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

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4428–4437 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 4428 Internal and External Pressures on Language Emergence: Least Effort, Object Constancy and Frequency Diana Rodr´ıguez Luna University of Amsterdam diana.rodrigzluna@gmail.com Edoardo Maria Ponti LTL, University of Cambridge ep490@cam.ac.uk Dieuwke Hupkes∗ ILLC, University of Amsterdam d.hupkes@uva.nl Elia Bruni∗ Universitat Pompeu Fabra elia.bruni@gmail.com Abstract In previous work, artificial agents were shown to achieve almost perfect accuracy in referen- tial games where they have to communicate to identify images. Nevertheless, the result- ing communication protocols rarely display salient features of natural languages, such as compositionality. In this paper, we propose some realistic sources of pressure on commu- nication that avert this outcome. More specif- ically, we formalise the principle of least ef- fort through an auxiliary objective. More- over, we explore several game variants, in- spired by the principle of object constancy, in which we alter the frequency, position, and luminosity of the objects in the images. We perform an extensive analysis on their effect through compositionality metrics, diagnostic classifiers, and zero-shot evaluation. Our find- ings reveal that the proposed sources of pres- sure result in emerging languages with less re- dundancy, more focus on high-level concep- tual information, and better abilities of gener- alisation. Overall, our contributions reduce the gap between emergent and natural languages. 1 Introduction One of the key requirements for a machine to be intelligent is its ability to communicate in natu- ral language (Mikolov et al., 2018). While super- vised approaches with labelled texts have recently achieved unprecedented performances in several applications (Chen et al., 2017, inter alia), they still neglect fundamental components of natural communication, such as the speakers’ intention and the function of their utterances (Clark, 1996). This functional aspect of language instead is captured by multi-agent games (Kirby, 2002), in which agents have to communicate about some shared input space (e.g. images). Agents usually ∗Shared senior authorship manage to communicate with success, measured in terms of task accuracy (Mordatch and Abbeel, 2018; Cohoi et al., 2018, inter alia), if the set- ting is fully cooperative (Cao et al., 2018). How- ever, several studies have shown that the emerged languages rarely display features inherent to natu- ral languages, such as compositionality of mean- ing1 and generalisation to novel objects (e.g. Kot- tur et al., 2017; van der Wal et al., 2020). For in- stance, agents might develop protocols to refer to specific pixel values, rather than concept-level in- formation (Bouchacourt and Baroni, 2018). Referential games are a perfect controlled envi- ronment to study how sources of pressure on the agents affect the ‘naturalness’ of emergent lan- guages. Previous work has proposed to limit the memory of neural agents across turns of dialogue (Kottur et al., 2017) or to soft-constrain the ac- tive vocabulary size (Mordatch and Abbeel, 2018). However, these constraints seem at odds with the capacity of human memory. In this work, we pro- pose a set of yet unexplored but more realistic …

Content truncated to 500 words