HKUST-KnowComp / DISCOS-commonsense

Codes for the WWW2021 paper: DISCOS: Bridging the Gap between Discourse Knowledge and Commonsense Knowledge (https://arxiv.org/abs/2101.00154).

DISCOS-commonsense

This is the github repo for The Web Conference (WWW) 2021 paper DISCOS: Bridging the Gap between Discourse Knowledge and Commonsense Knowledge.

Check out the follow-up work, a benchmark that evaluates the performance of transforming discourse knowledge to commonsense knowledge: (EMNLP 2021, CSKB Population) Benchmarking Commonsense Knowledge Base Population with an Effective Evaluation Dataset, and the code repositry.

How to train the Commonsense Knowledge Graph Population (CKGP) model

Here is the instruction for learning the CKGP model. We use the filtered graph as introduced in Section 5.1.1 for this experiment.

First, git clone this repo, and then download the prepared aligned graph from here. The data/graph_cache folder contains the data that can be directly used for training and testing. The data/graph_raw_data is the graph file that we got after aligning ATOMIC and ASER, with pre-defined negative edges for the CKGP task. The data/infer_candidates folder contains the candidate (h, r, t) tuples to be scored by our BertSAGE model.

Next install dependencies. Recommended python version is 3.8+.

pip install -r requirements.txt

Note that to load the files in data/graph_cache requires the same dependencies as in the requirements.txt file. E.g., you need to install the transformers package with version 3.4.0.

Next train the BertSAGE model. For example here is the command to train with the oReact relation:

python -u BertSAGE/train.py --model graphsage \
    --load_edge_types ASER \
    --neg_prop 1 \
    --graph_cach_path data/graph_cache/neg_{}_{}_{}.pickle \
    --negative_sample prepared_neg \
    --file_path data/graph_raw_data/G_aser_oReact_1hop_thresh_100_neg_other_20_inv_10.pickle

For other relations, you could find the corresponding .pickle file from data folder.

For the inference part, you could run after training:

python -u BertSAGE/infer.py --gpu 0 --model graphsage \
    --model_path models/G_aser_oReact_1hop_thresh_100_neg_other_20_inv_10/graphsage_best_bert_bs64_opt_SGD_lr0.01_decay0.8_500_layer1_neighnum_4_graph_ASER_acc.pth \
    --infer_path data/infer_candidates/G_aser_oReact_1hop_thresh_100_neg_other_20_inv_10.npy \
    --graph_cach_path data/graph_cache/neg_prepared_neg_ASER_G_aser_oReact_1hop_thresh_100_neg_other_20_inv_10.pickle

The Acquired Knowledge Graph DISCOS-ATOMIC

By populating the knowledge in ATOMIC to the whole ASER, we can acquire a large-scale ATOMIC-like knowledge graph by selecting the tuples scored by BertSAGE over 0.5. Also, we present the acquisition results of DISCOS under the setting of COMET, i.e., given h and r to generate t. The new knowledge graph can be downloaded here.

The 3.4M if-then knowledge is populated using the whole graph of ASER-core, without the neighbor filtering. You may find the processed training graph and inference candidates here.