Learning Representations for Zero-Shot Retrieval over Structured Data

TL;DR

This paper introduces a novel architecture for zero-shot retrieval of relevant structured data tables for question-answering systems, enabling better handling of large pools of candidate tables without prior table identification.

Contribution

It proposes a new retrieval-based approach for identifying relevant tables in structured data question-answering, moving beyond assumptions of known tables.

Findings

Demonstrates effective retrieval of relevant tables from large candidate pools.

Improves question-answering accuracy over structured data sources.

Provides a scalable architecture for zero-shot table retrieval.

Abstract

Large Scale Question-Answering systems today are widely used in downstream applications such as chatbots and conversational dialogue agents. Typically, such systems consist of an Answer Passage retrieval layer coupled with Machine Comprehension models trained on natural language query-passage pairs. Recent studies have explored Question Answering over structured data sources such as web-tables and relational databases. However, architectures such as Seq2SQL assume the correct table a priori which is input to the model along with the free text question. Our proposed method, analogues to a passage retrieval model in traditional Question-Answering systems, describes an architecture to discern the correct table pertaining to a given query from amongst a large pool of candidate tables.