Introduction
Taxonomies are fundamental for organizing knowledge into hierarchical structures, such as Siberian tiger $\rightarrow$ tiger $\rightarrow$ feline. While manual resources like WordNet are highly accurate, they are costly to maintain and often lack coverage for niche or rapidly evolving domains. Existing automated tools typically rely on "Web-scale" data redundancy to identify frequent patterns. However, in specialized fields with limited corpora, these frequentist methods struggle due to data sparsity. TAXIFY is an unsupervised system designed to build high-quality "is-a" taxonomies from small, domain-specific text collections by leveraging clustering and novel graph-pruning techniques.
The TAXIFY Framework
TAXIFY operates through four distinct phases to overcome the limitations of small datasets:
- Initial Construction: The system identifies "seed" relations using Hearst patterns (e.g., "X such as Y"). To maintain domain relevance, a custom domain-specificity score filters out generic terms.
- Syntactic Inference: Multi-word concepts are automatically linked to their linguistic heads (e.g., "whale shark" is assigned "shark" as a supertype).
- Clustering-based Inference: To discover relationships between concepts that do not co-occur in Hearst patterns, TAXIFY employs distributional semantics. It clusters concepts based on their usage in the text. As shown in the figure below, if "Munich" and "Rome" share similar contexts and "Rome" is known to be a "city," TAXIFY infers that "Munich" is also a "city."
- Graph-based Pruning: To clean noisy automated extractions, TAXIFY uses a "long path" intuition. It removes direct edges that bypass established long chains of intermediate concepts, as these are often over-generalization errors (e.g., linking "poodle" directly to "organism" instead of "dog").
Key Results and Evaluation
The system was evaluated across five diverse corpora, including biomedical (DDI, PMC) and Wikipedia-based (Animals, Plants, Vehicles) datasets. TAXIFY was compared against state-of-the-art systems K&H and OntoLearn.
As illustrated in the performance comparison below, TAXIFY consistently outperforms baseline methods in both precision and the scale of correct relations discovered. On the DDI dataset, it identified significantly more valid relations than K&H while maintaining higher accuracy. Ablation studies revealed that clustering-based inference increased the number of correct edges by 25.9% to 68.6%, while the pruning strategy improved precision by an average of 8% to 15%.
Furthermore, when compared to the expert-curated MeSH taxonomy, approximately 53% of TAXIFY’s drug relations either matched MeSH exactly or provided finer-grained detail than the gold standard.
Discussion and Future Work
While TAXIFY excels on focused datasets, its precision slightly decreased on the broader PMC corpus due to "pollution" from unrelated sub-hierarchies. Future research will explore integrating TAXIFY into downstream tasks like question answering and combining local corpora with broader web data to fill knowledge gaps.
Summary updated