Internal and external pressures on language emergence: least effort, object constancy and frequency

In previous work, artificial agents were shown to achieve almost perfect\naccuracy in referential games where they have to communicate to identify\nimages. Nevertheless, the resulting communication protocols rarely display\nsalient features of natural languages, such as compositionality. In this paper,\nwe propose some realistic sources of pressure on communication that avert this\noutcome. More specifically, we formalise the principle of least effort through\nan auxiliary objective. Moreover, we explore several game variants, inspired by\nthe principle of object constancy, in which we alter the frequency, position,\nand luminosity of the objects in the images. We perform an extensive analysis\non their effect through compositionality metrics, diagnostic classifiers, and\nzero-shot evaluation. Our findings reveal that the proposed sources of pressure\nresult in emerging languages with less redundancy, more focus on high-level\nconceptual information, and better abilities of generalisation. Overall, our\ncontributions reduce the gap between emergent and natural languages.\n