This is the pytorch implementation of our paper "Language Generation with Multi-hop Reasoning on Commonsense Knowledge Graph". The arxiv version of the paper could be found here.
python version >= 3
torch version >= 1.4.0
transformers == 2.8.0
nltk == 3.4.5
networkx == 2.1
spacy == 2.2.1
torch-scatter == 2.0.5+${CUDA}
For torch-scatter, ${CUDA} should be replaced by either cpu, cu92, cu101 or cu102 depending on your PyTorch installation.
For more information check here.
Preprocessed datasets can be downloaded from here.
Unzip the file and move it to data.
Extract English ConceptNet and build graph.
cd data
wget https://s3.amazonaws.com/conceptnet/downloads/2018/edges/conceptnet-assertions-5.6.0.csv.gz
gzip -d conceptnet-assertions-5.6.0.csv.gz
cd ../preprocess
python extract_cpnet.py
python graph_construction.pyPreprocessing multi-hop relational paths for the model. Set $DATA to either anlg, eg, story.
export DATA=eg
python ground_concepts_simple.py $DATA
python find_neighbours.py $DATA
python filter_triple.py $DATADownload the pre-trained GPT-2 model.
mkdir -p models
cd models
mkdir -p gpt2-small
wget -O pytorch_model.bin https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-pytorch_model.bin
wget -O vocab.json https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-vocab.json
wget -O merges.txt https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-merges.txt
wget -O config.json https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-config.jsonAdd special tokens to vocabulary.
cd scripts
python add_special_tokens.pyThe following command is an example to train the model on the trarining set and evaluate on the development set. Set $DATA_TYPE to either anlg, eg, story.
export DATA_TYPE={anlg, eg, story}
export ROOT_PATH=..
export DEVICE=1
CUDA_VISIBLE_DEVICES=${DEVICE} \
python3 main.py \
--train_data_file ${ROOT_PATH}/data/${DATA_TYPE}/train \
--dev_data_file ${ROOT_PATH}/data/${DATA_TYPE}/dev \
--test_data_file ${ROOT_PATH}/data/${DATA_TYPE}/test \
--graph_path 2hops_100_directed_triple_filter.json \
--output_dir ${ROOT_PATH}/models/${DATA_TYPE}/grf-${DATA_TYPE} \
--source_length 32 \
--target_length 32 \
--model_type gpt2 \
--model_name_or_path ${ROOT_PATH}/models/gpt2-small \
--do_train \
--per_gpu_train_batch_size 16 \
--per_gpu_eval_batch_size 16 \
--workers 7 \
--seed 42 \
--evaluate_metrics bleu \
--overwrite_output_dir \
--num_train_epochs 3 \
--learning_rate 1e-5 \
--aggregate_method max \
--alpha 3 \
--beta 5 \
--gamma 0.5 \
--weight_decay 0.0 \
--warmup_ratio 0.0 \
--logging_steps 20 \Note: Some scatter operations in torch_scatter are currently non-deterministic and not controlled by the random seed due to the usage of atomic operations.
See the discussions here.
So different runs may vary slightly.
To reproduce our results, you can directly download our checkpoints and evaluate the model.
The following command is an example to run the inference on the test set. Set $DATA_TYPE to either anlg, eg, story.
export DATA_TYPE={anlg, eg, story}
export ROOT_PATH=..
export DEVICE=1
CUDA_VISIBLE_DEVICES=${DEVICE} \
python3 main.py \
--train_data_file ${ROOT_PATH}/data/${DATA_TYPE}/train \
--dev_data_file ${ROOT_PATH}/data/${DATA_TYPE}/dev \
--test_data_file ${ROOT_PATH}/data/${DATA_TYPE}/test \
--graph_path 2hops_100_directed_triple_filter.json \
--output_dir ${ROOT_PATH}/models/${DATA_TYPE}/grf-${DATA_TYPE} \
--source_length 32 \
--target_length 32 \
--model_type gpt2 \
--model_name_or_path ${ROOT_PATH}/models/gpt2-small \
--do_eval \
--per_gpu_train_batch_size 16 \
--per_gpu_eval_batch_size 16 \
--workers 7 \
--seed 42 \
--evaluate_metrics bleu \
--overwrite_output_dir \
--aggregate_method max \
--gamma 0.5 \We also provide generated results in results/ on the three tasks and you can directly evaluate the results with our evaluation scripts.
To evaluate the generated results on anlg and eg, run the following commands. Set $DATA to either anlg, eg.
export DATA={eg, anlg}
python evaluation/eval.py --dataset ${DATA} --output_dir results/grf-${DATA}.txtTo evaluate the generated results on story, run the following commands.
python eval_story.py results/grf-story.txt@inproceedings{ji2020language,
title = "Language Generation with Multi-Hop Reasoning on Commonsense Knowledge Graph",
author = "Ji, Haozhe and Ke, Pei and Huang, Shaohan and Wei, Furu and Zhu, Xiaoyan and Huang, Minlie",
booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing",
year = "2020",
}
Please kindly cite our paper if you find this paper and the codes helpful.
Many thanks to the Github repository of Transformers, fairseq and KagNet. Part of our codes are modified based on their codes.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent