On the Word Boundaries of Emergent Languages Based on Harris's Articulation Scheme

Ryo Ueda, Taiga Ishii, Yusuke Miyao

Keywords: Emergent Communication, Emergent Language, Unsupervised Word Segmentation, Harris's Articulation Scheme

TL;DR: This paper investigates whether Harris's articulation scheme (HAS) also holds in emergent languages.

Abstract:

The purpose of this paper is to investigate whether Harris's articulation scheme (HAS) also holds in emergent languages.
HAS is thought to be a universal property in natural languages that articulatory boundaries can be obtained from statistical information of phonems alone, without referring to word meanings.
Emergent languages are artificial communication protocols that arise between agents in a simulated environment and have been attracting attention in recent years.
It is considerd important to study the structure of emergent languages and the similarity to natural languages.
In this paper, we employ HAS as an unsupervised word segmentation method and verify whether emergent languages arising from signaling games have meaningful boundaries.
Our experiments showed that the emergent languages arising from signaling games satisfy some preconditions for HAS.
However, it was also suggested that the HAS-based segmentation boundaries are not necessarily semantically valid.