Computational Semantics for AI: How Meaning Actually Gets Built

Computational Semantics for AI: How Meaning Actually Gets Built

Introduction: “What does this word mean?”—and why AI needs the answer

If you’ve ever tried explaining a word like “red,” you’ve probably felt the same frustration as Kennington’s students: you can point at examples, but you can’t easily explain the meaning in a way that feels complete. People can use language effortlessly, yet it’s surprisingly hard to say what meaning “is” in a principled way.

This new primer on computational semantics—based on Casey Kennington’s paper available at arXiv:2608.25022—dives into how different fields try to answer that question. The paper is motivated by a very practical problem: transformer-based language models (think ChatGPT and Gemini) are increasingly used as if they “understand” language. To use them responsibly, we need to be more precise about what meaning is, how it gets represented, and where these systems likely break.

Kennington’s big goal is to map meaning across multiple perspectives—linguistics, philosophy, psychology, neuroscience, and social science—and then connect those ideas to three semantic theories that are often treated as computational options: formal semantics, grounded semantics, and distributional semantics. Along the way, it compares how humans learn language (including the roles of attention, joint attention, emotion, embodiment, and intention) with how today’s AI learns from text.

The punchline is not “AI can’t understand anything.” It’s more interesting than that: AI can manipulate language well, but the nature of the “meaning” it learns may be missing key ingredients that humans use—especially grounding in the world, emotional connotation, and embodied experience.

Why This Matters: Meaning isn’t just text—it’s a learning curriculum

The timing on this research is excellent because we’re at a weird inflection point: society is deploying language models in high-stakes settings (customer support, tutoring, coding assistance, therapy-like conversations, policy drafting), yet our evaluation habits still mostly test fluency and task success, not whether the system has built the right kind of meaning.

Kennington’s framing matters right now because it highlights a mismatch that shows up in real deployments: language models may produce responses that are “reasonable-sounding,” while lacking the grounding and interactive common ground that humans rely on. That doesn’t mean every failure will be obvious. It means the failures can be subtle—like over-trusting an explanation that is semantically shallow, or treating a model’s “I know what apples are” as evidence of real understanding (more on that later).

A scenario you could apply today

Imagine you’re building an AI assistant for a hospital helpdesk. Patients ask, “What does that diagnosis mean?” and the assistant responds with careful definitions. If the assistant is trained only on text, it can still produce definitions and medical-sounding explanations. But patient understanding depends on more than definitions: it depends on shared context, emotional valence (fear, uncertainty), and the patient’s lived experience (what they’ve actually seen, felt, or been told).

Kennington’s work suggests a design principle: don’t just ask whether the model can produce semantic content—ask whether it has access to the kinds of signals that ground meaning in humans. In practice, that could mean:
- retrieving relevant patient-specific context (conversational grounding),
- using multimodal or interactive components when appropriate (grounding),
- and calibrating the assistant’s confidence because meaning-depth can’t be inferred from linguistic fluency alone.

How this builds on prior AI research (instead of replacing it)

This primer builds toward a familiar direction in AI: neuro-symbolic AI and hybrid systems—models that combine learned representations with more structured reasoning. Kennington explicitly ends with the “two semantic problems” framing:
1. the symbol grounding problem (how meaning becomes connected to the world), and
2. the neuro-symbolic problem (how to combine computation on representations with meaningful reasoning).

The paper’s distinctive contribution is the emphasis that “meaning” isn’t one thing. It’s a composite of multiple mechanisms—formal logic, world grounding, distributional learning, plus human-like learning features like joint attention and intention. That suggests the safest path isn’t to pick one semantic theory and declare victory; it’s to combine them in a way that respects how humans learn.

Formal semantics: when meaning looks like logic (and what that misses)

Formal semantics tries to treat meaning as something computers can manipulate with logic. The intuition is: computers already run on logic-like structures—so why not map language onto formal representations?

From bits to meaning: logic as the “native language” of computation

Kennington walks through the bridge from logic gates to computational logic, then into First Order Logic (FOL). In this view, words and phrases can be translated into logical statements using operators like AND, OR, quantifiers, and negation.

For example, “there exists a gray thing” becomes something like:
- ∃x.gray(x)

And combining requirements can look like:
- ∃x.big(x) ∧ gray(x) ∧ elephant(x)

This is attractive because it gives you compositional meaning: you can build complex meaning from smaller parts, at least in principle.

The real issue: logic still doesn’t tell you what “gray” means

Here’s the big shortcoming Kennington emphasizes: FOL gives you structure, not grounding. In the formula above, the machine still needs a way to know when gray(x) should be true. That’s not solved by logic itself; it needs an interpretation layer (sets, models, perhaps modal logic, etc.). In other words, formal semantics provides the skeleton of meaning, but not necessarily the “muscle” connecting it to reality.

This is why formal methods often work best when you already have structured knowledge: a labeled ontology, a controlled domain, or explicit truth conditions. Without that, the model can manipulate representations perfectly while still lacking genuine semantic awareness.

Grounded semantics: “meaning as experience,” not just symbols

Kennington next tackles grounded semantics, which is basically the claim that symbols don’t have meaning by themselves. They have meaning when they’re connected—grounded—into experience.

The symbol grounding problem, in plain language

The paper points to Stevan Harnad’s classic formulation of the symbol grounding problem: symbols can be manipulated correctly, yet still fail to “mean” anything in the way humans mean it.

Kennington connects this to Searle’s Chinese Room style argument: imagine you can map input symbols to output symbols using a dictionary, but you never actually understand the content. From the outside, behavior looks correct; from the inside, there’s no grounded understanding.

The key phrase Kennington uses is: text is an ungrounded placeholder. You can use the word “chair,” but the written word doesn’t give you what a chair is like to sit on.

Concrete vs abstract concepts: a continuum of grounding need

Grounded semantics gets practical when you ask: which words need grounding most?

Kennington divides meanings along a concreteness–abstractness continuum:
- Concrete words (like red, ball, chair) refer to physical things and typically require grounding through vision, touch, and action.
- Abstract words (like democracy, utopia) don’t map neatly to direct sensory input; they’re learned via other concepts, social practices, and linguistic definitions.

Importantly, Kennington notes this isn’t a binary. It’s a spectrum. Even abstract words often have “partial” sensory structure in their meaning (metaphors, emotional connotation, partial feature overlap).

Why this matters for AI today

This framing has immediate implications for evaluation. If your AI model has “meaning” learned only from text, then:
- it may excel at definitional semantics (formal-ish, dictionary-like understanding),
- but it can be weaker on experience-dependent meaning (what “relief” feels like after drinking water, what “sitting” affords, how emotional connotation changes interpretation).

Kennington doesn’t claim that grounded semantics is easy to implement—he explicitly calls out that grounding a model’s language into the physical world is not automatic, and current approaches can be arbitrary (for example, using vision features because they’re available, not because they capture the semantic role the grounding should play).

Distributional semantics: meaning as “the company words keep”

Distributional semantics flips the problem. Instead of grounding words in the world, you ground them in language usage.

“You know a word by the company it keeps”

Kennington relies on a familiar intuition from Firth: words derive much of their meaning from co-occurrence patterns. If a word appears in similar contexts to another word, it likely has related meaning.

Computationally, that becomes embedding spaces:
- words become vectors (points in an n-dimensional space),
- similar meanings become nearby points,
- distances correlate with semantic similarity.

From one-hot vectors to word2vec and beyond

The paper outlines the evolution:
- one-hot vectors represent words as distinct units (everything is far from everything else—too crude),
- early distributional embeddings compress co-occurrence information into sparse vectors (computationally heavy),
- word2vec (Mikolov et al.) learned dense embeddings using a prediction objective, shaping vectors into a useful geometry.

Kennington then notes a limitation: word-level embeddings struggle with sentence meaning. His phrasing captures the core issue: you can’t cram the meaning of an entire sentence into a single word vector and expect composition to work automatically.

Transformers as the “composition upgrade”

This is where the primer lands its modern relevance. Transformers (Vaswani et al., 2017) are trained to predict words from context, and the representations they learn capture not just word-level associations but phrase/sentence/document-level patterns through attention-based computation.

Kennington’s comparison of historical NLP shifts is clear:
- earlier language models were mostly about sequence statistics,
- transformer models became capable of broader generalization tasks after pre-training and fine-tuning.

In practice, this is why “one model does many tasks” became the new paradigm.

The key limitation for today’s LLMs: text isn’t meaning

Kennington makes the point bluntly: text is not language; text is not meaning. It’s a representation of language used by humans, but not the experience that gives words their full semantic weight.

He gives a memorable example: ask a text-trained system something like “Have you ever seen an apple?” It will correctly report it’s never seen anything. Meanwhile, “apples mean something” because humans have tasted, seen, grown, and cared about apples—meaning that lives in embodied experience.

Humans learn meaning through interaction: attention, joint attention, emotion, intention

So what’s missing? Kennington’s answer is: the semantic learning environment matters. Humans acquire language through situated, interactive dialogue long before literacy.

The “setting” of language acquisition

Kennington draws on Clark’s account of early language learning, emphasizing the child’s environment:
- situated: people share space and time,
- shared attention: gestures and gaze let learners coordinate,
- speech: spoken dialogue is primary (reading comes later),
- joint activities: language is embedded in lived interaction.

This connects directly to meaning: words get associated with what the learner and caregiver are attending to together.

Joint attention: why it’s not optional for first words

Attention isn’t just “nice to have”—Kennington argues it’s foundational. Children learn early words when caregivers direct attention to an object, and crucially when caregivers and children have joint attention: both know the other is attending to the same target.

That gives you a mechanism for referential meaning that pure text training lacks.

Conversational grounding: meaning updates mid-conversation

Kennington also emphasizes that meaning evolves during interaction. Two people can start with different assumptions about a word and then repair them through clarification requests, rephrasing, and shared understanding. He quantifies a bit here: in a corpus analysis, around 3.5% of dialogue turns had clarification requests; other studies report higher rates in spontaneous dialogue (around 3–6%).

For AI builders, the takeaway is practical: ambiguity and meaning repair are normal, not edge cases. A system that ignores conversational grounding is missing a core part of how humans stabilize meaning.

Emotion and embodiment: the parts of meaning that don’t show up in definitions

Kennington argues that separating cognition from emotion makes language modeling easier—but it also makes meaning shallower. Many words carry emotional valence (positive/negative tendencies), and emotional connotations are especially important for abstract terms.

He frames embodiment as equally important: if meaning depends on perception, action, and feeling, then a system without a body will learn a “type” of meaning that may not transfer cleanly to human-like understanding.

This is one reason he’s skeptical of purely text-trained semantic competence: it can approximate linguistic meaning, but not the full meaningfulness humans build through embodied life.

Key Takeaways

• Meaning isn’t a single method. Kennington argues for a composite view: formal structure, grounding in experience, and distributional learning each capture part of the story.
• Formal semantics gives logic, not understanding by itself. FOL can represent structure and composition, but predicates still need interpretation/grounding to reflect real-world meaning.
• Grounded semantics ties words to lived experience. Symbols become meaningful when connected to sensory-motor and emotional experience—especially for concrete concepts, though everything sits on a continuum.
• Distributional semantics scales and works—yet “text” isn’t “meaning.” Transformers learn powerful context-based representations, but without world experience, they can miss grounding-heavy semantic components.
• Human language learning depends on interaction mechanisms like situated dialogue, attention and joint attention, conversational grounding, intention, and emotional/embodied development.
• For AI today: if you want more than fluent responses, design systems that incorporate (a) interactive context, (b) grounding signals (multimodal or situated when possible), and (c) calibration for semantic depth—don’t assume linguistic competence implies grounded understanding.
• Future direction: the paper points toward hybrid approaches (neuro-symbolic plus grounding) to address both the symbol grounding and neuro-symbolic challenges simultaneously.

If you want, tell me what kind of AI system you’re thinking about (chatbot, tutoring tool, robotics, summarization, agent planning), and I can suggest which semantic ingredients from the paper matter most—and what failure modes to test first.