Linking Tabular Columns to Unseen Ontologies

Abstract

We introduce a novel approach for linking table columns to types in an ontology unseen during training. As the target ontology is unknown to the model during training, this may be considered a zero-shot linking task at the ontological level. This task is often a requirement for businesses that wish to semantically enrich their tabular data with types from their custom or industry-specific ontologies without the benefit of initial supervision. In this paper, we describe specific approaches and provide datasets for this new task: training models on open domain tables using a broad source ontology and evaluating them on increasingly difficult tables with target ontologies having different levels of type granularity. We use pre-trained Transformer encoder models and a range of encoding strategies to explore methods of encoding increasing amounts of ontological knowledge, such as type glossaries and taxonomies, to obtain better zero-shot performance. We demonstrate these results empirically through extensive experiments on three new public benchmark datasets.

A Appendix

1.1 A.1 Model Predictions

The following tables below shows examples of predictions returned by our proposed model built using a pretrained TinyBERT encoder. This model is trained using Wikidata labels and is asked to predict from the DBpedia target ontology for the top two tables. For the bottom two tables, the model predicts from the UMLS Semantic Network (UMLS SN).

The first row in the block titled Top model prediction returns model predictions using Type labels only. The second row returns predictions using Type labels and associated glossaries. The final row in this block returns predictions using our proposed encoding strategy. Note that the BioDivTab benchmark does not contain table metadata.