Unsupervised Translation of Emergent Communication

Emergent Communication (EC) provides a unique window into the language systems that emerge autonomously when agents are trained to jointly achieve shared goals. However, it is difficult to interpret EC and evaluate its relationship with natural languages (NL). This study employs unsupervised neural machine translation (UNMT) techniques to decipher ECs formed during referential games with varying task complexities, influenced by the semantic diversity of the environment. Our findings demonstrate UNMT's potential to translate EC, illustrating that task complexity characterized by semantic diversity enhances EC translatability, while higher task complexity with constrained semantic variability exhibits pragmatic EC, which, although challenging to interpret, remains suitable for translation. This research marks the first attempt, to our knowledge, to translate EC without the aid of parallel data. Introduction Emergent communication (EC) describes the phenomenon in which AI agents develop communication protocols to achieve shared goals (Giles and Jim 2003; Kasai, Tenmoto, and Kamiya 2008; Boldt and Mortensen 2024; Brandizzi 2023) . This capacity has garnered attention due to its significant potential for understanding the complexities of language formation and evolution within multi-agent systems. Yet, ECs remain largely opaque, difficult to interpret and translate into human-readable forms. Although efforts to understand EC have been made, interpretability remains elusive. Traditional approaches, such as topographic similarity (Brighton and Kirby 2006) , measure the correlation between message and input distances, provide a coarse measure of EC structures, but do not capture the full nuances of EC compositionality. Recent advancements aim to use natural language (NL) to analyze EC (Xu, Niethammer, and Raffel 2022; Carmeli, Belinkov, and Meir 2024; Chaabouni et al. 2020a). Another line of research aims to translate EC into NL using parallel data (EC-NL pairs) to gain a more intuitive understanding (Andreas, Dragan, and Klein 2017; Yao et al. 2022) . However, this approach may introduce certain biases in the translation process.