Multi-Hop Table Retrieval for Open-Domain Text-to-SQL
Abstract: The open-domain text-to-SQL task aims to retrieve question-relevant tables from massive databases and then generate SQL. However, the performance of current methods is limited by single-hop retrieval, and existing multi-hop retrieval of other tasks cannot be directly applied due to two primary challenges: error cascades and the tendency to retrieve tables similar to the retrieved ones but irrelevant to the question. Therefore, we propose the multi-hop table retrieval with rewrite and beam search (Murre). To reduce error cascades, our method employs beam search to select multiple tables at each hop. To avoid retrieving similar but irrelevant tables, we remove the retrieved information from the question, guiding the retriever to focus on unretrieved tables. We conduct experiments on two open-domain text-to-SQL datasets, achieving an average improvement of 5.7% over the previous state-of-the-art results.
Paper Type: Long
Research Area: Semantics: Lexical and Sentence-Level
Research Area Keywords: Syntax: Parsing, NLP Applications
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 2923
Loading