TCube: Domain-Agnostic Neural Time series Narration

This repository contains the code for the paper: "TCube: Domain-Agnostic Neural Time series Narration" (to appear in IEEE ICDM 2021).

The PLMs used in this effort (T5, BART, and GPT-2) are implemented using the HuggingFace library (https://huggingface.co/) and finetuned to the WebNLG v3 (https://gitlab.com/shimorina/webnlg-dataset/-/tree/master/release_v3.0) and DART (https://arxiv.org/abs/2007.02871) datasets.

Clones of both datasets are available under /Finetune PLMs/Datasets in this repository.

The PLMs fine-tuned to WebNLG/DART could not be uploaded due to the 1GB limitations of GitLFS. However, pre-made scripts in this repository (detailed below) are present for convientiently fine-tuning these models.

The entire repository is based on Python 3.6 and the results are visaulized through the iPython Notebooks.

Dependencies

Interactive Environments

• notebook
• ipywidgets==7.5.1

Deep Learning Frameworks

• torch 1.7.1 (suited to your CUDA version)
• pytorch-lightning 0.9.0
• transformers==3.1.0

NLP Toolkits

• sentencepiece==0.1.91
• nltk

Scientific Computing, Data Manipulation, and Visualizations

• numpy
• scipy
• sklearn
• matplotib
• pandas
• pwlf

Evaluation

• rouge-score
• textstat
• lexical_diversity
• language-tool-python

Misc

• xlrd
• tqdm
• cython

Please make sure that the aforementioned Python packages with their specified versions are installed in your system in a separate virtual environment.

Data-Preprocessing Scripts

Under /Finetune PLMs in this repository there are two scripts for pre-processing the WebNLG and DART datasets:

preprocess_webnlg.py
preprocess_dart.py

These scripts draw from the original datasets in /Finetune PLMs/Datasets/WebNLGv3 and /Finetune PLMs/Datasets/DART and prepare CSV files in /Finetune PLMs/Datasets breaking the original datasets into train, dev, and test sets in the format required by our PLMs.

Fine-tuning Scripts

Under /Finetune PLMs in this repository there are three scripts for fine-tuning T5, BART, and GPT-2:

finetuneT5.py
finetuneBART.py
finetuneGPT2.py

Visualization and Evaluation Notebooks

In the root directory are 10 notebooks. For the descriptions of the time-series datasets used:

Datatsets.ipynb

For comparisons of segmentation and regime-change detection algorithms:

Error Determination.ipynb
Regime Detection.ipynb
Segmentation.ipynb
Trend Detection Plot.ipynb

For the evaluation of the TCube framework on respective time-series datasets:

T3-COVID.ipnyb
T3-DOTS.ipnyb
T3-Pollution.ipnyb
T3-Population.ipnyb
T3-Temperature.ipnyb

Citation and Contact

If any part of this code repository or the TCube framework is used in your work, please cite our paper. Thanks!

Contact: Mandar Sharma ([email protected]), First Author.