Note : Please run each script from this directory only since it contains relative paths
Here, each <X> represents an SQL-to-Text method
gazp-- represents "Grounded Adaptation for Zero-shot Executable Semantic Parsing" (Zhong et al. EMNLP 2020)l2s-- represents "Learning to Synthesize Data for Semantic Parsing" (Wang et al. NAACL 2021)snowball-- represents "Logic-Consistency Text Generation from Semantic Parses" (Shu et al. ACL 2021)
Each train_<X>.sh script does the following :
- Preprocesses the spider data appropriately and stores the preprocessed data in
../../data/sql-to-text/<X>/directory - Trains the appropriate model using the preprocessed data and saves the intermediate and best checkpoints in
../models/sql-to-text/<X>/directory
Each infer_<X>.sh script does the following :
- Reads spider workloads from
../../data/spider_groups/ - Generates text corresponding to each SQL using appropriate SQL-to-Text method (uses model checkpoints from the location where
train_<X>.shsaves them) - Splits the (SQL, generated-text) pairs into train-dev splits for further SmBop finetuning
Implementation of each method can be found in ../../sql-to-text/ directory