Inferring Speakers’ Communicative Intents through Iterative Reasoning

Understanding Others’ Intentions: A Pragmatic Approach to Language
Humans have an incredible ability to understand others’ intentions, even when they don’t explicitly state their goals. This pragmatic understanding is crucial for effective communication and helps us navigate complex social situations. Researchers have developed models to capture this ability, using concepts like "iterative inference" to describe how we reason about other agents’ beliefs.
The Rational Speech Acts (RSA) model, built upon by the authors, frames communication as a Bayesian reasoning process between listeners and speakers. The listener, represented by π0 L, maps utterances to actions based on their literal meanings. A pragmatic speaker, πS, chooses an utterance that is likely to lead a literal listener to the intended action. This selection process is guided by the probability of the literal listener taking the right action upon hearing the chosen utterance.
In everyday life, we encounter situations where language deviates from its literal meaning. For instance, when describing colors, people might use different terms depending on the context. In this context, a speaker might choose to describe purple as blue by reasoning that a naive listener might resolve purple to the second color in the row. This recursive reasoning allows us to understand each other’s intentions and communicate more effectively.
The authors extend the RSA model by incorporating an approximate inference procedure π(a | s; R, β) that takes into account a reward function and computational budget β. This allows agents to make decisions based on suboptimal or even unsuccessful plans, which is a key human skill present in children as young as 18 months (Meltzoff, 1995).
In summary, this article focuses on pragmatic language understanding and the ability to infer speakers’ communicative intents from their utterances. The authors propose a model based on iterative inference, which enables agents to reason about each other’s beliefs and select appropriate actions. This approach can help us better comprehend how humans communicate and navigate complex social situations.