This folder contains a viewer tool for written Dutch image descriptions, collected for our replication paper: On task effects in NLG corpus elicitation: a replication study using mixed effects modeling. This tool allows you to look through the images and their descriptions in your browser.
If you use this code or data, please cite:
@inproceedings{miltenburg2019task,
Author = {Emiel van Miltenburg and Merel van de Kerkhof and Ruud Koolen and Martijn Goudbeek and Emiel Krahmer},
Booktitle = {Proceedings of the 12th International Conference on Natural Language Generation},
Publisher = {ACL},
Title = {On task effects in NLG corpus elicitation: a replication study using mixed effects modeling},
Year = {2019}}
Here are the versions that I have used. Others probably work, but remain untested.
- Python 3.6.6
- Flask 1.0.2
Usage is fairly straightforward. The Python script will start up a server on localhost. You can visit the locally hosted website in your browser.
- Run
python view_descriptions.py - Visit http://127.0.0.1:5002/ in your browser.
The viewer requires an active internet connection as long as you haven't downloaded the images. It will automatically download each image upon viewing the descriptions for that image.
If you want to download all images first, run the following command:
python download_images.py
The tool currently uses URLs from MS COCO. If these ever stop working, try the Visual Genome URLs instead.
All the relevant data files to run this tool are stored in ./data/. These are:
images.csv- This file contains identifiers and URLs of all the images in MS COCO and the Visual Genome dataset.written_annotations.json- This file contains all the written descriptions collected in our experiment.