The E2E NLG module leverages the seq2seq framework for end-to-end natural language generation from meaning representations (MRs). This is a work in progress...
In the e2e_nlg folder, optionally put your input files in the data folder. Run main.py in one of the following ways to run the training, or the evaluation (only after the training has been run):
python run_task.py --train [path_to_trainset] [path_to_devset]
python run_task.py --test [path_to_testset]
Replace [path_to_trainset], [path_to_devset], [path_to_testset] with relative paths to your trainset, devset, or testset, respectively. They are expected to be CSV files with two columns (their headers must be mr and ref, respectively), the first containing the MRs, and the second containing the corresponding reference utterances.
Once the training is done, the model folder will contain files describing the model, which will be used for evaluation. Therefore, you are not to modify them.
Finally, the evaluation produces output files in the predictions folder. The predictions.txt file contains raw results, while the predictions_final.txt is produced during the postprocessing step.
- Python libraries: tensorflow, numpy, pandas, nltk, networkx
- NLTK modules: perluniprops, punkt
- install using the following command: python -c "import nltk; nltk.download('[module_name]')"