Game-Based Video-Context Dialogue

This repository contains Twitch-FIFA dataset and code for: Game-Based Video-Context Dialogue

Twitch-FIFA dataset is collected from twitch.tv platform and is subject to research-only purposes in accordance with the Twitch privacy policy. Note that the included MIT license only applies to our code and models.

Twitch-FIFA Dataset Download - Google Drive

• You can download the full processed dataset along with inception-v3 frame-level videos features from here.
• You can download the raw videos at 3fps from here.

Dataset Statistics:

Code and Models

This code is tested on python 2.7 and tensorflow 1.2.

Setup:

Install all the required packages from requirements.txt file.

create "data" directory and place the downloaded processed data in that directory. The structure should look like as descibed below:

.
├── ...
├── data                   
│   ├── fifa        
│   │    ├── train_word.json
│   │    ├── ...                
└── ...

Run Code:

use run.sh script to train and test models

Pretrained Models

Download the pretrained models for TriDAF discriminative model from here.

Citation

If you find this dataset and code helpful, please consider citing the following paper:

@inproceedings{pasunuru2018gamebased,
    title={Game-Based Video-Context Dialogue},
    author={Pasunuru, Ramakanth and Bansal, Mohit},
    booktitle={EMNLP},
    year={2018}
}