Learning Metadata-Agnostic Representations for Text-to-SQL In-Context Example Selection
This addresses the challenge of efficient example selection for complex tasks like text-to-SQL, offering a domain-agnostic solution that improves accuracy with lower latency, though it is incremental as it builds on existing in-context learning methods.
The paper tackled the problem of selecting optimal demonstrations for in-context learning in text-to-SQL tasks by proposing MARLO, a method that aligns natural language questions and SQL queries in a shared embedding space, resulting in an average improvement of +2.9% in execution accuracy on the Spider benchmark compared to generic embedding models.
In-context learning (ICL) is a powerful paradigm where large language models (LLMs) benefit from task demonstrations added to the prompt. Yet, selecting optimal demonstrations is not trivial, especially for complex or multi-modal tasks where input and output distributions differ. We hypothesize that forming task-specific representations of the input is key. In this paper, we propose a method to align representations of natural language questions and those of SQL queries in a shared embedding space. Our technique, dubbed MARLO - Metadata-Agnostic Representation Learning for Text-tO-SQL - uses query structure to model querying intent without over-indexing on underlying database metadata (i.e. tables, columns, or domain-specific entities of a database referenced in the question or query). This allows MARLO to select examples that are structurally and semantically relevant for the task rather than examples that are spuriously related to a certain domain or question phrasing. When used to retrieve examples based on question similarity, MARLO shows superior performance compared to generic embedding models (on average +2.9\%pt. in execution accuracy) on the Spider benchmark. It also outperforms the next best method that masks metadata information by +0.8\%pt. in execution accuracy on average, while imposing a significantly lower inference latency.