Skip to content


Repository files navigation

Multi-Hop Graph Relation Networks (EMNLP 2020)

License: MIT

This is the repo of our EMNLP'20 paper:

Scalable Multi-Hop Relational Reasoning for Knowledge-Aware Question Answering
Yanlin Feng*, Xinyue Chen*, Bill Yuchen Lin, Peifeng Wang, Jun Yan and Xiang Ren.
EMNLP 2020.
*=equal contritbution

This repository also implements other graph encoding models for question answering (including vanilla LM finetuning).

  • RelationNet
  • R-GCN
  • KagNet
  • GConAttn
  • KVMem
  • MHGRN (or. MultiGRN)

Each model supports the following text encoders:

  • LSTM
  • GPT
  • BERT
  • XLNet
  • RoBERTa


We provide preprocessed ConceptNet and pretrained entity embeddings for your own usage. These resources are independent of the source code.

Note that the following reousrces can be download here.

ConceptNet (5.6.0)

Description Downloads Notes
Entity Vocab entity-vocab one entity per line, space replaced by '_'
Relation Vocab relation-vocab one relation per line, merged
ConceptNet (CSV format) conceptnet-5.6.0-csv English tuples extracted from the full conceptnet with merged relations
ConceptNet (NetworkX format) conceptnet-5.6.0-networkx NetworkX pickled format, pruned by filtering out stop words

Entity Embeddings (Node Features)

Entity embeddings are packed into a matrix of shape (#ent, dim) and stored in numpy format. Use np.load to read the file. You may need to download the vocabulary files first.

Embedding Model Dimensionality Description Downloads
TransE 100 Obtained using OpenKE with optim=sgd, lr=1e-3, epoch=1000 entities relations
NumberBatch 300 entities
BERT-based 1024 Provided by Zhengwei entities


Run the following commands to create a conda environment (assume CUDA10):

conda create -n krqa python=3.6 numpy matplotlib ipython
source activate krqa
conda install pytorch=1.1.0 torchvision cudatoolkit=10.0 -c pytorch
pip install dgl-cu100==0.3.1
pip install transformers==2.0.0 tqdm networkx==2.3 nltk spacy==2.1.6
python -m spacy download en


1. Download Data

First, you need to download all the necessary data in order to train the model:

git clone
bash scripts/

The script will:

2. Preprocess

To preprocess the data, run:


By default, all available CPU cores will be used for multi-processing in order to speed up the process. Alternatively, you can use "-p" to specify the number of processes to use:

python -p 20

The script will:

  • Convert the original datasets into .jsonl files (stored in data/csqa/statement/)
  • Extract English relations from ConceptNet, merge the original 42 relation types into 17 types
  • Identify all mentioned concepts in the questions and answers
  • Extract subgraphs for each q-a pair

The preprocessing procedure takes approximately 3 hours on a 40-core CPU server. Most intermediate files are in .jsonl or .pk format and stored in various folders. The resulting file structure will look like:

└── data/
    ├── cpnet/                 (prerocessed ConceptNet)
    ├── glove/                 (pretrained GloVe embeddings)
    ├── transe/                (pretrained TransE embeddings)
    └── csqa/
        ├── train_rand_split.jsonl
        ├── dev_rand_split.jsonl
        ├── test_rand_split_no_answers.jsonl
        ├── statement/             (converted statements)
        ├── grounded/              (grounded entities)
        ├── paths/                 (unpruned/pruned paths)
        ├── graphs/                (extracted subgraphs)
        ├── ...

3. Hyperparameter Search (optional)

To search the parameters for RoBERTa-Large on CommonsenseQA:

bash scripts/ csqa roberta-large

To search the parameters for BERT+RelationNet on CommonsenseQA:

bash scripts/ csqa bert-large-uncased

4. Training

Each graph encoding model is implemented in a single script:

Graph Encoder Script Description
None w/o knowledge graph
Relation Network
R-GCN Use --gnn_layer_num and --num_basis to specify #layer and #basis
KagNet Adapted from, still tuning

Some important command line arguments are listed as follows (run python {lm,rn,rgcn,...}.py -h for a complete list):

Arg Values Description Notes
--mode {train, eval, ...} Training or Evaluation default=train
-enc, --encoder {lstm, openai-gpt, bert-large-unased, roberta-large, ....} Text Encoer Model names (except for lstm) are the ones used by huggingface-transformers, default=bert-large-uncased
--optim {adam, adamw, radam} Optimizer default=radam
-ds, --dataset {csqa, obqa} Dataset default=csqa
-ih, --inhouse {0, 1} Run In-house Split default=1, only applicable to CSQA
--ent_emb {transe, numberbatch, tzw} Entity Embeddings default=tzw (BERT-based node features)
-sl, --max_seq_len {32, 64, 128, 256} Maximum Sequence Length Use 128 or 256 for datasets that contain long sentences! default=64
-elr, --encoder_lr {1e-5, 2e-5, 3e-5, 6e-5, 1e-4} Text Encoder LR dataset specific and text encoder specific, default values in utils/
-dlr, --decoder_lr {1e-4, 3e-4, 1e-3, 3e-3} Graph Encoder LR dataset specific and model specific, default values in {model}.py
--lr_schedule {fixed, warmup_linear, warmup_constant} Learning Rate Schedule default=fixed
-me, --max_epochs_before_stop {2, 4, 6} Early Stopping Patience default=2
--unfreeze_epoch {0, 3} Freeze Text Encoder for N epochs model specific
-bs, --batch_size {16, 32, 64} Batch Size default=32
--save_dir str Checkpoint Directory model specific
--seed {0, 1, 2, 3} Random Seed default=0

For example, run the following command to train a RoBERTa-Large model on CommonsenseQA:

python --encoder roberta-large --dataset csqa

To train a RelationNet with BERT-Large-Uncased as the encoder:

python --encoder bert-large-uncased

To reproduce the reported results of MultiGRN on CommonsenseQA official set:

bash scripts/

5. Evaluation

To evaluate a trained model (you need to specify --save_dir if the checkpoint is not stored in the default directory):

python {lm,rn,rgcn,...}.py --mode eval [ --save_dir path/to/directory/ ]

Use Your Own Dataset

  • Convert your dataset to {train,dev,test}.statement.jsonl in .jsonl format (see data/csqa/statement/train.statement.jsonl)
  • Create a directory in data/{yourdataset}/ to store the .jsonl files
  • Modify and perform subgraph extraction for your data
  • Modify utils/ to support your own dataset
  • Tune encoder_lr,decoder_lr and other important hyperparameters, modify utils/ and {model}.py to record the tuned hyperparameters


Scalable Multi-Hop Relational Reasoning for Knowledge-Aware Question Answering (EMNLP 2020)






No packages published

Contributors 4