This project generates icon language models from corpus-less icon set. This projects provides the code used for the process and experiments described in Compositional Language Modeling for Icon-Based Augmentative and Alternative Communication written by Dudy and Bedrick. In the paper you can find more details about the motivation and the context for this research.
@inproceedings{dudy-bedrick-2018-compositional,
title = "Compositional Language Modeling for Icon-Based Augmentative and Alternative Communication",
author = "Dudy, Shiran and
Bedrick, Steven",
booktitle = "Proceedings of the Workshop on Deep Learning Approaches for Low-Resource {NLP}",
month = jul,
year = "2018",
address = "Melbourne",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/W18-3404",
doi = "10.18653/v1/W18-3404",
pages = "25--32"
}
The code was written by Shiran Dudy.
Additional requirements to run the code are:
- python3
- nltk for python
- An icons' meta data file. Each icon should have its name and synonyms associated with it in the file. Here, I assume the XML file is the icons metadata file of Symbolstix.
- A corpus that contains textual sentences in which the icons (name/synonyms) are found.
- A pretrained word embedding set. The icon terms (mainly the synonym should be represented in this set).
Optional, but recommended is to follow the fast version of the process (fond in the appropriate folder). There you will be required to have
- slurm scheduler
- gpus
The dataprep part of the code generates the simulated icon corpus. This part requires slurm for the fast_version.
The modeling part trains and tests the icon language models. This part requires slurm and gpu for the fast_version
The modeling part can be replaced by any alternative language modeling learning method. I slightly modified the basic approach borrowed from word langauge modeling and adapted it to large vocabularies using nce.
