Interpretable Neural Semantic Parsing with Semantic Anchors as Intermediate Supervision

Code for the AAAI 2023 paper "Unveiling the Black Box of PLMs with Semantic Anchors: Towards Interpretable Neural Semantic Parsing".

Prerequisites

Dependencies

Below are the Python packages our implementation relies on.

torch                       1.9.0+cu111
torchaudio                  0.9.0
torchvision                 0.10.0+cu111
transformers                4.16.2
sqlite                      3.31.1
babel                       2.9.1

Datasets

The data we used in the experiments are respectively from: Overnight (link), KQA Pro (link) and WikiSQL (link). The preprocessed version is available here. Please refer to their original release license for proper usage.

# Extract the compressed files into the pre-specified locations
# e.g. Overnight dataset should appear in: "./data/overnight/data"
tar -zxvf data.tar.gz 

PLM Checkpoints

In case Huggingface Hub may update the model checkpoints over time, we also upload the initial PLM checkpoints (without any fine-tuning) used in our experiments here.

Usage

For running BART and T5 we prepared two separate folders "./bart/" and "./t5/" that can be regarded as Python sub-modules. Below we give a tutorial for running BART models with Overnight dataset but you may simply replace the "bart" with "t5" for running T5, and modify "overnight" into "kqapro" or "wikisql" for running on the other 2 datasets. For running the baselines models, you may simply remove the "--customized" and "--hybrid" flags.

Preprocess

python -m bart.preprocess \
--input_dir ./data/overnight/data/ \
--output_dir ./exp_files/overnight/ \
--config ./data/overnight/config.py \
--model_name_or_path ./bart-base \
--customized --supervision_form hybrid

Training

# For multi-GPU training
python -m torch.distributed.run \
--rdzv_backend=c10d 
--nproc_per_node=8 \
-m bart.train \
--input_dir ./exp_files/overnight/ \
--output_dir ./exp_results/overnight/ \
--config ./data/overnight/config.py \
--model_name_or_path ./bart-base/ \ 
--batch_size 64 \
--num_train_epochs 50 \
--customized --hybrid 

# For single-GPU training
python -m bart.train \
--input_dir ./exp_files/overnight/ \
--output_dir ./exp_results/overnight/ \
--config ./data/overnight/config.py \
--model_name_or_path ./bart-base/ \
--batch_size 64 \
--num_train_epochs 50 \
--customized --hybrid 

Inference

# Please run inference with single GPU only 
python -m bart.inference \
--input_dir ./exp_files/overnight/ \
--output_dir ./exp_results/overnight/ \
--model_name_or_path ./bart-base/ \
--ckpt ./exp_results/overnight/checkpoint-best/ \
--config ./data/overnight/config.py \
--customized --hybrid