Predicting Appropriate Semantic Web Terms from Words

Lushan Han, Tim FininAAAI 2008

Abstract

The Semantic Web relies on the use of shared ontologies to ensure data interoperability. However, the vast number of existing namespaces and the technical complexity of RDF present a significant barrier to entry for human users. This paper presents a system designed to bridge the gap between natural language and formal ontologies by recommending appropriate Semantic Web terms based on simple English keywords. By integrating linguistic expansion via WordNet, real-world usage statistics from the Swoogle search engine, and a graph-optimized Naïve Bayes classifier, the system achieves over 90% accuracy in its top ten suggestions.

Introduction

The Resource Description Framework (RDF) is the foundational building block of the Semantic Web. In RDF, every concept is identified by a unique URI composed of a namespace (the ontology) and a local name (the specific term). While this structure ensures precision, it creates a "cold start" problem for users: selecting the correct URI requires an intimate knowledge of thousands of existing ontologies. For instance, a user wishing to describe a "Person" must choose between numerous competing ontologies, such as the Friend of a Friend (FOAF) vocabulary. Without a standardized selection process, the interoperability of the Semantic Web is compromised.

This research, originally presented at AAAI 2008, proposes a recommendation engine that translates natural language terms into statistically likely RDF namespaces by leveraging the collective intelligence of existing web data.

Methodology

The proposed system employs a multi-stage pipeline to transform a user's keyword into a ranked list of ontology terms.

1. Semantic Expansion

To account for the variability of natural language, the system first performs semantic expansion. Using WordNet, the system identifies synonyms and related concepts for the user's input. For example, the term "associate" might be expanded to include "relate" or "colleague," ensuring that the search for candidate ontologies is not limited by the user's specific word choice.

2. Candidate Retrieval and Filtering

The expanded term set is used to query Swoogle, a specialized search engine for the Semantic Web. Swoogle provides access to millions of RDF documents and thousands of ontologies. To ensure the quality of recommendations, the system utilizes Swoogle’s "OntoRank" algorithm to prioritize popular and well-established ontologies over obscure or poorly defined ones.

3. Contextual Prediction via Naïve Bayes

The core of the recommendation engine is a Naïve Bayes (NB) classifier. The system does not treat terms in isolation; instead, it considers the "context" of the user's current work. This context may include other namespaces already in use or a general domain description. The classifier calculates the conditional probability that a specific namespace is appropriate given the existing context.

Graph-Optimized Classification

Standard Naïve Bayes classification becomes computationally prohibitive when dealing with tens of thousands of classes (namespaces). To solve this, the authors introduced a namespace co-occurrence graph.

The Semantic Web is characterized by "locality"—certain ontologies are frequently used together (e.g., FOAF and Dublin Core), while others never overlap. By treating each RDF document in the Swoogle crawl as a "clique" in a sparse graph, the authors were able to prune the search space. When predicting a new namespace, the system only considers candidates that are neighbors of the existing context nodes in the graph. This "least connected" shortcut—starting the search with the namespace that has the fewest neighbors—allows for massive speedups in processing time.

Evaluation and Results

The system was trained and tested using a massive dataset derived from a Swoogle crawl, consisting of 2.5 million RDF documents, 10,000 ontologies, and 20,000 distinct namespaces.

Using ten-fold cross-validation, the researchers evaluated the system's ability to "fill in the blank" when a namespace was removed from a document. The results demonstrated high precision:

  • Top-1 Accuracy: The correct namespace was the first suggestion 55.6% of the time.
  • Top-3 Accuracy: 67.5%.
  • Top-10 Accuracy: 91.9%.

Furthermore, the system proved highly efficient. Despite the scale of the dataset, the graph model could be constructed in approximately 30 seconds on modest 2008-era hardware (a 3.0 GHz Celeron processor), demonstrating its potential for real-time application in ontology editors and data entry tools.

Discussion and Future Work

While the system is highly effective, it relies on the Naïve Bayes independence assumption, which posits that the presence of one namespace is independent of another. In reality, RDF data is highly structured, and future iterations could benefit from analyzing actual RDF triples (subject-predicate-object relationships) rather than simple co-occurrence. Additionally, further research is needed to balance the trade-offs between high recall (providing many options) and high precision (providing a single definitive answer).

By leveraging the collective patterns of the existing Semantic Web, this system provides a scalable path toward making RDF authoring more intuitive for human users.

Summary updated

BibTeX
@inproceedings{han2008predicting,
  author = {Lushan Han and Tim Finin},
  title = {{Predicting Appropriate Semantic Web Terms from Words}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
}
Figure