"Knowledge Is Flat: A Seq2Seq Generative framework For Various Knowledge Graph Completion".

• Compatible with PyTorch 1.11.0 and Python 3.x.
• Dependencies can be installed using requirements.txt.

• We use WN18RR, FB15k-237N, FB15k-237, ICEWS14 and NELL-One dataset for knowledge graph link prediction.
• All the preprocessed data are included in the ./data/processed/ directory. Alternatively, you can download the raw dataset into ./data/raw/ and run ./preprocess.sh to generate the processed data.
• Raw data source:
  • WN18RR, FB15k-237
  • ICEWS14
  • reformatted NELL-One

To enable a quick evaluation, we upload the trained model. Download the checkpoint folders to ./checkpoint/, and run the evaluation commandline for corresponding dataset.

The results are:

• Install all the requirements from ./requirements.txt.

• Commands for reproducing the reported results:

  WN18RR

  python3 main.py -dataset 'WN18RR' \
                  -lr 0.001 \
                  -epoch 100 \
                  -batch_size 64 \
                  -src_descrip_max_length 40 \
                  -tgt_descrip_max_length 10 \
                  -use_soft_prompt \
                  -use_rel_prompt_emb \
                  -seq_dropout 0.1 \
                  -num_beams 40 \
                  -eval_tgt_max_length 30 \
                  -skip_n_val_epoch 30 \
  
  # evaluation commandline:
  python3 main.py -dataset 'WN18RR' \
                  -src_descrip_max_length 40 \
                  -tgt_descrip_max_length 10 \
                  -use_soft_prompt \
                  -use_rel_prompt_emb \
                  -num_beams 40 \
                  -eval_tgt_max_length 30 \
                  -model_path path/to/trained/model \
                  -use_prefix_search

  FB15k-237N

  python3 main.py -dataset 'FB15k-237N' \
                  -lr 0.001 \
                  -epoch 50 \
                  -batch_size 32 \
                  -src_descrip_max_length 80 \
                  -tgt_descrip_max_length 80 \
                  -use_soft_prompt \
                  -use_rel_prompt_emb \
                  -seq_dropout 0.2 \
                  -num_beams 40 \
                  -eval_tgt_max_length 30 \
                  -skip_n_val_epoch 30
  
  # evaluation commandline:
  python3 main.py -dataset 'FB15k-237N' \
                  -src_descrip_max_length 80 \
                  -tgt_descrip_max_length 80 \
                  -use_soft_prompt \
                  -use_rel_prompt_emb \
                  -num_beams 40 \
                  -eval_tgt_max_length 30 \
                  -model_path path/to/trained/model \
                  -use_prefix_search  

  FB15k-237

  python3 main.py -dataset 'FB15k-237' \
                  -lr 0.001 \
                  -epoch 40 \
                  -batch_size 32 \
                  -src_descrip_max_length 80 \
                  -tgt_descrip_max_length 80 \
                  -use_soft_prompt \
                  -use_rel_prompt_emb \
                  -seq_dropout 0.2 \
                  -num_beams 40 \
                  -eval_tgt_max_length 30 \
                  -skip_n_val_epoch 20
  
  # evaluation commandline:
  python3 main.py -dataset 'FB15k-237' \
                  -src_descrip_max_length 80 \
                  -tgt_descrip_max_length 80 \
                  -use_soft_prompt \
                  -use_rel_prompt_emb \
                  -num_beams 40 \
                  -eval_tgt_max_length 30 \
                  -model_path path/to/trained/model \
                  -use_prefix_search 

  ICEWS14

  python3 main.py -dataset 'ICEWS14' \
                  -lr 0.0005 \
                  -epoch 100
                  -batch_size 32 \
                  -src_descrip_max_length 40 \
                  -tgt_descrip_max_length 40 \
                  -temporal  \
                  -use_soft_prompt \
                  -use_rel_prompt_emb \
                  -seq_dropout 0.1 \ 
                  -num_beams 40 \
                  -eval_tgt_max_length 26 \
                  -skip_n_val_epoch 50
  
  # evaluation commandline:
  python3 main.py -dataset 'ICEWS14' \
                  -src_descrip_max_length 40 \
                  -tgt_descrip_max_length 40 \
                  -temporal  \
                  -use_soft_prompt \
                  -use_rel_prompt_emb \
                  -num_beams 40 \
                  -eval_tgt_max_length 26 \
                  -model_path path/to/trained/model \
                  -use_prefix_search  

  NELL-One

  python3 main.py -dataset 'NELL' \
                  -lr 0.0005 \
                  -epoch 30 \
                  -batch_size 128 \
                  -src_descrip_max_length 0 \
                  -tgt_descrip_max_length 0 \
                  -use_soft_prompt \
                  -use_rel_prompt_emb \
                  -num_beams 40 \
                  -eval_tgt_max_length 25 \
                  -skip_n_val_epoch 15
  
  
  # evaluation commandline:
  python3 main.py -dataset 'NELL' \
                  -src_descrip_max_length 0 \
                  -tgt_descrip_max_length 0 \
                  -use_soft_prompt \
                  -use_rel_prompt_emb \
                  -num_beams 40 \
                  -eval_tgt_max_length 25 \
                  -model_path path/to/trained/model \
                  -use_prefix_search  

  • -src_descrip_max_length denotes the training batch size
  • -src_descrip_max_length denotes the maximum description length for source entity during training
  • -tgt_descrip_max_length denotes the maximum description length for target entity during training
  • -eval_tgt_max_length denotes the maximum description length for generation during inference
  • -use_soft_prompt denotes the option whether to use soft prompt
  • -use_rel_prompt_emb denotes the option whether to use relation-specific soft prompt (need to enable -use_soft_prompt)
  • -seq_dropout denotes the value for sequence dropout
  • -use_prefix_search denotes to use constrained decoding method
  • -temporal denotes the dataset is for temporal knowledge graph completion
  • -skip_n_val_epoch denotes the number of training epochs without evaluation (evaluation is costly due to the auto-regressive decoding)

If you used our work or found it helpful, please use the following citation:

@inproceedings{KG_S2S,
    title = "Knowledge Is Flat: A Seq2Seq Generative Framework for Various Knowledge Graph Completion",
    author = "Chen, Chen  and
      Wang, Yufei  and
      Li, Bing  and 
      Lam, Kwok-Yan",
    booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
}