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

Conversational text-to-SQL is designed to translate multi-turn natural language questions into their corresponding SQL queries. Most advanced conversational methods are incompatible with generative pre-trained models (PLMs), such as T5. In this paper, we present a two-stage unified MultI-task Generation frAmework (MIGA) that leverages PLMs’ ability tackle text-to-SQL. the pre-training stage, MIGA first decomposes main task several related sub-tasks and then unifies them same sequence-to-sequence (Seq2Seq) paradigm task-specific prompts boost from multi-task training. Later in fine-tuning propose four perturbations alleviate error propagation problem. tends achieve state-of-the-art performance on two benchmarks (SparC CoSQL). We also provide extensive analyses discussions shed light some new perspectives for