uier / fact2vec

Introduction

Fact2Vec is a part of the Erato system (https://erato.idvxlab.com/project/). It is a pretrained embedding model used for coverting a data fact (i.e., a data story piece) into a vector presentation that captures the semantics of the fact. Those logically related data facts will be lined up and close to each other in the vector space. Details about the model is described in the following paper:

@article{sun2022erato,
    title={Erato: Cooperative Data Story Editing via Fact Interpolation},
    author={Sun, Mengdi and Cai, Ligan and Cui, Weiwei and Wu, Yanqiu and Shi, Yang and Cao, Nan},
    journal = {IEEE Transactions on Visualization and Computer Graphics},
    year = {2022},
    publisher={IEEE}
}

Training Corpus

The model was trained based on 100 high-quality data stories that were manually authored by experts using the Calliope system (https://datacalliope.com). All of these stories consist of 5 data facts with diverse fact types. They were designed by following either the time-oriented narrative structure or the parallel structure. 300 fact trigrams were extracted from these stories as the training set. Each of them consisted of 3 succeeding data facts in the original story. The data are available at (dataset/storypieces.csv)

Use the pretrained model

on GPU

import torch
net = torch.load('fact2vec.pth')
embedding = net(fact)

on CPU

import torch
net = torch.load('fact2vec.pth',map_location='cpu')
net = net.to(torch.device('cpu'))
torch.no_grad()
embedding = net(fact)

Requirements and Dependencies

• Python 3.7.3
• Pytorch 1.9.0
• Transformers 4.9.1
• Sentence-transformers 2.0.0
• CUDA ≥10.2 (Optional)