Symbol Grounding Problem

How the symbols inside a machine come to refer to real things in the world, and so mean something.

This is a problem cognitive scientist Stevan Harnad posed in Physica D in 1990: for a system that only manipulates symbols according to rules, how do those symbols get their own meaning, rather than depending entirely on a person to interpret them? His example is trying to learn Chinese from a Chinese-Chinese dictionary alone: every word is defined using other words, and no amount of looking things up ever touches an actual object. Harnad's proposed solution is to ground the most basic symbols in perception: build sensory representations of the outside world first, then learn features that distinguish categories; symbols are the names of these categories, and complex concepts are built by combining them. Embodied AI often treats this as part of the theoretical case for “why a body is needed,” and it is frequently cited in discussions of language grounding, meaning connecting words to vision and action.

ExampleGiven the instruction “bring me the red cup,” a large language model can parse the sentence, but it only counts as grounded for the robot once “the red cup” is tied to a specific object in the camera feed and “bring” is tied to a sequence of grasping actions.

Also called

Symbol Grounding

Related

Language Grounding · Embodied Cognition · Disembodied AI · Three Schools of AI · SayCan

Sources

The Symbol Grounding Problem (Harnad, Physica D 1990)
Symbol grounding problem - Wikipedia