Abstract
Knowledge bases like DBpedia are foundational for semantic applications but often suffer from incompleteness. This paper explores the potential of mining structured HTML tables from the Web to automatically discover and rank new attributes. We propose a two-step pipeline: first, aligning noisy table columns into distinct attribute clusters using fuzzy schema matching and graph partitioning; second, ranking these clusters based on their co-occurrence with existing ontology properties. Our results on the WDC Web Tables Corpus demonstrate that this approach effectively surfaces useful new attributes, significantly outperforming simple exact-matching baselines.
Introduction
Ontologies and knowledge graphs serve as the backbone for numerous semantic applications, yet manual curation is slow and often misses structured data already present on the Web. While much research focuses on extraction from unstructured text, HTML tables represent a vast, untapped resource of structured data. However, these tables are noisy, diverse, and lack a unified schema. To utilize them for ontology augmentation, we must solve two primary challenges: attribute alignment—grouping disparate column headers that represent the same concept—and relevance ranking—identifying which discovered attributes are truly valuable for the target ontology.
Methodology
The proposed pipeline consists of two primary stages:
- Attribute Matching and Clustering: We map web tables to an existing ontology (DBpedia) using T2K Match. For unmapped columns, we calculate similarities using label-based (string similarity), instance-based (data value similarity), and key/value-based techniques. These similarities form a graph, which is then partitioned into clusters representing unique attributes using algorithms like the Center algorithm.
- Attribute Ranking: Discovered clusters are ranked using co-occurrence metrics, including Conditional Probability, Schema Consistency, and Schema Coherency (based on Point-wise Mutual Information). We also evaluate the impact of weighting these scores by the source webpage's PageRank to account for domain authority.
Results
Evaluation on the T2D Gold Standard and the 147-million table WDC corpus revealed several key insights into the effectiveness of the pipeline.
In the matching phase (Figure 1), label-based matching significantly outperformed instance-based methods (Cosine and Containment) on the gold standard, likely due to the presence of high-quality headers in that dataset. For the clustering of these matches, the Center algorithm was found to be the most effective partitioning strategy, yielding the best F1-scores for reconstructing known attribute clusters at relatively low similarity thresholds.
Regarding ranking (Figure 2), metrics that consider the context of existing attributes—Schema Consistency and Coherency—generally outperformed isolated frequency counts in terms of both precision and recall. Interestingly, incorporating PageRank (the "+ PR" variants) did not consistently improve results and in some cases degraded performance, suggesting that a website's overall domain authority is not a direct proxy for the semantic utility of its specific tables.
Conclusion
Web tables represent a rich resource for ontology augmentation. By employing fuzzy clustering and context-aware ranking, we can reliably discover and integrate new attributes into knowledge bases. The success of the "Remove One" experiment, where the system successfully rediscovered known attributes like "population," further validates the pipeline's ability to surface objectively useful properties. Future work will explore extending this pipeline to discover new classes and relations beyond simple attributes.
Summary updated