KagNet: Knowledge-Aware Graph Networks

News: We released a more general-purpose LM-GNN reasoning framework, MHGRN, which includes more options for text/graph encoders. It also matches the current state-of-the-art performance (76.5% acc) on the offical CommonsenseQA test set. We won't maintain this repo, so please follow the new repo.

Introduction

This codebase is an implementation of the proposed KagNet model for commonsense reasoning (EMNLP-IJCNLP 2019).

Overall Workflow
GCN + LSTM-based Path Encoder + Hierarchical Path Attention

Install Dependencies

sudo apt-get install graphviz libgraphviz-dev pkg-config
conda create -n kagnet_test python==3.6.3
conda activate kagnet_test
# which python
# which pip
pip install torch torchvision 
pip install tensorflow-gpu==1.10.0
conda install faiss-gpu cudatoolkit=10.0 -c pytorch -n kagnet_test 
pip install nltk
conda install -c conda-forge spacy -n kagnet_test
python -m spacy download en
pip install jsbeautifier
pip install networkx
pip install dgl
pip install pygraphviz
pip install allennlp

Datasets downloading


cd datasets
mkdir csqa_new

wget -P csqa_new https://s3.amazonaws.com/commensenseqa/train_rand_split.jsonl
wget -P csqa_new https://s3.amazonaws.com/commensenseqa/dev_rand_split.jsonl 
wget -P csqa_new https://s3.amazonaws.com/commensenseqa/test_rand_split_no_answers.jsonl


python convert_csqa.py csqa_new/train_rand_split.jsonl csqa_new/train_rand_split.jsonl.statements
python convert_csqa.py csqa_new/dev_rand_split.jsonl csqa_new/dev_rand_split.jsonl.statements
python convert_csqa.py csqa_new/test_rand_split_no_answers.jsonl csqa_new/test_rand_split_no_answers.jsonl.statements

Preprocess ConceptNet and embedding files

cd ../conceptnet
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
python extract_cpnet.py


cd ../triple_string
python triple_string_generation.py

# get concept and relation embeddings with frequency and vocab files
cd ../embeddings/
cd glove/
wget http://nlp.stanford.edu/data/glove.6B.zip
unzip glove.6B.zip
rm glove.*.zip
cd ../
python glove_to_npy.py  
python create_embeddings_glove.py

Concept Grounding

# concept grounding: core concept recognition (find mentioned concepts)
 
cd ../grounding/
python batched_grounding.py generate_bash "../datasets/csqa_new/train_rand_split.jsonl.statements"
bash cmd.sh
python batched_grounding.py combine "../datasets/csqa_new/train_rand_split.jsonl.statements"
python prune_qc.py ../datasets/csqa_new/train_rand_split.jsonl.statements.mcp

python batched_grounding.py generate_bash "../datasets/csqa_new/dev_rand_split.jsonl.statements"
bash cmd.sh
python batched_grounding.py combine "../datasets/csqa_new/dev_rand_split.jsonl.statements"
python prune_qc.py ../datasets/csqa_new/dev_rand_split.jsonl.statements.mcp

# python batched_grounding.py generate_bash "../datasets/csqa_new/test_rand_split.jsonl.statements"
# bash cmd.sh
# python batched_grounding.py combine "../datasets/csqa_new/test_rand_split.jsonl.statements"

Schema Graph Construction

cd ../pathfinder/
python graph_construction.py

python batched_pathfinding.py generate_bash "../datasets/csqa_new/train_rand_split.jsonl.statements.mcp"
bash cmd.sh
python batched_pathfinding.py combine "../datasets/csqa_new/train_rand_split.jsonl.statements.mcp"

python batched_pathfinding.py generate_bash "../datasets/csqa_new/dev_rand_split.jsonl.statements.mcp"
bash cmd.sh
python batched_pathfinding.py combine "../datasets/csqa_new/dev_rand_split.jsonl.statements.mcp"

# Pruning 

python path_scoring.py train
python path_scoring.py dev

python path_pruning.py train
python path_pruning.py dev

cd ../graph_generation
python graph_gen.py train
python graph_gen.py test

Train KagNet based on extracted BERT embeddings

cd ../baselines/

bash train_csqa_bert.sh
python extract_csqa_bert.py --bert_model bert-large-uncased --do_eval --do_lower_case --data_dir ../datasets/csqa_new --eval_batch_size 60 --learning_rate 1e-4  --max_seq_length 70 --mlp_hidden_dim 16 --output_dir ./models/ --save_model_name bert_large_b60g4lr1e-4wd0.01wp0.1_1337 --epoch_id 1 --data_split_to_extract train_rand_split.jsonl --output_sentvec_file ../datasets/csqa_new/train_rand_split.jsonl.statements.finetuned.large --layer_id -1 
python extract_csqa_bert.py --bert_model bert-large-uncased --do_eval --do_lower_case --data_dir ../datasets/csqa_new --eval_batch_size 60 --learning_rate 1e-4  --max_seq_length 70 --mlp_hidden_dim 16 --output_dir ./models/ --save_model_name bert_large_b60g4lr1e-4wd0.01wp0.1_1337 --epoch_id 1 --data_split_to_extract dev_rand_split.jsonl --output_sentvec_file ../datasets/csqa_new/dev_rand_split.jsonl.statements.finetuned.large --layer_id -1

cd ../models/
python main.py

Citation

@inproceedings{kagnet-emnlp19,
  author    = {Bill Yuchen Lin and Xinyue Chen and Jamin Chen and Xiang Ren},
  title     = {KagNet: Knowledge-Aware Graph Networks for Commonsense Reasoning.},
  booktitle = {Proceedings of EMNLP-IJCNLP},
  year      = {2019},
}

Remarks

Feel free to email yuchen[dot]lin[at]usc[dot]edu if you have any questions and need help.

INK-USC / KagNet