Fuzzy Semantic Labeling of Semi-structured Numerical Datasets

Abstract

SPARQL endpoints provide access to rich sources of data (e.g. knowledge graphs), which can be used to classify other less structured datasets (e.g. CSV files or HTML tables on the Web). We propose an approach to suggest types for the numerical columns of a collection of input files available as CSVs. Our approach is based on the application of the fuzzy c-means clustering technique to numerical data in the input files, using existing SPARQL endpoints to generate training datasets. Our approach has three major advantages: it works directly with live knowledge graphs, it does not require knowledge-graph profiling beforehand, and it avoids tedious and costly manual training to match values with types. We evaluate our approach against manually annotated datasets. The results show that the proposed approach classifies most of the types correctly for our test sets.