An Investigation into the Feasibility of the Semantic Web

Zhengxiang Pan, Abir Qasem, Jeff HeflinAAAI 2006

Introduction

The Semantic Web envisions a global, machine-understandable data infrastructure where information is linked across decentralized sources. However, realizing this vision requires overcoming two primary hurdles: heterogeneity (the lack of a central authority for naming) and scalability (the massive volume of decentralized data). This paper investigates the feasibility of large-scale reasoning by introducing the Description Logic Database (DLDB) system, which leverages relational databases and "ontology perspectives" to manage millions of triples efficiently.

System Architecture

The core of the proposed solution is the DLDB system, designed to bridge the gap between expressive reasoning and high-performance data storage.

The DLDB system architecture, illustrating the interaction between the RDBMS backbone, the DL Reasoner, and the API layers.

DLDB employs a three-part strategy to achieve scalability:

  1. Relational Backbone: Instead of in-memory processing, data is stored in a PostgreSQL database using a schema-aware approach where classes and properties map to specific tables.
  2. Hybrid Reasoning: A Description Logic (DL) reasoner (Racer Pro) pre-computes class hierarchies during ontology loading. These hierarchies are materialized as database views, allowing complex queries to be executed as optimized SQL joins.
  3. Ontology Perspectives: To resolve heterogeneity without a global model, users select a "basis ontology." Mapping ontologies then align disparate vocabularies using OWL axioms, creating a localized, consistent view of the data that avoids the "poison pill" problem of global inconsistencies.

Experimental Evaluation

The system was tested against a massive real-world dataset from Swoogle’s 2005 index, comprising approximately 45 million triples across 343,977 documents and over 41,000 ontologies. The entire index was housed in an 8GB database, demonstrating the efficiency of the RDBMS-backed approach.

Results and Discussion

The experiments demonstrated that reasoning over millions of triples is computationally viable. A key finding was the "network effect" of mappings: by linking common concepts across different ontologies, the system significantly increased data recall. For instance, in one large-scale test, mapping multiple versions of a Person concept allowed the system to aggregate over 1.2 million results from diverse sources.

Experimental results showing query performance (ms) and result counts across various ontology perspectives.

As shown in the performance data, most queries resolved in under one second. Even complex integrated queries involving multiple mapped ontologies—such as Query 1, which maps FOAF ($O_f$), SWRC ($O_s$), and others—remained efficient, taking approximately 5 seconds to aggregate hundreds of results. This performance is largely due to the RDBMS handling the heavy lifting of join operations via pre-computed views, which is significantly faster than traditional recursive reasoning.

Conclusion

While DLDB proves the feasibility of medium-scale Semantic Web reasoning, several challenges remain. The system currently has gaps in handling transitive properties and instance-level identity (the "same-as" problem). Furthermore, as the web scales toward billions of triples, manual mapping will need to be replaced by automated ontology alignment. Nevertheless, the success of the "perspective" framework suggests a practical path forward for managing the inherent messiness of decentralized data.

Summary updated

BibTeX
@inproceedings{pan2006investigation,
  author = {Zhengxiang Pan and Abir Qasem and Jeff Heflin},
  title = {{An Investigation into the Feasibility of the Semantic Web}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure