[Paper Review] Mesure de la similarité entre termes et labels de concepts ontologiques

Van Tien Nguyen, Christian Sallaberry+1 more|arXiv (Cornell University)|Jul 24, 2013

Data Management and Algorithms16 references4 citations

TL;DR

This paper proposes a hybrid string similarity metric, Liuppa(i,j), that combines two character-based string metrics (Jaro-Winkler) at both token and character sequence levels to improve similarity measurement between ontological labels and terms. It achieves superior performance—ranking third on the Cohen dataset—especially when word order matters, and enables accurate geospatial entity typing in geographic information retrieval by matching terms like 'chemin de fer touristique' to ontology labels such as 'voie ferrée touristique'.

ABSTRACT

We propose in this paper a method for measuring the similarity between ontological concepts and terms. Our metric can take into account not only the common words of two strings to compare but also other features such as the position of the words in these strings, or the number of deletion, insertion or replacement of words required for the construction of one of the two strings from each other. The proposed method was then used to determine the ontological concepts which are equivalent to the terms that qualify toponymes. It aims to find the topographical type of the toponyme.

Motivation & Objective

• To develop a robust method for measuring similarity between ontological labels and natural language terms, particularly in geographic information retrieval contexts.
• To address the challenge of matching terms with multi-word labels where word order significantly affects meaning.
• To integrate a hybrid similarity metric into an ontology-based geospatial typing pipeline for named spatial entities.
• To enable automatic parameter tuning of the similarity metric using a representative training sample.
• To support ontology matching workflows by enhancing the initial, label-based comparison phase before structural alignment.

Proposed method

• The method employs a meta-metric framework, Liuppa(i,j), which combines two base string metrics (i and j) at two levels: token-level and character-sequence level.
• It computes similarity by applying the first metric (e.g., Jaro-Winkler) to individual tokens and the second metric (e.g., Jaro-Winkler) to the full string sequences, then combines the results.
• The approach accounts for word order, common tokens, and edit operations (insertions, deletions, replacements) between strings.
• A parameter tuning phase uses a representative sample to select the optimal metric combination and threshold (ε = 0.84) for best performance.
• The final metric is applied to match terms extracted from travel narratives to geospatial ontology labels for entity typing.
• The method is evaluated on two datasets: a custom geospatial term-label dataset and the Cohen dataset for record linkage.

Experimental results

Research questions

• RQ1 How can string similarity between multi-word terms and ontological labels be improved when word order is semantically significant?
• RQ2 What combination of string metrics at different abstraction levels (tokens vs. character sequences) yields the highest accuracy in matching terms to ontology labels?
• RQ3 Can a hybrid similarity metric be automatically parameterized using a representative training sample for robust generalization across datasets?
• RQ4 How does the proposed method compare to existing string metrics in terms of performance on standard benchmark datasets?
• RQ5 To what extent can this similarity metric enhance the accuracy of geospatial entity typing in natural language documents?

Key findings

• The Liuppa(JaroWinkler, JaroWinkler) metric with ε = 0.84 achieved the best performance among 14 evaluated metrics on the custom geospatial term-label dataset.
• The method ranked third among 15 metrics on the Cohen dataset for record linkage, outperforming many standard approaches, especially in cases where word order matters.
• The hybrid approach significantly improved similarity measurement for terms like 'chemin de fer touristique' and 'voie ferrée touristique', enabling correct geospatial entity typing.
• The method demonstrated strong adaptability, as the optimal parameter set (metric pair and threshold) could be automatically determined from a representative sample.
• The approach proved effective in real-world geographic information retrieval, successfully linking terms from travel narratives to appropriate ontology concepts such as 'voie de communication'.
• The meta-metric framework Liuppa(i,j) is generalizable and applicable to other domains requiring ontology-based term matching, including database record linkage.