Natural Language Does Not Emerge ‘Naturally’ in Multi-Agent Dialog

By Satwik Kottur, José M. F. Moura, Stefan Lee, Dhruv Batra. 2017

239 citations recorded by the source corpus.

A number of recent works have proposed techniques for end-to-end learning of communication protocols among cooperative multi-agent populations, and have simultaneously found the emergence of grounded human-interpretable language in the protocols developed by the agents, learned without any human supervision! In this paper, using a Task & Talk reference game between two agents as a testbed, we present a sequence of ‘negative’ results culminating in a ‘positive’ one – showing that while most agent-invented languages are effective (i.e. achieve near-perfect task rewards), they are decidedly not interpretable or compositional. In essence, we find that natural language does not emerge ‘naturally’,despite the semblance of ease of natural-language-emergence that one may gather from recent literature. We discuss how it is possible to coax the invented languages to become more and more human-like and compositional by increasing restrictions on how two agents may communicate.

Subfields: Artificial Intelligence, Cultural Studies

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Cultural Studies

Citations in the corpus

Measuring Compositionality in Representation Learning (2019, background): “As in the previous section, existing work argues for a relationship between compo-sitionality and generalization, claiming that agents need compositional communication protocols to generalize to unseen referents (Kottur et al., 2017; Choi et al., 2018).”
Rethinking Action Spaces for Reinforcement Learning in End-to-end Dialog Agents with Latent Variable Models (2019, background): “havior: the response decoder deviates from human language and generates utterances that are incomprehensible (Lewis et al., 2017; Das et al., 2017; Kottur et al., 2017).”
Talk the Walk: Navigating New York City through Grounded Dialogue (2018, background): “…communication, where, agents communicate (in natural language or otherwise) in order to solve a task, with respect to their shared environment (Lazaridou et al., 2016; Das et al., 2017b; Mordatch & Abbeel, 2017; Evtimova et al., 2017; Lewis et al., 2017; Strub et al., 2017; Kottur et al., 2017).”
Evaluating Machines by their Real-World Language Use (2020, background): “However, ensuring that machines communicate in standard English is difficult, as there is usually a more efficient machine-language coding scheme for the task (Kottur et al., 2017).”
Multitasking Inhibits Semantic Drift (2021, background): “Work in emergent communication has found that reinforcement learning tends not to learn policies with natural language-like properties (Kottur et al., 2017), although population-based training has been found to be helpful (Gupta et al., 2019).”
Multi-agent Communication meets Natural Language: Synergies between Functional and Structural Language Learning (2020, background): “Chaabouni et al. (2019) show that protocols found through emergent communication, unlike natural language, do not conform to Zipf’s Law of Abbreviation; Kottur et al. (2017) find that communication protocols do not follow composiar X iv :2 00 5.”
Interactive Grounded Language Acquisition and Generalization in a 2D World (2018, background): “This suggests that neural networks tend to exploit data in an unexpected way to achieve tasks if no constraints are imposed (Kottur et al., 2017).”
TarMAC: Targeted Multi-Agent Communication (2018, background): “While (Foerster et al., 2016; Jorge et al., 2016; Das et al., 2017; Kottur et al., 2017; Mordatch & Abbeel, 2017; Lazaridou et al., 2017) constrain agents to communicate with discrete symbols with the explicit goal to study emergence of language, our work operates in the paradigm of learning a…”
Aligning Faithful Interpretations with their Social Attribution (2020, background): “Kottur et al. (2017); Subramanian et al. (2020) describe failure cases of Trojans where communicating modules encode information unintuitively, despite an interpretable interface between them.”
Learning when to Communicate at Scale in Multiagent Cooperative and Competitive Tasks (2018, background): “Lee et al. (2017) demonstrated that agents speaking different languages can learn to translate in referential games.”
Emergent Communication of Generalizations (2021, background): “The ability to communicate accurately, even on unseen concepts, is not necessarily indicative of more systematic communication; generalization without compositional language is a common ﬁnding in the literature [1, 6, 16, 22].”
Interpretable and Pedagogical Examples (2017, background): “A number of techniques have been suggested to encourage interpretability, such as limiting symbol vocabulary size (Mordatch & Abbeel, 2017), limiting memorization capabilities of the speaker (Kottur et al., 2017), or introducing auxiliary tasks such as image labelling based on supervision data (Laza”
Deep Reinforcement Learning: An Overview (2017, background): “-to-end optimization with deep RL for goal-driven and visually grounded dialogue systems for GuessWhat?! game.Das et al.(2017) proposed to learn cooperative Visual Dialog agents with deep RL. See alsoKottur et al. (2017). SeePasunuru and Bansal(2017) for video captioning. 5.5 NEURAL ARCHITECTURE DES”
Iterated learning for emergent systematicity in VQA (2021, background): “Most approaches to learning a discrete communication protocol between agents use reinforcement learning (Foerster et al., 2016; Lazaridou et al., 2017; Kottur et al., 2017; Jorge et al., 2016; Havrylov & Titov, 2017).”
Compositionality and Generalization In Emergent Languages (2020, methodology): “…of neural networks, particularly in linguistic tasks where humans rely on compositionality (e.g., Fodor and Lepore, 2002; Marcus, 2003; van der Velde et al., 2004; Brakel and Frank, 2009; Kottur et al., 2017; Lake and Baroni, 2018; Andreas, 2019; Hupkes et al., 2019; Resnick et al., 2019).”
Interactive Learning from Activity Description (2021, methodology): “Finally, our work also fundamentally differs from work on (RL-based) cooperative emergent language (Lazaridou et al., 2017; Havrylov & Titov, 2017; Das et al., 2017; Evtimova et al., 2018; Kottur et al., 2017) in that we assume the teacher speaks a fixed, well-formed language, whereas in that work t”
TuringAdvice: A Generative and Dynamic Evaluation of Language Use (2021, background): “However, ensuring that machines communicate in standard English is di ﬃ cult, as there is usually a more e ﬃ cient machine-language coding scheme for the task (Kottur et al., 2017).”
Emergent Translation in Multi-Agent Communication (2017, background): “rge et al., 2016), adaptive communication bandwidth (Havrylov & Titov, 2017) and multi-turn communication with a variable-length conversation (Evtimova et al., 2017), and simple compositionality (Kottur et al., 2017; Mordatch & Abbeel, 2017). Meanwhile, Andreas et al. (2017) proposed a model”