TexRel: a Green Family of Datasets for Emergent Communication with Relations

Hugh Perkins

Keywords: emergent communication, compositionality, metrics, language model, dataset, green, relations

Abstract:

We propose a new dataset TexRel as a playground for the study of emergent communications, in particular for relations.
TexRel provides rapid training and experimentation, whilst being sufficiently large to avoid overfitting.
We use TexRel to compare Sender models, compare with a related dataset, examine the effect of changing meaning space size, and perform a case-study for using \sys/ in place of symbolic input.

One-sentence Summary: TexRel is a dataset for the study of emergent communications, in particular for relations.