NaRuto: An advanced system that can automatically acquire planning models from narrative texts
NaRuto is an innovative system that initially extracts structured events from text and subsequently generates planning-language-style action models based on predictions of commonsense event relations, as well as textual contradictions and similarities in an unsupervised manner. It comprises two stages.
Python 3.7+
Packages:
transformers==3.0.2
torch==1.5.1
pytorch-lightning==0.8.1
pandas
nltk
spacy
allennlp
tensorboard
psutil
sacrebleu
rouge-score
tensorflow_datasets
faiss
streamlit
elasticsearch
nlp
torchtext
A pipeline for processing input text and returning events:
- event_extraction.py: Get event occurrences from text. Data type should be specified: movieplot or goodnews.
- resolve_coref.py: coreference resolution. Map the entity mentions that refer to previous occurred entity names.
- POS_DEP_tagging.py: POS tagging and dependency parsing via StandofordcoreNLP toolkit.
- phrasal_verb.py: Detect phrasal verbs if exist.
- subevent_detect.py: Detect argument events and conditional events from the event occurrences.
- pipeline_event_representation_acquire.py: Run this pipeline python file equals to run all the above files one by one as a pipeline.
Finetune COMET-BM:
Go to the COMET-BM folder, and run bash run.sh. For running, please specify the directory (--data_dir) where you download & save the ATOMIC-2020 data.
The results will be saved to COMET-BM/results/.
Run python COMET-BM/inference.py for predicting using the local negation generation strategy.
Additionally, run python COMET-BM/global_negation.py for generating global negated preconditions and effects.
If you feel the code helpful, please cite:
@inproceedings{li2024naruto,
title={NaRuto: Automatically Acquiring Planning Models from Narrative Texts},
author={Li, Ruiqi and Cui, Leyang and Lin, Songtuan and Haslum, Patrik},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={38},
number={18},
pages={20194--20202},
year={2024}
}