Automated Alignment and Extraction of Bilingual Domain Ontology for Cross-Language Domain-Specific Applications

Jui-Feng Yeh, Chung-Hsien Wu, Ming-Jun Chen, Liang-Chih YuCOLING 2004

TL;DR

This paper introduces a novel method for automatically creating bilingual, domain-specific ontologies by aligning existing knowledge bases like WordNet and HowNet. This approach aims to improve cross-language information access and knowledge reuse by generating more comprehensive and semantically rich ontologies.

What Problem Does This Paper Address?

In today's increasingly globalized world, effective information exchange and communication across different languages are crucial. This necessitates robust multilingual lexical resources and ontologies, especially for natural language processing tasks like machine translation and cross-language information retrieval. While manual ontology construction is labor-intensive and error-prone, existing automatic methods often struggle with integrating well-defined source ontologies. This paper tackles the challenge of automatically building aligned bilingual domain-specific ontologies that capture richer structural and semantic information than individual language-dependent ontologies.

Methods

The proposed approach involves two main stages:

  1. Bilingual Ontology Alignment: This stage aligns WordNet (English) and HowNet (Chinese) to create a universal bilingual ontology. The alignment is based on the co-occurrence of words in parallel corpus sentence pairs. A key aspect is mapping HowNet's "primary features" (which define categories and relations like "is-a") to WordNet's synsets. The process uses a statistical approach to estimate the probability of a Chinese word being related to a specific English synset.
  2. Domain Ontology Extraction: For specific applications (e.g., medical domain), a domain-specific ontology is extracted from the bilingual ontology. This is achieved using an "island-driven algorithm" and a domain-specific corpus. The algorithm identifies "operative nodes" based on TF-IDF scores in the domain corpus compared to a contrastive corpus. It then activates inoperative nodes if their ancestors or descendants are operative, effectively expanding the domain concept coverage. Finally, domain-specific terminologies and axioms (e.g., from a medical encyclopedia) are integrated into the extracted ontology.

Datasets

  • WordNet: An existing English lexical database.
  • HowNet: An existing Chinese lexical database.
  • Sinorama Database: Used as a bilingual language parallel corpus for computing word co-occurrences.
  • Medical Domain Corpus: Used for extracting the medical domain ontology.
  • Contrastive (Out-of-Domain) Corpus: Used in conjunction with the domain corpus for identifying operative nodes.
  • Medical Encyclopedia: Used to derive domain-specific axioms and terminologies for the medical domain.
  • Benchmark Ontology: Constructed by domain experts (doctors and pharmacologists) based on UMLS, used for evaluating the proposed ontology.

Key Insights and Ideas

  • Leveraging Complementary Knowledge Bases: By aligning WordNet and HowNet, the approach creates a bilingual ontology that benefits from the structural and semantic coverage of both resources, leading to richer information.
  • Island-Driven Algorithm for Domain Extraction: This algorithm is a novel way to extract domain-specific concepts by leveraging the hierarchical structure of the universal ontology and identifying key terms based on corpus statistics. It helps overcome data sparsity issues.
  • Integration of Axioms: Including domain-specific axioms alongside terminologies significantly enhances the ontology's utility for specific applications, providing deeper semantic context.
  • Evaluation Framework: The paper proposes a structured evaluation framework with both taxonomic and non-taxonomic relation measures, and a benchmark ontology, which is crucial for assessing the quality of automatically generated ontologies.

Key Results

  • The proposed approach successfully extracts an aligned bilingual domain-specific ontology.
  • Taxonomic relation similarity: 0.57
  • Non-Taxonomic relation similarity: 0.68
  • In a medical web-mining system evaluation, the ontology-based system outperformed a baseline system using synonym expansion, particularly in recall rate. For instance, at a recall level of 0.7, the ontology-based system achieved 68% precision compared to 30% for the baseline.

Limitations and Future Work

  • Limitations: The paper notes that the number of words mapped to the same concept in the upper layer of the ontology is larger than in the lower layer, attributing this to terminologies typically appearing in lower layers. The evaluation of the domain application was limited to a medical domain.
  • Future Work: Not explicitly detailed, but the success in the medical domain suggests potential for applying this approach to other domains and further refining the extraction and alignment algorithms. The paper also implies that the overall performance of information retrieval systems is directly related to the quality of the ontology.

Summary updated

BibTeX
@inproceedings{yeh2004automated,
  author = {Jui-Feng Yeh and Chung-Hsien Wu and Ming-Jun Chen and Liang-Chih Yu},
  title = {{Automated Alignment and Extraction of Bilingual Domain Ontology for Cross-Language Domain-Specific Applications}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2004},
}
Figure