MIGA: A Unified Multi-Task Generation Framework for Conversational Text-to-SQL

Authors: Yingwen Fu, Wenjie Ou, Zhou Yu, Yue Lin

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate MIGA on two conversational text-to-SQL benchmarks Spar C and Co SQL. Comparisons against several advanced approaches show that MIGA tends to achieve state-of-the-art (SOTA) in two datasets based on Interaction Match (IM) accuracy. Extensive analyses provide some new insights into the proposed framework and some takeaways for future conversational text-to-SQL studies.
Researcher Affiliation Collaboration 1 Guangdong University of Foreign Studies, Guangzhou, China 2 Net Ease Games AI Lab, Guangzhou, China 3 Columbia University
Pseudocode Yes Algorithm 1: Training Process of MIGA
Open Source Code No The paper mentions using Hugging Face's Transformers and PyTorch, with links to their respective repositories and the T5 model, but does not provide a link or explicit statement for the open-sourcing of MIGA's specific implementation code.
Open Datasets Yes We use context-independent dataset Spider (Yu et al. 2018), conversational datasets Spar C (Yu et al. 2019b), Co SQL (Yu et al. 2019a) for the pre-training stage.
Dataset Splits Yes Detailed data statistics on training and evaluation. Since Spider is a single-turn dataset, experimental data involving interactions are treated as ignored for this dataset (marked with / ).
Hardware Specification Yes We implement MIGA based on Hugging Face s Transformers1 with Py Torch2 in 2 NVIDIA A100 GPUs.
Software Dependencies No The paper mentions using 'Hugging Face s Transformers' and 'Py Torch' and initializing with 'T5 model of 3B size', but does not specify their version numbers (e.g., PyTorch 1.9 or Transformers 4.x).
Experiment Setup Yes Epochs and learning rates for 2 training steps are set as (15, 50) and (1e-4, 5e-5). Batch size and optimizer are set as 64, and Ada Factor (Shazeer and Stern 2018) respectively. Following BERT (Kenton and Toutanova 2019), β in SQL perturbation is 15%. As for the other probability α, we set it as 0.15 in this paper after searching.