Evaluating SQL selection/projection over table embeddings

Word embeddings are a powerful technique for representing and analyzing textual data in natural language processing (NLP) tasks. Notably, they possess a word-analogy property that represents realworld relationships through geometric operations in the embedding space. We study this property in the setting where embeddings are generated from relational databases. Our study aims to determine whether existing methods to obtain embeddings of relational tables preserve the inherent relationships of the relational model, akin to the word-analogy property in natural language text. By treating the learned vector space itself as an execution substrate, we ask a simple yet unexplored question: can an embedding faithfully stand in for the DBMS when answering basic SQL? To test this hypothesis, we develop a framework that assesses the capability of embeddings of relational data to answer SQL queries involving selection and projection. This framework encompasses the generation of embeddings, the querying of these embeddings through SQL, and the evaluation of the results. Our findings indicate that embedding methods that pretrain on the table can capture and reflect the original table/row/attribute relationships.