Towards Tabular Embeddings, Training the Relational Models

Correctly identifying the semantic label of a table column (e.g. artist ) or the domain of a tuple (e.g. Song) is crucial for data science tasks, such as schema matching, data cleaning and discovery. Existing data preparation and integration systems are known to make mistakes that need to be corrected by humans, which is labor-intensive and expensive especially at scale [19], [20], [25], [9], [8]. The accuracy is also known to suffer on dirty data. In this paper we define and evaluate tabular embeddings to help increase the accuracy.Embeddings, is a well-known dimensionality reduction technique, usually applied to represent (one-dimensional) text documents/sentences as vectors for further analytics [28], [7], [23]. Using embeddings lowers dimensionality, as well as the models trained with embeddings often exhibit higher accuracy, compared to the same trained without such, because the embeddings vectors store the context information. Our 2-dimensional embeddings are more suitable for tables rather than text. We justify their efficiency on fundamental tasks on tables such as classifying columns and tuples.We perform an extensive experimental evaluation and compare the Neural Network trained with tabular embeddings against the same trained without such embeddings. We report significant accuracy gains in tuple classification when using the Neural Network trained with our tabular embeddings - up to 17.6% delta in F-measure for Songs compared with the same without our embeddings. For training and evaluating all our embeddings and models, we use a large-scale WebTables dataset having ≈15 million tables coming from ≈ 248K English Web sources [12].