README

This is the code for: RDF-to-Text Generation with Graph-augmented Structural Neural Encoders (IJCAI 2020)

WebNLG Dataset1 preprocess

• The train/dev/test data processed with entity masking can be found under directory "data/preprocess_input"
  • train/dev/test-webnlg-all-delex.triple
  • train/dev/test-webnlg-all-delex.lex
  • If you want to restore the "ent_id: ent_name" information, please use the data train/dev/test_gtrlstm.pkl
• Run "build_dataset.py" to use the processed data above to build input dataset
  • seq_train/test/dev.pt
  • gcn_train/test/dev.pt
  • gtr2_train/test/dev.pt
  • gcn_gtr2_train/test/dev.pt

Train model

• The configuration files for each model is in "utils/".
• If you want to train the bi-GCN model on WebNLG dataset1, copy the "config_webnlg1_gcn.py" to "config.py" and run "main_PG.py".

Translate

• Set the mode = 'translate' and load_model = True in config files and run "main_PG.py" to generate texts.
• The generated text can be found in "data/translate_result"

Unmask the eid in generated text

• run "data/translate_result/ent_unmask.py", the file “gcn_translate_43.txt” will be transformed to “gcn_translate_43_relex.txt”.

Evaluate the generated text

• run "webnlg_eval_scripts/webnlg_gcnonmt_relexicalise.py" using the arguments as follows:

-i
../data/
-f
../data/translate_result/gcn_translate_43_relex.txt
-p
test
-c
seen
-l

• generate a final file: “gcn_translate_43_relex_relexicalised_predictions.txt”

• run "calculate_bleu_dev.sh" to calculate BLEU scores