Linguistics Seminar Talk - Mutsumi Imai

Note that this is an online talk and will be held on Zoom. Please use this link to join:

https://ucl.zoom.us/j/93367144409

Meeting ID: 933 6714 4409

Title: Symbol Grounding in humans: The roles of iconicity and abductive inference in the acquisition and evolution of language

Abstract:

In 1990, Steven Harnad first proposed the “symbol grounding problem” in the context to criticize Artificial Intelligence using a symbolic approach at that time. He pointed out that symbols cannot acquire meanings through transformations of other symbols, arguing that symbols must be connected to the world, especially to the body (Harnad, 1990). Interestingly, after 30 years, people may think that the Symbol Grounding problem has been overcome by recent Generative AIs, but no matter how fluent and natural the generated language is, words are still not anchored to the world or the body.

In this talk, I address the process of symbol grounding in humans considering both language acquisition and language evolution. Language may be understood as a socially shared symbolic system (Nelson & Shaw, 2002). In my view, to solve the “symbol grounding problem” (Harnad, 2003), we need to understand not only how children map their physical and sensory experiences to the first set of symbols but also how they go from the “here and now” and climb the ladder into an inter-connected system of socially shared, more or less abstract concepts, which are largely language-specific.

I consider two abilities of human children that play key roles in language learning. One is the ability to detect cross-modal iconicity between sound and other sensory modalities. However, as I show, this ability is not sufficient to initiate language learning. Language learning, particularly learning of word meanings, requires chains of abductive inference, in which different sources of knowledge and perceptual and contextual cues must be properly weighed and combined (Imai, 2017).

I present the results of two research programs from my laboratory. I will first show that preverbal infants can detect sound-meaning correspondence of novel words and discuss the implications of these findings (Asano et al., 2015; Yang et al., 2019). I then show that human infants have the ability (or disposition) to make abductive inferences at the age of 8 months, and this divides humans and non-human animal species in their ability to learn language (Imai et al., 2021). In the end, I will explore how these two elements—the ability to detect iconicity across different sensory modalities and the ability to draw abductive inference—might have kick-started the use of language in our ancestors.