Is this a theory of English grammar?

Not in this entry. The source context concerns represented mental objects in AI knowledge representation.

Does storing a sentence make it true?

No. An agent can hold an incorrect expression, and a record can represent mention rather than belief.

Can the system reason about equivalent expressions?

Yes, if its formal rules support that reasoning. Equivalence is not automatic string identity.

AI Glossary / Graphs and symbolic AI / Advanced

Syntactic theory

Definition

In knowledge representation, a syntactic theory of mental objects represents the contents of beliefs or knowledge using expressions treated as objects, such as character strings. The system can reason about which expressions an agent holds without automatically identifying every expression that refers to the same thing. This entry uses that AI sense of “syntactic theory,” rather than the linguistic sense of a theory of sentence structure.

What it means

A fact about the world and a fact about someone's belief are different. Two names can refer to one person, while an observer recognizes only one of those names. Automatically substituting the names inside every belief report would incorrectly attribute knowledge to that observer.

The syntactic approach preserves descriptions as expressions. A quoted sentence can be the object of a relation such as Holds(agent, expression). This representation does not establish that human thought literally consists of stored text, or that storing an expression makes it true.

How it works

Define an object language for the expressions whose contents will be represented. Then provide a way to name or quote those expressions within a language that talks about them. For example, the expression Approved(Item7) can be represented by a string containing those symbols.

Specify relations that connect agents with represented expressions, and define which operations are permitted. A system might record that an agent explicitly received an expression, derived it using an allowed rule, or withdrew it after an update.

Keep expression identity separate from denotation. Two different strings may describe the same object, while the expressions remain different strings. If the model needs an agent to recognize an alias or infer an equivalent proposition, supply a justified rule or update. Unrestricted assumptions about truth, knowledge, quotation, and self-reference can create logical difficulties; a full formal theory needs carefully stated semantics and inference principles.

Three practical examples

1. Agent modeling: a simulation records which descriptions a participant has encountered. It avoids giving the participant an alias relationship known only to the simulator.

2. Explanation systems: a symbolic assistant distinguishes an explicit stored statement from another statement that could be inferred after applying a rule. This helps explain the provenance of an attributed belief.

3. Training scenarios: a tutor models a learner as having accepted one formula without assuming the learner immediately recognizes every logically equivalent formulation. The representation must state how learner knowledge is approximated.

Think of it this way

Think of two index cards with different names on them. A researcher knows both names identify the same author, but another reader has seen only one card. Keeping the cards distinct helps describe what that reader has actually encountered.

The analogy is limited: real beliefs are richer than cards, and a formal expression is not automatically a faithful model of a person's understanding. The technique provides representational control, not direct access to someone's mind.

Worked example

Imagine two names, Mira and Pilot7, denote the same person in the world model: Mira = Pilot7. Let the expression CanFly(Pilot7) describe that person's flying ability. A model of observer Lee contains Holds(Lee, “CanFly(Pilot7)”).

The two expression strings “CanFly(Pilot7)” and “CanFly(Mira)” are different sequences of symbols. Equality of the people named Mira and Pilot7 does not make these strings identical. Therefore the stored Holds fact alone does not supply Holds(Lee, “CanFly(Mira)”).

This does not deny the ordinary world-level inference: from CanFly(Pilot7) and Mira = Pilot7, a suitable first-order world theory can infer CanFly(Mira). The point is that an inference available to the outside modeler need not already be attributed to Lee.

If Lee later learns the identity, the agent model may apply an explicitly defined alias-substitution rule and add the second expression. Without that rule or update, it remains unrecorded. “Unrecorded” does not mean Lee believes the negation; the example uses a limited explicit-expression store, not a complete psychological theory or automatic closed-world belief model.

Illustrative case study

Fictional case: AliasLab builds a training simulation in which participants identify equipment under internal and customer-facing names. The simulator knows the mapping, but trainees discover it gradually.

The team initially merges every name in every participant record. This makes simulated trainees appear to know mappings they have not encountered. Engineers separate world identity facts from each participant's represented expressions and introduce explicit learning events. Reviewers inspect those events before attributing a new belief. The fictional case demonstrates the modeling distinction without claiming that string storage fully captures learning.

Common misconceptions and mistakes

Do not confuse a statement with the quoted representation of that statement. Approved(Item7) concerns an item; a string naming that expression is a different kind of object.

Different strings need not describe different realities. Conversely, equal denotations do not force identical represented beliefs.

A string store does not automatically provide a logic of belief. Rules for implication, consistency, updates, and attributed reasoning must be specified.

Avoid assuming that an agent knows every consequence of its information. Whether that closure is appropriate is a modeling choice, and unrestricted principles involving self-reference can be problematic.

Business use

The approach is relevant when systems must distinguish the organization's complete information from the information attributed to a particular participant or agent.

Document what a stored expression means: explicitly received, currently believed, derivable, or merely mentioned. These are different statuses. Preserve provenance and update rules so that an audit can explain why the system attributes a statement to an agent. Treat the model as a designed approximation and evaluate it against the intended task, rather than interpreting its records as a definitive account of human cognition.

Check your understanding

Why does Mira = Pilot7 not by itself give Lee the second Holds fact?

Answer: The equality concerns the named person. The Holds relation in this example records distinct expressions, and no rule has yet attributed the alias-based inference to Lee.

Related terms

Knowledge representation
Reification (computer science)
Modal logic
Logical omniscience
propositional attitude
quotation
reification
referential opacity
denotation

Learn next

Study referential opacity next, then compare an explicit-expression model with a possible-worlds model of knowledge. For a small alias example, list world facts, represented agent expressions, and permitted inference rules separately. This makes the source of each attributed conclusion visible.

Inference rules

Sources

Primary reference: Russell and Norvig, AIMA corrected bibliographical notes, https://aima.cs.berkeley.edu/errata/aima-341.pdf. The authors' description of syntactic mental-object representation and its logical cautions was read directly. The Lee/Mira example is an original restricted model; the title's linguistic ambiguity is resolved explicitly above.

Previous: Syntactic sugar Next: Syntax