1: \begin{abstract}

2: Text-to-SQL technology has significant applications in realizing database query through natural language, with no requirement for learning SQL grammar. Nevertheless, the challenge is that modeling alignment between database information and consideration in a certain query is not obvious. Text-to-SQL parsing is proposed as novel Grammar Pre-training (GP) to decode deep relations between database and question. To adequately learn the internal relationship of SQL grammar, the decoder is pre-trained independently of the encoder. Subsequently, the robustness of the model is improved and convergence is accelerated. Flooding level is adopted to obtain the non-zero training loss and avoid local extrema problems. Ultimately, we achieved better performance on Spider, a cross-DB Text-to-SQL dataset (72.8\% dev, 69.8\% test)by encoding the sentence with ${\bf G{\scriptstyle RAPPA}}$ and RAT-SQL model. By reducing the average loss by 78.9\%, the variance is only 0.8\% of the previous model while training. Moreover, experiments proved that this technique converges much faster and has excellent robustness.

3: \end{abstract}

4: