A Tabular Open Data Search Engine Based on Word Embeddings for Data Integration

Abstract

Nowadays, open data has become a prominent information source for creating value-added product and services. Actually, open data portal initiatives are adopted by most of the governments to supply their public sector information, usually in the form of tabular data such as spreadsheets or CSV files. Most open data portals allow reusers to retrieve tabular open data by means of a keyword-based search engine over metadata. However, these search engines rely on the (not so often good enough) metadata quality, which must be complete, descriptive, and representative of the tabular open data content. Moreover, keyword-based search is not always an adequate solution for retrieving open data, since it does not consider the tabular nature of (most) open data and search results can be useless for reusers (e.g., when they attempt to find open data to be integrated with a given tabular dataset). To overcome these problems, this paper presents Search!, a search engine that enables users to pose an input query table to retrieve adequate tabular open data to be integrated with. To do so, semantic searches are performed by leveraging word embeddings to compute the similarity between column names and cell contents of tabular data. The relevance criteria established in the search engine aims to retrieve a ranking of tabular open datasets suitable for completion and augmentation, and thus, enabling integration with the input query table.