Semi-Automatic Ontology Extension Using Spreading Activation

Wei Liu, Albert Weichselbraun, Arno Scharl, Elizabeth ChangwebLyzard 2005

Abstract

Building domain-specific ontologies is a critical but labor-intensive task in the development of the Semantic Web. This paper presents a system designed to semi-automatically extend and refine ontologies by mining textual data from international online media. By expanding a small seed ontology, the system creates a semantic network through co-occurrence analysis, trigger phrase detection, and disambiguation using the WordNet lexical dictionary. The core of the approach is a "spreading activation" algorithm that identifies the most relevant candidates for inclusion. Using a "climate change" seed ontology as a case study, we demonstrate how this hybrid method integrates statistical and linguistic analysis to produce a robust, hierarchical knowledge network.

The Knowledge Acquisition Bottleneck

In the Semantic Web, ontologies serve as the backbone for shared understanding, defining concepts and relationships so that humans and machines can communicate effectively. However, creating these maps manually is expensive and slow, often failing to keep up with the rapid evolution of information. This "knowledge acquisition bottleneck" necessitates automated tools that can assist domain experts by suggesting new terms and hierarchical relationships from massive text corpora.

System Architecture and Methodology

The proposed system follows a multi-stage pipeline that combines statistical analysis with linguistic rules to bridge the gap between raw text and structured knowledge.

Figure 1: The system architecture for semi-automatic ontology extension, showing the flow from lexical analysis to concept positioning.

1. Corpus Mining and Lexical Analysis

The process begins with a small, expert-validated "seed" ontology. For this study, a crawling agent gathered 50MB of textual data from 156 international news sites. The Lexical Analyzer then processes this corpus using:

  • Co-occurrence Analysis: Utilizing the Log Likelihood Algorithm to find terms frequently appearing near seed terms.
  • Trigger Phrase Analysis: Identifying linguistic patterns such as "X such as Y" or "X is a type of Y" to infer relationships.
  • WordNet Integration: Using the lexical database to disambiguate meanings and identify initial synonyms.

2. Spreading Activation: The "Secret Sauce"

The system constructs a semantic network where nodes represent terms and links are weighted based on "trustworthiness." To prioritize domain-specific discovery over general dictionary definitions, WordNet links are assigned a lower weight (0.2) compared to corpus-derived links.

In the spreading activation phase, "energy" is injected into the seed nodes and allowed to flow through the network. Terms that accumulate the highest activation levels are flagged as the most relevant candidates. This mechanism acts as a consensus layer, naturally integrating "knowledge-poor" statistical methods with "knowledge-rich" linguistic rules.

3. Relationship Confirmation

Once candidates are identified, the system determines their placement in the hierarchy using Subsumption Analysis (checking if one term's appearance is a subset of another's) and Head Noun Analysis (e.g., identifying that "Carbon Dioxide" is a type of "Dioxide").

Key Results and Discussion

The system successfully expanded a 7-node "Climate Change" seed into a detailed hierarchy. The spreading activation process identified highly relevant terms such as Kyoto Protocol, Carbon Dioxide, and Emissions as the strongest candidates for inclusion.

Figure 4: The resulting extended ontology for the "Climate Change" domain, showing newly discovered concepts and their relationships.

As seen in the final output, the system correctly identified that "Greenhouse Gas" subsumes "Methane" and "Carbon Dioxide." It also established broader contextual links between "Global Warming" and concepts like "Arctic," "Species," and "Ice."

Limitations

Despite its success, the system faces challenges with non-hierarchical relationships. For instance, while it correctly identified "Scientists" as relevant to "Greenhouse Gas," it could not define the specific nature of the link, labeling it simply as a general "relation." Additionally, high-frequency terms like "EPA" (Environmental Protection Agency) were occasionally pulled in due to high co-occurrence, even though they are organizations rather than scientific concepts.

Conclusion and Future Work

This research demonstrates that spreading activation is an effective heuristic for integrating diverse knowledge sources into a unified ontology extension process. By balancing statistical evidence from news corpora with linguistic data from WordNet, the system provides a scalable way to maintain domain-specific knowledge. Future work will focus on using machine learning to dynamically adapt trustworthiness weights, further reducing the need for manual tuning by domain experts.

Summary updated

Abstract

This paper describes a system to semi-automatically extend and refine ontologies by mining textual data from the Web sites of international online media. Expanding a seed ontology creates a semantic network through co-occurrence analysis, trigger phrase analysis, and disambiguation based on the WordNet lexical dictionary. Spreading activation then processes this semantic network to find the most probable candidates for inclusion in an extended ontology. Approaches to identifying hierarchical relationships such as subsumption, head noun analysis and WordNet consultation are used to confirm and classify the found relationships. Using a seed ontology on "climate change " as an example, this paper demonstrates how spreading activation improves the result by naturally integrating the mentioned methods.
BibTeX
@inproceedings{liu2005semi,
  author = {Wei Liu and Albert Weichselbraun and Arno Scharl and Elizabeth Chang},
  title = {{Semi-Automatic Ontology Extension Using Spreading Activation}},
  booktitle = {webLyzard},
  year = {2005},
}
Figure