Abstract

Finding relevant tables is challenging and commonly performed using keyword, join, or union search. However, these are not always able to retrieve all the relevant tables since they usually require some form of exact matches or structural similarity. Semantic table search has recently been proposed as a novel technique that exploits knowledge graphs and various forms of semantic similarity for the retrieval of semantically relevant data lake tables. Specifically, semantic table search computes the relevance of a table based on the contained entity mentions without requiring strict structural similarities. This enables the discovery of a much larger set of relevant tables compared to existing approaches. In this paper, we demonstrate Jazero, the first semantic search system for data lakes featuring various semantic similarity metrics. Jazero uses the query-by-example paradigm, where a user provides an example table containing data of interest, enabling entity-centric discovery of data lake tables w.r.t. the query entity tuples. Jazero further enables scalable semantic table search with search space pre-filtering using the popular hierarchical navigable small world index over
different vector representations for knowledge graph entities, leading to an average runtime improvement of 87.4%. This corresponds to an average runtime of 3.3s. Jazero allows users to experience this novel discovery paradigm with various entity representations and similarity functions and to perform data discovery in multiple data lake instances. Jazerohas proven to retrieve near-disjoint results to keyword search, whilst retaining the same NDCG ranking performance.

Acknowledgements

This research was partially funded by the Danish Council for Independent Research (DFF), grant agreement no. DFF 804800051B, the Poul Due Jensen Fond, and DataGEMS, funded by the European Union’s Horizon Europe Research and Innovation programme, grant agreement no. 101188416.