FERN: Cross-Benchmark Table Retrieval for Text-to-SQL with Frozen Retrievers and Learned Ranking
FERN put every required table in the top five for 95.6% of Spider questions, despite being trained only on BIRD.
Its learned ranker raised that result from 83.9% using candidates supplied by three frozen retrievers, without a generative LLM. Top-five table recall was close to RASL’s published results on Spider and BIRD with evidence hints, but fell 8.2 points behind on BIRD without hints. Transfer was uneven: FERN trailed both BEAVER-trained retrievers on retrieving every required table and fell well short on Fiben at fifteen tables, with different question sets limiting that comparison. Median retrieval latency on BIRD dev was 208 milliseconds per question on one A100, without hosted model calls.
ArXiv · AI/CL/LG's note
Its learned ranker raised that result from 83.9% using candidates supplied by three frozen retrievers, without a generative LLM. Top-five table recall was close to RASL’s published results on Spider and BIRD with evidence hints, but fell 8.2 points behind on BIRD without hints. Transfer was uneven: FERN trailed both BEAVER-trained retrievers on retrieving every required table and fell well short on Fiben at fifteen tables, with different question sets limiting that comparison. Median retrieval latency on BIRD dev was 208 milliseconds per question on one A100, without hosted model calls.
ArXiv · AI/CL/LG's note
score 5