What If Language Understanding Starts With a Dictionary Instead of a Model?
LLM → pattern matching → guessing → hallucination.
That’s the standard pipeline we accept today.
What if understanding language doesn’t start with prediction… but with structure?
🧠 The thing everyone skips
Every modern AI system eventually scales into:
- more parameters
- more data
- more compute
And yet the same problem keeps showing up:
Meaning is still unstable.
Not compute. Not storage.
Meaning.
📚 The overlooked system already exists
A dictionary already contains something interesting:
- A–Z defined
- 0–9 defined
- symbols defined
- every word has meaning
- every word connects to other words
It’s not random text.
It’s a structured semantic network.
Not perfect.
But structured.
✍️ The simple realization
No human knows every word in a dictionary.
But humans still learn language.
So the question becomes:
What if understanding is not stored… but constructed?
🧩 The direction I started exploring
Instead of building a model that predicts language, I started exploring something else:
A system that:
- starts from basic symbols (A–Z, 0–9, characters)
- builds into spelling → words → grammar → meaning
- uses a dictionary as the grounding layer
- connects meaning through structured relationships
- learns progressively like a curriculum
Not guessing.
Tracing meaning step by step.
🌊 The core idea (Kitana)
Kitana is not a traditional language model.
It is a cognitive system where:
- knowledge is structured (dictionary grounding layer)
- learning is progressive (like schooling)
- meaning is connected (graph / “tank” structure)
- understanding is dynamic, not stored facts
- reasoning comes from relationships, not prediction
⚠️ Still early
Right now it’s unstable.
Language is messy:
- slang
- ambiguity
- contradictions
- exceptions
And I’m still testing how far structure can go before it breaks.
But one pattern keeps repeating:
The system keeps returning to definitions instead of guesses.
🔥 Final thought
Maybe language understanding doesn’t start with intelligence.
Maybe it starts with:
structure strong enough to make intelligence emerge.
Top comments (6)
Hello senior developer @sylwia-lask 😊 do you see the chances of hallucination being reduced to the nearest minimum if an engine like Kitana is schooled into knowledge the way a human child is, rather than just guessing text?
If we communicate in plain, strictly structured English, can this structural approach push the error margin of communication correctness close to zero, or do you believe a baseline level of hallucination is unavoidable?
Thanks for the tag! 🙂 This is a really interesting concept.
I do think an approach like this could reduce the number of errors, especially by grounding meaning more explicitly. But nothing comes for free. A system built around definitions and structure would have its own challenges: ambiguity of language, incomplete knowledge, contradictory definitions, and of course... missing context. 🙂
Even humans, who learn progressively like children, make mistakes and sometimes arrive at wrong conclusions.
That's why my guess is that the future will be a hybrid of these approaches rather than one replacing the other entirely. Prediction-based models, structured knowledge, and verification mechanisms will probably complement each other instead of competing.
I agree with you. The future will likely be a hybrid rather than a competition 🙏
The circularity is what I'd want to see you tackle next. A dictionary defines "run" using words like "move" and "fast", which are themselves defined by other words, so tracing meaning eventually loops back on itself with no ground floor. Humans break out of that loop with sensory experience, a kid learns "hot" by touching something hot, not by reading the definition. Where does Kitana's grounding actually bottom out, or is the dictionary both the start and the floor? Also curious how it handles a word like "bank", where structure alone can't tell you river or money without the surrounding sentence.
On Grounding:
Meaning is not grounded in sensory experience like “touching hot.” It is grounded in structure. A word is understood when traversal through its definition graph closes without hitting a dead end. That closure is the signal of understanding. If traversal fails, the system does not infer or guess — it returns “unknown / not yet learned.”
On Context (“bank”):
Words are not resolved in isolation. Meaning is selected through context-driven traversal. Surrounding words activate specific regions of the definition graph and constrain the path that can be taken. For example, in the presence of “river” or “water,” the traversal naturally routes through the geographic sense of “bank,” not the financial one. Context does not redefine meaning — it selects the correct traversal path within the existing structure. 😊
Permit me to add a few assumptions behind Kitana's direction:
Incomplete knowledge is a curriculum problem, not necessarily an architectural failure.
Human mistakes don't automatically prove machine mistakes must persist at the same rate.
The model assumes valid, structured, reasonably clear input. Humans themselves struggle when communication is vague or poorly expressed.
If communication quality improves, understanding quality should improve too. We already see this principle in mathematics, programming languages, legal writing, and scientific notation.
Kitana's goal is not to know everything, but to learn, connect, and communicate from what it knows in a traceable way.
Kitana doesn't aim to be more knowledgeable than humans, only close enough to feel alike and reliable.
Error is not impossible even among humans. The objective is simply to reduce it as close to the practical minimum as structure, learning, and verification allow.
And perhaps the most amusing part of the experiment: if successful, Kitana could exist as a small offline system, measured in megabytes rather than billions of parameters. 😊