Table Retrieval Does Not Necessitate Table-specific Model Design

Anonymous

Abstract: Tables are an important form of structured data for both human and machine readers alike, providing answers to questions that cannot, or cannot easily, be found in texts. Recent work designs special models and trains for table-related tasks such as table-based question answering and table retrieval. Though effective, they add model-data dual complexity to generic text solutions and obscure which elements are truly beneficial. In this work, we focus on the task of table retrieval, and ask: ``are table-specific model designs necessary for table retrieval, or can a text-generic model be effectively used to achieve a similar result?’’ We start by analyzing NQ-table, a set of table-answerable questions in the Natural Questions (NQ) dataset, and find 90\% of the questions can match tables in content with little concern for table structure. Motivated by this, we experiment with a general-purpose Dense Passage Retriever (DPR) for text and a special-purpose Dense Table Retriever (DTR) for tables. We show that DPR, without any design for or training on tables, can perform comparably well to the state-of-the-art DTR model, and neither adding DTR-like table-specific embeddings nor perturbing cell orders lead to significant changes. Both results strongly indicate that table retrieval does not necessitate table-specific model design, as well as the potential of directly applying powerful text-generic retrievers to structured tables.

Paper Type: long