A Primer on Computational Semantics for Artificial Intelligence Systems

This paper examines the fundamental nature of linguistic meaning (semantics) and explores how it is approached across different scientific and philosophical disciplines. The author contrasts how humans acquire language through physical, social, and emotional experiences with how modern transformer-based language models (LLMs) learn through the statistical analysis of massive text corpora. The central argument is that while LLMs are highly effective at symbolic communication, they lack the grounding in the physical world that gives language its true meaning for humans.

The paper outlines three primary frameworks for computational semantics:

• Formal Semantics: Uses logic and set theory to represent meaning, treating language like a computer program. While elegant, it struggles to map abstract logical forms to the messy, physical reality of the world.
• Grounded Semantics: Argues that meaning is derived from sensory experience and interaction with the environment. It addresses the Symbol Grounding Problem, which posits that symbols (words) cannot carry meaning on their own without being linked to physical referents.
• Distributional Semantics: The foundation of modern LLMs, which assumes that a word's meaning can be inferred from the company it keeps (its context in text). The author argues that while this approach is powerful, it remains ungrounded because text is merely a symbolic representation of language, not language itself.

The author emphasizes that human language acquisition is a deeply social, embodied, and incremental process. Children learn language through joint attention, physical exploration, and emotional engagement with caregivers. Because LLMs are trained on static text, they miss out on the multimodal experiences—such as the physical sensation of an object or the emotional valence of a word—that are essential for human-like comprehension. The paper suggests that the path forward for AI involves a hybrid, neuro-symbolic approach that integrates formal logic with grounded, multimodal learning.