• Data formatted and processed in preprocessed.[train|val|test].tsv include the following fields:

  • "title" of the papper
  • "entities" extracted
  • "entities type" following the order of "entities"
  • "text" is the equivalent text with placeholders of entities
  • the last part is the appearance order of the "entities", although the authors stated in this that the field is not used

• lastGraph2_sents.tsv and order preprocess* seem to be the same, but only the later is used in code

This repository contains the source code of our paper, Text Generation from Knowledge Graphs with Graph Transformers, which is accepted for publication at NAACL 2019.

Training:

python3.6 train.py -save <DIR>

Use --help for a list of all training options.

To generate, use

python3.6 generator.py -save <SAVED MODEL>

with the appropriate model flags used to train the model

To evaluate, run

python3.6 eval.py <GENERATED TEXTS> <GOLD TARGETS>

The AGENDA dataset is available in a user-friendly json format in /data/unprocessed.tar.gz Preprocessed data is also available in /data.

If this work is useful in your research, please cite our paper.

@inproceedings{koncel2019text,
  title={{T}ext {G}eneration from {K}nowledge {G}raphs with {G}raph {T}ransformers},
  author={Rik Koncel-Kedziorski, Dhanush Bekal, Yi Luan, Mirella Lapata, and Hannaneh Hajishirzi},
  booktitle={NAACL},
  year={2019}
}