Emerging Grounded Shared Vocabularies Between Human and Machine, Inspired by Human Language Evolution

摘要

Building conversational AI systems has the goal to teach machines to understand human language and respond naturally. The most common way to train agents to produce and interpret natural language is currently by exposing them to large quantities of data. Although this has resulted in advances in many areas, these systems typically have little understanding of how language is related to the real world (Mordatch and Abbeel, 2018), known as the grounding problem. Also, most conversational agents are trained in isolation, while humans are social animals, deeply embedded in culture and surrounded by others. Complex human behaviors, like language, evolved in socio–cultural contexts and could not exist without a variety of minds using and transmitting these behaviors.To overcome this problem, researchers in Computational Linguistics have started modeling emerging communication setups, in which novel signals are created by interacting agents (Lazaridou et al., 2018; Mordatch and Abbeel, 2018; Chaabouni et al., 2019; ter Hoeve et al., 2021). However, the findings in such models do not always match what is found in similar experiments with humans, and features found in human language often do not emerge (Lazaridou et al., 2020).The mechanisms that influence the emergence of communication and linguistic structure have been studied in the field of Language Evolution. Although the precise origins of human language are widely debated, computer simulations (Boer, 2006; Kirby, 2017; Steels, 2012a) and experiments in which humans use novel communication signals (Scott-Phillips and Kirby, 2010; Galantucci and Garrod, 2010; Kirby et al., 2014), have revealed some key mechanisms that drive the initial emergence of a novel language and the gradual appearance of more complex linguistic structure. We review relevant findings and proposeto apply methods that confirm the importance of including micro–societies of interacting minds to the emergence of novel human–machine communication systems.A major insight from these studies is that language adapts to human biases and how it is learned and used(Kirby et al., 2014, 2015). Similarly, current language models also exhibit biases, free order case-marking languages are for example more challenging to model than fixed-order languages (Bisazza et al., 2021). As such, we suggest that language used in human–machine communication should also evolve more naturally, resulting in a grounded communication system adapted to biases and constraints of human and machine learning. We moreover emphasize the importance of co–development of shared vocabularies by conversational partners (human or AI–based). Doing so might result in a dynamic communication system that is natural to humans and artificial conversational agents. We propose to follow a process of several steps, displayed in figure 1. Starting from random behaviors, a signal–meaning mapping emerges from shared interactions (section 2) which become more structured through horizontal and vertical transmission (section 3) and eventually evolve into an adaptive communication system (sections 4 & 5).Successful communication happens when the coordinated actions of all participants adhere to the grounding criterion: that interlocutors agree that they have understood what was meant for the current purposes (Clark and Brennan, 1991). This requires a vocabulary that is (partially) aligned between interlocutors of a conversation (Pickering and Garrod, 2004). The emergence of which starts with agreeing on what kind of (initially random) behaviors should be interpreted as communicative and what they refer to (box 1 & 2 in figure 1).Experiments with human participants have been conducted to study the emergence of novel communication forms (Galantucci, 2005; Steels, 2006; Scott-Phillips et al., 2009; Galantucci and Garrod, 2010). Here, participants need to invent and negotiate novel signals to solve a communicative or cooperative task. Albeit often bound to the startin