providing an unsupervised, ontology-driven framework designed to automatically annotate heterogeneous HTML documents with semantic metadata. By transforming messy web content into structured RDF data, the system enables precise, context-aware querying.
The SHIRI-Annot Framework
The SHIRI-Annot approach consists of a two-phase process: the Locating Phase and the Annotating Phase.
System Architecture
The overall architecture of the system facilitates the transition from semi-structured document inputs to a queryable Knowledge Base through a series of automated steps.
- Locating Phase: The system identifies "concept instances" (e.g., dates, names, or research fields) using domain-independent, unsupervised tools. It leverages C-Pankow, which utilizes Hearst patterns and search engine results for categorization, and the Senellart technique, which uses external databases like DBLP to identify specific entities.
- Annotating Phase: This phase represents the core innovation. The system utilizes a domain ontology to generate "Annotation Rules" based on Horn first-order logic.
Semantic Annotation Model
The system treats HTML tags as "structural units." When one unit is nested within another, the system infers a semantic relationship based on the document's hierarchy.
To handle complex data, the authors introduce specialized metadata classes:
- Structural Units: These formalize nesting; for example, a list item (
<li>) within a list (<ul>) implies a relationship between the contained concepts. - MetaData Classes: Classes such as
PartOfSpeechare used for units containing multiple distinct concepts, whileSetOfX(e.g.,SetOfTopic) handles lists of similar concepts. - Relation Inference: If the ontology lacks a specific relationship between nested items, the system applies an
unnamedRelationlabel to preserve the link for future SPARQL queries.
Experimental Evaluation
The system was evaluated using a corpus of 444 HTML documents from 33 scientific conference websites. These documents were pre-processed to ensure they were well-formed before undergoing annotation.
The preliminary study focused on the Topic concept, which is particularly challenging due to its varied presentation formats (e.g., bulleted lists vs. inline text). The system achieved:
- Precision: 84.3%
- Recall: 65.1%
The authors anticipate that recall will improve as logic rules for the PartOfSpeech metadata—which accounted for nearly 24% of test cases—are fully implemented.
Conclusion and Future Work
SHIRI-Annot demonstrates that document structure can effectively imply semantic relationships. By formalizing nesting and handling "inseparable" concept blocks through specialized metadata, the system allows for robust query rewriting.
Future work will focus on reducing the system's dependency on initial locating tools, integrating advanced Natural Language Processing (NLP) for better entity recognition in unstructured text, and expanding the framework to other domains such as e-commerce.
Summary updated