DeepAlignment: Unsupervised Ontology Matching with Refined Word Vectors

Proceedings of NAACL-HLT 2018, pages 787–798 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics DeepAlignment: Unsupervised Ontology Matching With Refined Word Vectors Prodromos Kolyvakis1, Alexandros Kalousis2, Dimitris Kiritsis1 1 ´Ecole Polytechnique F´ed´erale de Lausanne (EPFL), Lausanne, Switzerland {prodromos.kolyvakis,dimitris.kiritsis}@epfl.ch 2Business Informatics Department, University of Applied Sciences, Western Switzerland Carouge, HES-SO, Switzerland alexandros.kalousis@hesge.ch Abstract Ontologies compartmentalize types and rela- tions in a target domain and provide the se- mantic backbone needed for a plethora of prac- tical applications. Very often different on- tologies are developed independently for the same domain. Such “parallel” ontologies raise the need for a process that will estab- lish alignments between their entities in order to unify and extend the existing knowledge. In this work, we present a novel entity align- ment method which we dub DeepAlignment. DeepAlignment refines pre-trained word vec- tors aiming at deriving ontological entity de- scriptions which are tailored to the ontol- ogy matching task. The absence of explicit information relevant to the ontology match- ing task during the refinement process makes DeepAlignment completely unsupervised. We empirically evaluate our method using stan- dard ontology matching benchmarks. We present significant performance improvements over the current state-of-the-art, demonstrat- ing the advantages that representation learning techniques bring to ontology matching. 1 Introduction Translation across heterogeneous conceptual sys- tems is an important challenge for cognitive sci- ence (Goldstone and Rogosky, 2002; Stolk et al., 2016). Ontology Matching constitutes the task of establishing correspondences between seman- tically related entities (i.e. classes and proper- ties) from different ontologies, as illustrated in Figure 1. Similarly, ontology matching is crucial for accomplishing a mutual understanding across heterogeneous artificial cognitive agents (Taylor, 2015). However, despite the many proposed solu- tions, it is widely accepted that there is no solution robust enough to deal with the high ontological linguistic variability (Shvaiko and Euzenat, 2008, Entity Conference Dinner Television Episode Entity MealMenu Banquet TV Episode Figure 1: Example of alignments (black lines) and mis- alignments (red crossed lines) between ontologies. 2013); hampering, thus, the discovery of shared meanings. Research in automatic ontology matching has focused on engineering features from terminolog- ical, structural, extensional (ontology instances) and semantic model information extracted from the ontological model. These features are then used to compute ontological entity similarities that will guide the ontology matching. Deriving such features for a given problem is an extremely time consuming task. To make matters worse, these features do not transfer in other domains. As Cheatham and Hitzler (2013) have recently shown, the performance of ontology matching based on different textual features varies greatly with the type of ontologies under consideration. At the same time, machine learning research is characterised by a shift from feature engineer- ing based approaches to feature and representa- tion learning as a result of the performance im- provements brought by deep learning methods. A by now classical example is the unsupervised learning of semantic word representations based on the distributional hypothesis (Harris, 1954), i.e. the assumption that semantically similar or re- lated words appear …

Content truncated to 500 words