thu-bpm / isesl-sql

The source code of paper "Semantic Enhanced Text-to-SQL Parsing via Iteratively Learning Schema Linking Graph" in KDD2022.

ISESL-SQL

The source code of paper "Semantic Enhanced Text-to-SQL Parsing via Iteratively Learning Schema Linking Graph" published at KDD 2022.

download data

spider: https://yale-lily.github.io/spider

put the data into the data/ directory

preprocess data

python3 -u preprocess/process_dataset.py --dataset_path data/train.json --raw_table_path data/tables.json --table_path data/tables.bin --output_path 'data/train.bin' --skip_large --semantic_graph
python3 -u preprocess/process_dataset.py --dataset_path data/dev.json --table_path data/tables.bin --output_path 'data/dev.bin' --skip_large --semantic_graph
python3 -u preprocess/process_graphs.py --dataset_path 'data/train.bin' --table_path data/tables.bin --output_path  data/train.rgatsql.bin
python3 -u preprocess/process_graphs.py --dataset_path 'data/dev.bin' --table_path data/tables.bin --output_path  data/dev.rgatsql.bin

train model

CUDA_VISIBLE_DEVICES=0 python scripts/text2sql.py --task lgesql_large --seed 999 --device 0 --plm google/electra-large-discriminator --gnn_hidden_size 512 --dropout 0.2 --attn_drop 0.0 --att_vec_size 512 --model rgatsql --output_model without_pruning --score_function affine --relation_share_heads --subword_aggregation attentive-pooling --schema_aggregation head+tail --gnn_num_layers 8 --num_heads 8 --lstm onlstm --chunk_size 8 --drop_connect 0.2 --lstm_hidden_size 512 --lstm_num_layers 1 --action_embed_size 128 --field_embed_size 64 --type_embed_size 64 --no_context_feeding --batch_size 35 --grad_accumulate 5 --lr 1e-4 --l2 0.1 --warmup_ratio 0.1 --lr_schedule linear --eval_after_epoch 120 --smoothing 0.15 --layerwise_decay 0.8 --max_epoch 200 --max_norm 5 --beam_size 5 --logdir logdir/run --train_path train --dev_path dev --training --optimize_graph --schema_loss 

Reference

@inproceedings{liu2022semantic,
  title={Semantic Enhanced Text-to-SQL Parsing via Iteratively Learning Schema Linking Graph},
  author={Liu, Aiwei and Hu, Xuming and Lin, Li and Wen, Lijie},
  booktitle={Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  pages={1021--1030},
  year={2022}
}

isesl-sql's Issues

#ISESL-SQL-main

Hello, when I was training the model, the step was always 0, right? At the same time, the training accuracy was 0：

Another question is, if the model is trained well, how can I input text and output SQL? thanks

script for inference

Which script can be used for inference, i.e. loading the best model and then getting the final SQL file?

Original Spider Dataset used in the Paper + Setup Instructions to Run on a New Database

Hey,

I tried following the setup instructions given in README.md but I think because of some changes in the Spider dataset these instructions are no longer valid.

For example,

python3 -u preprocess/process_dataset.py --dataset_path data/train.json --raw_table_path data/tables.json --table_path data/tables.bin --output_path 'data/train.bin' --skip_large --semantic_graph

There is no 'data/train.json' file in the Spider dataset. But it has 'data/train_spider.json' and 'data/train_others.json' files.

I tried changing the file name, but I get the following error.

Firstly, preprocess the original databases ... Traceback (most recent call last): File "/sensei-fs/users/saudi/text2sql/ISESL-SQL/preprocess/process_dataset.py", line 74, in <module> tables = process_tables(processor, tables_list, args.table_path, args.verbose) File "/sensei-fs/users/saudi/text2sql/ISESL-SQL/preprocess/process_dataset.py", line 26, in process_tables tables[each['db_id']] = processor.preprocess_database(each, verbose=verbose) File "/sensei-fs/users/saudi/text2sql/ISESL-SQL/preprocess/common_utils.py", line 100, in preprocess_database c = [w.lemma.lower() for s in doc.sentences for w in s.words] File "/sensei-fs/users/saudi/text2sql/ISESL-SQL/preprocess/common_utils.py", line 100, in <listcomp> c = [w.lemma.lower() for s in doc.sentences for w in s.words] AttributeError: 'NoneType' object has no attribute 'lower'

If possible, please upload the original dataset used in google drive and share the link.

Also, please provide the instructions to run the pipeline on a new database. Like what files need to be created in the data folder and what scripts to use.

Your help on this is much appreciated.

Best,
Saud Iqbal

#ISESL-SQL

According to README.md, when training the model, there is a problem, as follows：

But the “tables.bin” does not exist in the original spider dataset.
Thank you！

Code Release

Nice work, how about the code?

Question about oracle schema linking

Nice work！
I have noticed that the oracle schema linking information could bring a huge improvement for Spider and Spider-SYN in section3.6, I wonder that where does the oracle schema information come from for both of two datasets. Could you offer a link please?
(I guess that the oracle schema information of Spider come from SLSQL(EMNLP2020), and how about Spider-SYN? )

Besides, could you please offer more details about how you directly change the implicit graph matrix A(𝑡) in Eq. 7 based on the oracle information, becaucse I want to try this by myself. Thanks a lot.

Best wishes!