This repository is the official implementation of COLA: Contextualized Commonsense Causal Reasoning from the Causal Inference Perspective.
The paper is accepted to the main conference of ACL 2023.
An Overview of Our Method
Python version is 3.8.5
requirements:
allennlp_models==2.10.1
datasets==2.6.1
evaluate==0.3.0
matplotlib==3.6.1
numpy==1.23.4
pandas==1.5.0
pathos==0.2.9
statsmodels==0.13.2
torch==1.10.2
tqdm==4.64.1
transformers==4.20.0
ujson==5.5.0
You can install all requirements with the command
pip install -r requirements.txt
We extract a large dataset containing 800K pairs of events and their temporal order from RocStories. You can download it here download All licenses are subject to RocStories' original release.
The contextualized commonsense causal reasoing dataset COPES is in the dir ./COPES_data.
Our COLA Framework owns a few steps to conduct inference.
First, you need to fine-tune a temporal predictor, which learns temporal relations.
The code for this part is in mlm.py and mlm_util.py.
We provide an example command in the script run_script/finetune_mlm.sh. It fine-tunes
bert-base-uncased, you can change it to other models.
Second, you need to sample covariates for each event.
Run the script run_script/sample_covariates.sh.
After getting the covariates, you need to do Multistamp Covariate Sampling, which
tasks the union of covariates sampled at multiple timestamps.
Run the example command
python mix_covariate.py --input_path YOUR_COVARIATES --output_path YOUR_OUTPUT_PATH --copes_path ./COPES_data/COPES.json
To generate interventions of events, you first need to run SRL tool
to extract verbs and two arguments ARG0 and ARG1:
python SRL_parsing.py --data_path ./COPES_data/COPES.json \
--model_path structured-prediction-srl-bert --batch_size 512 \
--output_dir YOUR_OUTPUT_DIR
After generating covariates and interventions, you can predict the temporal relations between covariates, interventions and RocStories events.
For temporal relations in covariate balancing (between covariates and interventions/original causes),
run run_script/get_cov-inter.sh.
For average treatment effect (between interventions/original causes and caused events),
run run_script/get_inter-outcome.sh.
Finally, you can conduct causal inference based on temporal relations: run
run_script/evaluate.sh.
Some operations are computationally heavy (e.g., using GPT-J-6b), you can download our pre-computed results here download Temporal relations are predicted by fine-tuned bert-base-uncased.
If you find any typo or bug, please open an issue. This repo is maintained by Zhaowei Wang