DeLVe-SQL: Decoupled Latent Variable Models for Few-Shot Text-to-SQL

Anonymous

Abstract: Few-shot single-table text-to-sql tasks present considerable challenges due to the constraints of limited training data. Existing approaches primarily transform this problem into column-based classification tasks and utilize self-training methods to leverage unlabeled texts with pseudo-labels. The critical challenge, however, lies in selecting high-quality pseudo-labels and incorporating them effectively into model training. Past self-training techniques selected pseudo SQL predictions based on the probabilities yielded by column-specific classifiers. This approach may not align well with the original queries, especially given the limited performance of the few-shot classifier. To address these limitations, we introduce a novel approach DeLVe-SQL: a latent variable model specifically designed for few-shot text-to-sql tasks. This model effectively decouples textual and SQL semantics via distinct latent variables, enhancing the classifier's performance. Moreover, we apply an additional GPT2 decoder to take into account the reconstruction probabilities of the original query given pseudo SQL predictions, providing a more refined weighting of pseudo-labels. Our experiments, conducted on both open-domain and domain-specific benchmarks, demonstrate that our proposed method delivers promising results, outperforming existing methods in few-shot scenarios.

Paper Type: long

Research Area: Semantics: Sentence-level Semantics, Textual Inference and Other areas

Contribution Types: NLP engineering experiment, Approaches to low-resource settings, Approaches low compute settings-efficiency

Languages Studied: English, Chinese