Generated Knowledge Prompting

This repository contains the code for our ACL 2022 paper, Generated Knowledge Prompting for Commonsense Reasoning.

conda env create -f environment.yml
conda activate gkp

First, download the dataset for the following tasks: NumerSense, CommonsenseQA, CommonsenseQA 2.0, and QASC.

Next, use standardize/{task_name}_standardize.py to put the data in a unified format.

For convenience, we provide the standardized data at data/{task_name}/{split_name}.{task_name}.json

Use gpt3_generate_knowledge.py to generate knowledge for a task dataset. For example, to generate knowledge for the validation set of NumerSense, run

python knowledge/gpt3_generate_knowledge.py \
    --task numersense \
    --input_path data/numersense/validation.json \
    --output_path data/numersense/knowledge/knowledge_gpt3.validation.json \
    --prompt_path knowledge/prompts/numersense_prompt.txt

This will create a JSON file that contains the knowledge statements for each question in this dataset.

This step needs access to GPT-3. For convenience, we provide the knowledge files that were produced by our method and were used in downstream evaluation. These files are at data/{task_name}/knowledge/knowledge_gpt3.{split_name}.{task_name}.json

For the NumerSense task, use infer_numersense_t5.py to run inference with the generated knowledge. For example:

CUDA_VISIBLE_DEVICES=0 python inference/infer_numersense_t5.py \
    --model-type t5-11b \
    --input-path data/numersense/knowledge/knowledge_gpt3.validation.json

where --input-path is the knowledge file we produced in the previous step at the --output-path. This will create a JSON file that contains the inference results under data/{task_name}/inference/.

For the CommonsenseQA, CommonsenseQA 2.0, and QASC tasks, use infer_t5.py instead. For example:

CUDA_VISIBLE_DEVICES=0 python inference/infer_t5.py \
    --task csqa \
    --model-type allenai/unifiedqa-t5-11b \
    --input-path data/csqa/knowledge/knowledge_gpt3.dev.csqa.json

To get the baseline of your inference model (i.e. without using generated knowledge), use the standardized data file as --input-path:

CUDA_VISIBLE_DEVICES=0 python inference/infer_t5.py \
    --task csqa \
    --model-type allenai/unifiedqa-t5-11b \
    --input-path data/csqa/dev.csqa.json

Please check the argparse choices in the scripts for more parameters available.

On the QASC dataset, we find that the text-infilling question format is more suitable when using the off-the-shelf T5 as inference model. Therefore, we first convert the dataset to an infilling format:

python standardize/qasc_to_infill.py

For convenience, we include the infilling-format QASC dataset at data/qasc/{split_name}.qasc_infill.json.

For the knowledge generation step, still use the original standardized data file. For the inference step, use the knowledge file where questions are in the text-infilling format:

CUDA_VISIBLE_DEVICES=0 python inference/infer_t5.py \
    --task qasc \
    --model-type t5-11b \
    --input-path data/qasc/knowledge/knowledge_gpt3.dev.qasc_infill.json

Note that it is not needed to use the infilling-format data with finetuned T5 or any UnifiedQA inference models.

If you find this repository helpful, please consider citing our paper:

@article{liu2021generated,
  title={Generated knowledge prompting for commonsense reasoning},
  author={Liu, Jiacheng and Liu, Alisa and Lu, Ximing and Welleck, Sean and West, Peter and Bras, Ronan Le and Choi, Yejin and Hajishirzi, Hannaneh},
  journal={arXiv preprint arXiv:2110.08387},
  year={2021}
}