# Bridging Textual and Tabular Data for Cross-Domain Text-to-SQL Semantic Parsing

**Authors:** Xi Victoria Lin, Richard Socher, Caiming Xiong  
**arXiv:** [2012.12627](https://arxiv.org/abs/2012.12627)  
**Date:** 2020-12-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2012.12627v2/

## Key points

- Authors Xi Victoria Lin, Richard Socher, and Caiming Xiong introduce BRIDGE, a sequential architecture for cross-domain text-to-SQL semantic parsing.
- BRIDGE represents the question and DB schema as a tagged sequence enriched with cell values referenced in the question.
- The hybrid sequence is encoded using BERT with minimal additional layers, leveraging deep attention mechanisms for text-DB contextualization.
- BRIDGE achieves state-of-the-art performance on benchmarks like Spider (71.1% dev, 67.5% test) and WikiSQL (92.6% dev, 91.9% test).
- The model effectively captures cross-modal dependencies and demonstrates potential for generalization to other text-DB related tasks.
- BRIDGE emphasizes interpretability by utilizing anchor texts to enhance alignment between textual mentions and DB schema components.
- The implementation of BRIDGE is publicly available at \url{https://github.com/salesforce/TabularSemanticParsing}, providing a valuable resource for further exploration in this domain.

## Summary

, , , , 

In their paper titled "Bridging Textual and Tabular Data for Cross-Domain Text-to-SQL Semantic Parsing," authors Xi Victoria Lin, Richard Socher, and Caiming Xiong introduce BRIDGE, a sequential architecture designed to model dependencies between natural language questions and relational databases in cross-DB semantic parsing. The key innovation of BRIDGE lies in its representation of the question and DB schema as a tagged sequence, where certain fields are enriched with cell values referenced in the question. This hybrid sequence is then encoded using BERT with minimal additional layers, leveraging deep attention mechanisms for text-DB contextualization. Furthermore, BRIDGE incorporates a pointer-generator decoder that prioritizes schema-consistency driven search space pruning. This approach has enabled BRIDGE to achieve state-of-the-art performance on prominent cross-DB text-to-SQL benchmarks like Spider (71.1% dev, 67.5% test with ensemble model) and WikiSQL (92.6% dev, 91.9% test). Through detailed analysis, the authors demonstrate that BRIDGE effectively captures cross-modal dependencies and exhibits potential for generalization to other text-DB related tasks. The authors emphasize the interpretability of BRIDGE by utilizing anchor texts to enhance alignment between textual mentions and DB schema components. By maximizing the utilization of pre-trained language models like BERT, BRIDGE excels at linking text references with database structures. Overall, their findings suggest that BRIDGE is adept at handling natural language variations and structural patterns while achieving outstanding performance on challenging text-to-SQL benchmarks such as WikiSQL. This comprehensive study extends beyond traditional semantic parsing approaches by emphasizing the importance of bridging textual and tabular data effectively for enhanced performance in cross-domain applications. The implementation of BRIDGE is publicly available at \url{https://github.com/salesforce/TabularSemanticParsing}, providing researchers and practitioners with a valuable resource for further exploration in this domain.