Annotating web tables through ontology matching

Vasilis Efthymiou, Oktie Hassanzadeh, Mohammad Sadoghi, Mariano Rodriguez-MuroOM@ISWC 2016 · OpenReview

Introduction

The proliferation of data in HTML tables across the Web presents a significant opportunity for knowledge base (KB) enrichment. However, the lack of semantic structure in these tables makes machine interpretation difficult. This paper introduces a scalable framework that reinterprets Web table annotation as an ontology matching problem, leveraging mature alignment tools to link table data to KBs like DBpedia.

Problem Statement

Web table annotation involves linking table cells to entities and columns to properties in a KB. While humans can easily infer context, machines struggle with ambiguity. Existing solutions are often specialized and iterative, making them difficult to scale to the massive size of modern KBs. The authors investigate whether transforming tables into mini-ontologies allows the use of highly optimized, general-purpose ontology matching tools for this task.

Methodology

The proposed framework converts table annotation into a standard ontology alignment task through a three-stage pipeline:

  1. Semantic Modeling: Tables are transformed into structured ontologies. A two-pass scan identifies the "label column" (main entity) and distinguishes between literal data and entity references. Each row is then instantiated as a formal entity.
  2. Blocking (Candidate Selection): To ensure scalability, a pre-filtering step limits the KB search space to instances whose labels match strings found in the table.
  3. Ontology Matching: The generated "table ontology" and the filtered KB ontology are processed using LogMap, a logic-based matching tool.
  4. Scalability: The pipeline is implemented using MapReduce, enabling parallel processing of large-scale datasets.

Evaluation

The framework was evaluated using the T2D Gold Standard benchmark against DBpedia. Performance was compared with the state-of-the-art T2K tool and standard DBpedia lookup baselines.

Results

The experimental results demonstrate that general-purpose ontology matchers can achieve competitive performance, particularly in precision.

Table 1: Comparison of the proposed ontology matching framework against DBpedia lookup baselines and the state-of-the-art T2K tool.

As shown in Table 1, the ontology matching approach achieved a high precision of 0.89, nearly matching the specialized T2K tool (0.90). While the recall (0.57) and F-measure (0.70) are currently lower than T2K, the framework's speed is a significant advantage, processing the evaluation corpus in under four minutes via MapReduce.

Discussion and Future Work

The primary limitation identified is the "recall gap," largely attributed to the initial blocking step. The values in parentheses in Table 1 represent the raw blocking results (Recall: 0.71, Precision: 0.32), indicating that while the blocker captures many candidates, it also introduces significant noise that the matcher must filter. Future work will focus on refining the blocking strategy and establishing a dedicated Web table matching track within the Ontology Alignment Evaluation Initiative (OAEI).

Conclusion

By bridging Web table annotation and ontology matching, this research demonstrates that mature alignment technologies can be effectively repurposed for large-scale semantic data integration. The high precision and scalability of the MapReduce-based pipeline suggest a viable path for automated KB augmentation from the Web.

Summary updated

BibTeX
@inproceedings{efthymiou2016annotating,
  author = {Vasilis Efthymiou and Oktie Hassanzadeh and Mohammad Sadoghi and Mariano Rodriguez-Muro},
  title = {{Annotating web tables through ontology matching}},
  booktitle = {OM@ISWC},
  year = {2016},
  url = {https://openreview.net/forum?id=Wmt2oCqcQ6LC},
}
Figure