FERN: Cross-Benchmark Table Retrieval for Text-to-SQL with Frozen Retrievers and Learned Ranking
arXiv cs.CL · 11 Nov 2026, 04:39 AEST
FERN is a new approach for table retrieval in Text-to-SQL systems that combines frozen retrievers and learned ranking. It shows competitive recall rates compared to existing models like RASL across various datasets.
Key facts
- FERN (Feature-Engineered Ranking with No generative LLM) instead fuses three frozen retrievers into 100 candidates, describes each with up to 120 features, and reorders them with LightGBM rankers fitted once on BIRD's training questions.
- On Spider, where it was not fitted, its candidate pool alone reaches RASL's published recall at fifteen tables (98.0), and the ranker raises the share of questions with every gold table in the top five (complete@5) from 83.9% to 95.6%.
- On 121 real BEAVER questions, without BEAVER fitting, its recall at fifteen (82.1) falls between two BEAVER-trained retrievers (80.8, 84.5), but its complete@15 is below both (49.6 vs 54.5 and 62.0).
- On BIRD dev with evidence, non-retrieval features raise complete@5 by 7.7 points over retrieval signals alone.
- On BIRD dev, median latency is 208 ms per question on one A100, without hosted model calls.
Read aloud
FERN is a new approach for table retrieval in Text-to-SQL systems that combines frozen retrievers and learned ranking. It shows competitive recall rates compared to existing models like RASL across various datasets. FERN Feature-Engineered Ranking with No generative L L M instead fuses three frozen retrievers into one hundred candidates, describes each with up to one hundred twenty features, and reorders them with LightGBM rankers fitted once on BIRD's training questions. On Spider, where it was not fitted, its candidate pool alone reaches RASL's published recall at fifteen tables ninety eight point zero, and the ranker raises the share of questions with every gold table in the top five complete@five from eighty three point nine percent to ninety five point six percent. On one hundred twenty one real BEAVER questions, without BEAVER fitting, its recall at fifteen eighty two point one falls between two BEAVER-trained retrievers eighty point eight, eighty four point five, but its complete@fifteen is below both forty nine point six vs fifty four point five and sixty two point zero. On BIRD dev with evidence, non-retrieval features raise complete@five by seven point seven points over retrieval signals alone. On BIRD dev, median latency is two hundred eight ms per question on one A100, without hosted model calls.
Sources in this cluster
- FERN: Cross-Benchmark Table Retrieval for Text-to-SQL with Frozen Retrievers and Learned Ranking
arXiv cs.CL · 11 Nov 2026, 04:39 AEST