Symbol grounding problem

Suppose you had to learn Chinese from a Chinese–Chinese dictionary alone. Every word is defined by other words, which are defined by others. You could follow the definitions round for ever and never arrive at what any of them means. Stevan Harnad used this picture in 1990 to state the symbol grounding problem: a system that only relates symbols to more symbols has, so far, no meanings of its own. The meanings are in the heads of the people reading its output.

Where it comes from

The problem generalises Searle’s Chinese room, which argued that running a program cannot be enough for understanding, because a program is defined over the shapes of symbols and not over what they stand for. Harnad took the point as a design problem and asked what would have to be added. His answer was that some symbols must be connected, through the system’s own sensory and motor capacities, to the things they are about: the ability to pick out horses is what grounds “horse”, and other terms can then be built from grounded ones.

Positions

One response accepts the problem and holds that only an embodied system, acting in a world, could have grounded symbols. Another denies that grounding needs a body: meaning might be fixed by a term’s role in a vast web of inference, or by causal and historical links to a community of speakers who are themselves in touch with the world. A third doubts that there is a single thing called meaning for which one account is needed.

Language models

The problem has returned with large language models, which are trained almost entirely on text. On Harnad’s picture they are the dictionary and nothing more. Their defenders reply that text written by grounded speakers carries a great deal of structure about the world, and that models trained with images or connected to tools are no longer confined to text.

My own work does not depend on settling this. Whether or not a model’s words are grounded, a further question remains about each thing it asserts: whether its being true played a part in its being said (the Tracking Minimum).