WHAT PRESERVES THE EMERGENCE OF LANGUAGE? Anonymous authors Paper under double-blind review

Abstract

The emergence of language is a mystery. One dominant theory is that cooperation boosts language to emerge. However, as a means of giving out information, language seems not to be an evolutionarily stable strategy. To ensure the survival advantage of many competitors, animals are selfish in nature. From the perspective of Darwinian, if an individual can obtain a higher benefit by deceiving the other party, why not deceive? For those who are cheated, once bitten and twice shy, cooperation will no longer be a good option. As a result, motivation for communication, as well as the emergence of language would perish. Then, what preserves the emergence of language? We aim to answer this question in a brand new framework of agent community, reinforcement learning, and natural selection. Empirically, we reveal that lying indeed dispels cooperation. Even with individual resistance to lying behaviors, liars can easily defeat truth tellers and survive during natural selection. However, social resistance eventually constrains lying and makes the emergence of language possible.

1. INTRODUCTION

Unveiling the principles behind the emergence and evolution of language is attractive and appealing to all. It is believed that this research field is of great significance for promoting the development of enabling agents to evolve an efficient communication protocol (Nowak & Krakauer, 1999; Kottur et al., 2017; Chaabouni et al., 2019) or acquire existing one (Li & Bowling, 2019) , especially when interacting with humans. Previously, many studies have investigated some intriguing properties of language and their effects on the emergence of language (Andreas & Klein, 2017; Lazaridou et al., 2018; Mordatch & Abbeel, 2018) . The motivation behind these is that human language is considered as a remarkable degree of structure and complexity (Givon, 2013) and each character is the result of evolution, thus they believe that understanding the language itself is an indispensable step to take. Unlike existing work, we, from a different perspective, focus on a fundamental question that what made the emergence of language possible during evolution. One of the dominant theories in the community of emergent communication is: cooperation boosts language to emerge (Nowak & Krakauer, 1999; Cao et al., 2018) . Hence, there has been a surge of work investigating this field in cooperative multi-agent (mostly two agents) referential games (Lazaridou & Peysakhovich, 2017; Kottur et al., 2017; Das et al., 2017; Evtimova et al., 2018; Lazaridou et al., 2018) , a variant of the Lewis signaling game (David, 1969) . However, they seem to miss some basic elements in the human language. On one hand, human language emerges from the community, not just two persons, after all, language is learnable and can spread from one place to other (Dagan et al., 2020) . Studying a language in two-player games is like looking at the world through a keyhole. On the other hand, many works make an agreement that prior to the emergence of language some pre-adaptations occurred in the hominid lineage, and one of the candidates is the ability to use symbols (Deacon, 2003; Davidson, 2003; Christiansen & Kirby, 2003) . It seems understanding the emergence of symbolic signals is the key to approach the truth of the origin of language (Deacon, 1998) . However, chimpanzees have demonstrated a degree of language capacity by using arbitrary symbols as well as the ability for the cross-modal association, abstract thought, and displacement of thought in time (Meddin, 1979) . So why don't they have a language like us? One of the theory is selfishness has kept animal communication at a minimum (Ulbaek, 1998) . In more detail, if an individual can obtain a higher benefit by deceiving the other party in the cooperation, why not deceive? Once deception emerges, mistrust among individuals will haunt. For those who are cheated, once bitten and twice shy, cooperation will no longer be a good option. As a result, motivation for communication, as well as demands of the emergence of language will perish. But human beings are so special since we have overcome this kind of obstacle and evolved language. Then, what preserves the emergence of language? We aim to answer this question in a brand new framework of agent community, reinforcement learning (RL), and natural selection. We believe this process should occur in the pre-language period since lying is possible as long as agents can communicate. Therefore, our investigating communication protocol uses symbols to transmit meaning based on a social convention or implicit agreement. In this paper, we introduce several agents to form one community and allow natural evolution and elimination among them. Both liars (agents tell lies) and truth tellers (agents always tell truth) exist in the community. Each tournament, every agent is supposed to play a non-cooperative game (Nash Jr, 1950; Nash, 1951; Schelling, 1958; Binmore et al., 1986; Von Neumann & Morgenstern, 2007) with others. In our multi-round bargaining game, agents are required to reach an agreement about how many items to give out so that the total quantity can satisfy the market's demand and they can keep their loss to the minimum in the meantime. We believe this is a perfect fit for the nature of human beings and more common in the hominid lineage compared with the cooperation game. Importantly, during the process of natural selection, the fraction of liars and truth tellers may change from time to time and this allows us to observe what factor imposes influence on the motivation of communication, which is the prerequisite of the emergence of language. It is worthy of note that pre-language communication was subject to the constraints of Darwinian evolution. While linguistic change, which began in the post-language communicative era of hominid evolution, is by and large tied to society and culture (Givón & Malle, 2002; Li & Hombert, 2002) . Thus, we disregard the factors related to linguistic change since we are investigating motivation for communication from which language evolved. Moreover, apart from the normal setting mentioned above, we add up two more rules to further dig out. Firstly, we introduce a credit mechanism for truth tellers. In other words, we make sure truth tellers know the existence of liars which is one step in the evolution process. Specifically, every liar has credit in the mind of truth teller, and the credit varies with the profit of truth teller. Cooperation would be impossible between two agents as soon as the credit drops to negative. Secondly, an additional penalty will be brought in as a price of lying, and we consider it as social pressure for resisting lying behaviors. All in all, we want to make a thorough investigation about how the individual or social resistance to lying affects communication. Empirically, we show that in normal settings, two truth tellers can make a fair agreement, and liars can achieve a huge advantage over truth teller by telling lies. As for two liars, there is always a better liar that gains relative more than the other. In the credit setting, liars can learn a sophisticated lying strategy that deceives the credit mechanism and makes more profits meanwhile. In both settings, as time goes on, truth tellers seem not to show enough competition against liars and thus die out. In the society setting, we find out liars are afraid of lying if punishment is sufficiently large. This again proves the theory (Ulbaek, 1998): in the human lineage, social cooperation based on obligatory reciprocal altruism (Trivers, 1971) as well as a system which punishes people morally and physically for cheating has evolved. In such an environment language is finally possible.

2.1. GAME SETTINGS

We explore emergent language in the context of multi-round non-cooperative bargaining game (Nash Jr, 1950; Nash, 1951) as illustrated in Figure 1 . The core is binding cooperative strategy with the highest profit is impossible whereas selfish strategy sometimes can. In this case, the behavior of telling lies can be meaningful since it undermines cooperation and grabs more benefits from others. In the game, two agents i, j bargain over how to satisfy the demand of market. Agents are presented with N different items. They possess a fixed number of quantity for each item ({q i n } N n=1 , {q j n } N n=1 ), and have their own hidden utilities for N items ({u i n } N n=1 , {u j n } N n=1 ). Agents move sequentially. Suppose at bargaining round t, it is turn for agent i to be proposer and it makes proposal {p i t,n } N n=1 about how many to give out to the market, which means the other agent j should contribute the rest {d n -p i t,n } N n=1 to the market, where d n is the market demand for item n. Then agent j would

