VeeAlign: Multifaceted Context Representation Using Dual Attention for Ontology Alignment

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10780–10792 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 10780 VeeAlign: Multifaceted Context Representation Using Dual Attention for Ontology Alignment Vivek Iyer University of Edinburgh Edinburgh, UK viyer@ed.ac.uk Arvind Agarwal IBM Research New Delhi, India arvagarw@in.ibm.com Harshit Kumar IBM Research New Delhi, India harshitk@in.ibm.com Abstract Ontology Alignment is an important research problem applied to various fields such as data integration, data transfer, data preparation, etc. State-of-the-art (SOTA) Ontology Alignment systems typically use naive domain-dependent approaches with handcrafted rules or domain- specific architectures, making them unscal- able and inefficient. In this work, we pro- pose VeeAlign, a Deep Learning based model that uses a novel dual-attention mechanism to compute the contextualized representation of a concept which, in turn, is used to dis- cover alignments. By doing this, not only is our approach able to exploit both syntac- tic and semantic information encoded in on- tologies, it is also, by design, flexible and scalable to different domains with minimal ef- fort. We evaluate our model on four differ- ent datasets from different domains and lan- guages, and establish its superiority through these results as well as detailed ablation stud- ies. The code and datasets used are available at https://github.com/Remorax/VeeAlign. 1 Introduction Ontology alignment is the task of establishing correspondences between semantically related el- ements (i.e. classes and properties) from differ- ent ontologies. It is useful for many applications, particularly for data integration and data migra- tion. The problem has been extensively studied in the past decade, and the solutions have ranged from simple rule based systems (Faria et al., 2013; Jiang et al., 2016) to the ones incorporating ex- ternal knowledge (Hertling and Paulheim, 2012; Algergawy et al., 2011), as well as the most recent ones that use sophisticated deep learning based sys- tems (Kolyvakis et al., 2018; Wang et al., 2018; Jiménez-Ruiz et al., 2020; Xue et al., 2021). A common limitation of these systems is their inabil- ity to generalize to new data - rule based systems are based on handcrafted rules which may not cover all data scenarios in all datasets. On the other hand, the Deep Learning (DL) based systems (Kolyvakis et al., 2018; Wang et al., 2018) proposed so far, often have strong dependencies on domain-specific external knowledge bases such as lexicons and the- sauri. Moreover, they also underperform as com- pared to their rule-based counterparts. One of the primary reasons for this and in fact, also the depen- dency of DL architectures on external background knowledge is the lack of sufficient, usable training data. Ground truth alignments are typically very scarce in number, especially for smaller ontologies, making supervised training difficult. Moreover, the small number of “positive” alignment pairs as compared to the “negative” ones (i.e entity pairs that contain an alignment versus those that do not) leads to a class imbalance and further adds to the difficulties of supervised learning based solely on the reference alignments. For example, the con- ference dataset …

Content truncated to 500 words