This is the code associated with the paper:
Learning to map context dependent sentences to executable formal queries. Alane Suhr, Srinivasan Iyer, and Yoav Artzi. In NAACL, 2018. paper bib
I will be slowly adding and documenting the code. Please open an issue if you have any questions about the code or getting it to run.
If you want access to the data, code for preprocessing it, and the database to execute the queries, please email me (suhr@cs.cornell.edu). ATIS comprises LDC93S5, LDC94S19, and LDC95S26.
The main file is run.py. You can train with my parameters by calling train.sh, and evaluate by calling eval.sh (but will need to provide the save file name, and you can edit this script so that it selects segments from gold queries instead, as well as evaluating on different splits of the data).
Depending on your machine setup, you might not be able to run evaluation on all of the interactions on the GPU. The file cpu_full_interactions.txt contains the list of interactions that will only fit on CPU memory for an NVIDIA Tesla V100 GPU (16GB of RAM per GPU). You can add to this list if you are running into memory problems, and these interactions will be ignored during training and evalution. To predict on them, run evaluation with the --no_gpus flag set to true, and only these interactions will be evaluated (you can concatenate the output files for both runs to evaluate predictions on both sets of interactions).
You need Crayon installed and running. See this page for details. If you don't want to use Crayon, you can comment it out. You can use the provided scripts remove_experiments.py and delete_all_experiments.py in order to remove old experiments.
