PROCLAIM: An Unsupervised Approach to Discover Domain-Specific Attribute Matchings from Heterogeneous Sources

Abstract

Schema matching is a critical problem in many applications where the main goal is to match attributes coming from heterogeneous sources. In this paper, we propose PROCLAIM (PROfile-based Cluster-Labeling for AttrIbute Matching), an automatic, unsupervised clustering-based approach to match attributes of a large number of heterogeneous sources. We define the concept of attribute profile to characterize the main properties of an attribute using: (i) the statistical distribution and the dimension of the attribute's values, (ii) the name and textual descriptions related to the attribute. The attribute matchings produced by PROCLAIM give the best representation of heterogeneous sources thanks to the cluster-labeling function we defined. We evaluate PROCLAIM on 45,000 different data sources coming from oil and gas authority open data website3. The results we obtain are promising and validate our approach.

Weakly Supervised Named Entity Recognition for Carbon Storage Using Deep Neural Networks

“…They used a conventional approach with three different embeddings configurations tested using a BiLSTM-CRF architecture. Some other approaches are focused on unsupervised clustering-based technique to match attributes of a large number of heterogeneous sources as also proposed in [19] to identify entities.…”