Minimal symbol grounding through language statistics: An information theoretic approach quantifying uncertainty

One of the fundamental issues in the cognitive sciences concerns the question of how language attains meaning. The current study investigated how the extent to which language statistics and symbol grounding reduce uncertainty about word meaning can be quantified. Using Shannon's information theory, conditional entropy of the valence of a word was quantified without grounding (baseline), with minimal symbol grounding and with full grounding. By using a mini lexicon as a principled illustration, later extended to a full lexicon, the entropy of phonological cues, frequency cues, and word embeddings were compared to a symbol-grounding case when predicting word valence. Phonology and frequency provided small but measurable reductions in uncertainty, which word embeddings propagated through a lexical network, particularly once only a few seed words were grounded. The results show that minimal grounding, combined with language-statistical structure, can substantially reduce uncertainty about word valence. The evidence of minimal symbol grounding sheds light on the symbol-grounding problem and encourages new and quantifiable perspectives on past and present theories on language and cognition.