To Describe or Not to Describe? Benchmarking Database Representations for Schema Linking in Text-to-SQL

2026-04-12 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, medium

Summary

A study presented at PROPOR 2026 by Daiane Ucceli Kreitlow and Hilário Tomaz Alves de Oliveira investigates Schema Linking for Text-to-SQL systems, specifically for questions in Brazilian Portuguese. The research compares two schema representation strategies: natural-language descriptions generated by Large Language Models (LLMs) and representations derived from Data Definition Language (DDL) and Data Manipulation Language (DML) commands. Experiments were conducted on a Brazilian Portuguese adaptation of the Spider dataset, which includes over 200 databases. The evaluation involved several LLMs and embedding models, with results based on Hit@k metrics. The findings indicate that natural language descriptions consistently outperform DDL/DML-based representations, highlighting the superior effectiveness of LLM-generated schema descriptions for Schema Linking tasks in Text-to-SQL contexts.

Key takeaway

For AI Engineers developing Text-to-SQL systems, especially for non-English languages like Brazilian Portuguese, you should prioritize using Large Language Models to generate natural language descriptions of database schemas. This approach has been shown to consistently outperform DDL/DML-based representations in schema linking tasks, leading to more accurate identification of relevant databases, tables, and columns.

Key insights

LLM-generated natural language descriptions enhance Text-to-SQL schema linking more effectively than DDL/DML.

Principles

• Natural language descriptions improve schema linking.
• LLMs can generate effective schema representations.

Method

The study compared LLM-generated natural language descriptions against DDL/DML representations for schema linking on a Brazilian Portuguese Spider dataset using Hit@k metrics.

In practice

• Use LLMs for database schema descriptions.
• Prioritize natural language for schema linking.

Topics

• Text-to-SQL
• Schema Linking
• Large Language Models
• Database Representations
• Brazilian Portuguese

Best for: AI Engineer, Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• IndicDB -- Benchmarking Multilingual Text-to-SQL Capabilities in Indian Languages — IndicDB reveals a significant "Indic Gap" in LLM Text-to-SQL performance for Indian languages compared to English.
• Why NL-to-SQL Systems Fail: Lessons from building a Semantic SQL validation pipeline — Semantic correctness and robust validation are paramount for reliable NL-to-SQL systems, beyond mere syntax.
• A Larger Annotated Corpus of Portuguese Coreference — LLMs can annotate Portuguese coreference, but a new LLM-based system did not outperform existing rule-based methods.
• Textual Inference in Portuguese: Comparing Language Models — LLMs struggle with logic-sensitive semantic reasoning in Portuguese, highlighting the need for specialized benchmarks.
• IndicDB -- Benchmarking Multilingual Text-to-SQL Capabilities in Indian Languages — IndicDB benchmarks multilingual Text-to-SQL for Indic languages, revealing a significant performance gap compared to English.
• Dissecting the MAD Landscape OS Space with DBpedia — Neural Machine Translation can effectively bridge the semantic gap between natural language and formal query languages like SPARQL.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.