In classical logic, a proposition like "The sky is blue" possesses a stable, objective truth value determined by comparing the statement to the real world. However, natural language rarely operates in a vacuum of absolute, static truths. Consider the statement:
"I am standing here today."
If Alice says this in Paris on Tuesday, it means something entirely different than if Bob says it in Tokyo on Friday. Words like I, here, you, and today are indexicals—linguistic placeholders whose meaning and truth values shift dynamically based on the context of the speech act.
For decades, researchers have recognized that traditional logical frameworks break down when encountering indexicals because they treat sentences as abstract strings with fixed meanings. To bridge the gap between formal logic and natural language, a specialized research direction emerged: utterance-dependent logic. This paradigm shifts the evaluation of truth from the abstract statement itself to the specific instance of its utterance, formalizing how context binds variables to reality.
The Core Problem: The Failure of Static Logic
The fundamental challenge of knowledge representation in multi-agent environments is the tendency to conflate an abstract sentence with a specific instance of its utterance. As Robert C. Moore argued in Problems in Logical Form, a logical representation cannot be effectively utilized for automated reasoning until its definite references are resolved against the specific environment in which it was spoken.
Without this context-dependent layer, standard logical systems fall into inconsistencies. For example, in the classic "Knights and Knaves" logic puzzles—where knights always tell the truth and knaves always lie—an inhabitant saying "I am a knave" creates an immediate paradox if "I" is treated as a fixed constant. The system must be able to link the pronoun to the speaker's identity and character traits before evaluating the claim's validity.
Furthermore, natural language interpretation does not require an objective, real-world ontological truth to function. In On the Use of Linguistic Quantifying Operators in the Logico-Semantic Structure Representation of Utterances, Irena Bellert pointed out that semantic rules must operate identically whether a speaker is telling the truth, lying, or imagining. Meaning is governed by the speaker's purported beliefs and modal frames, meaning logic must adapt its quantifiers to accommodate subjective perspectives rather than absolute reality.
The Mechanism: Truth Under Utterance
To resolve these paradoxes without relying on ad hoc programming tricks, researchers introduced specialized predicates into first-order logic. In Proving Facts about "I" and Proving Facts about "|", Michael J. Miller et al. proposed a formal mechanism known as the Truth under Utterance predicate, formalized as $TU(p, t)$.
Instead of evaluating a statement $t$ in isolation, the $TU(p, t)$ predicate asserts that the statement $t$ is true if and only if the indexical placeholder "I" within $t$ is dynamically replaced by the identity of the agent $p$ who uttered it.
Standard Logic:
Statement: "I am a knight" --> Evaluated as a static string (Fails or causes paradox)
Utterance-Dependent Logic:
Utterance Instance --> TU(Agent_A, "I am a knight")
Dynamic Binding --> Replaces "I" with "Agent_A"
Situated Evaluation --> Evaluates: Knight(Agent_A) based on Agent_A's traits
By treating indexicals as placeholders within a logical schema rather than fixed constants, automated theorem provers can dynamically bind identities during the reasoning process. This allows the system to establish axioms governing how an agent's inherent character traits (such as reliability or truthfulness) dictate the objective truth of their declarations.
Nested Perspectives and Quasi-Indexicals
The problem deepens when agents reason about other agents' self-references. If John believes that "he himself is in danger," the phrase "he himself" functions as a quasi-indexical. It represents how John would refer to himself ("I am in danger") if we were to step inside his internal perspective.
To handle this, William J. Rapaport et al. developed a semantic network approach in Quasi-Indexical Reference in Propositional Semantic Networks. Their framework explicitly separates an agent's internal self-beliefs from an external observer's representation of that agent. By capturing this indirectness, the system prevents catastrophic inference errors—such as mistakenly attributing a belief about John to someone else who happens to use the pronoun "he."
Modern Challenges: Indexicals in Language Models
While the foundational rules of indexical binding were mapped out using formal logic, evaluating these capabilities in modern AI systems remains a critical frontier. Large Language Models (LLMs) excel at standard third-person coreference resolution (e.g., matching "he" to "John"), but they struggle significantly with shifting indexical contexts.
In Un-considering Contextual Information: Assessing LLMs' Understanding of Indexical Elements, Metehan Oguz et al. demonstrated that while frontier models successfully resolve the word "I," they frequently fail on words like "you" and "here." Instead of applying strict grammatical rules to track who is speaking to whom, LLMs often ignore the structural mechanics of the speech act, defaulting instead to superficial, misleading contextual "vibes."
This challenge is compounded in multilingual settings. As highlighted in the Proceedings of the First Workshop on Natural Language Processing for Turkic Languages (SIGTURK 2024), many languages feature complex grammatical rules governing indexical shift, where the perspective of a pronoun changes entirely depending on whether it is embedded in direct or indirect speech. Evaluating a model's capacity to navigate these shifts serves as a crucial benchmark for true contextual understanding.
Ultimately, whether through first-order logic predicates or modern evaluation benchmarks, the core insight of this research direction remains the same: language cannot be decoupled from the act of speaking. To truly understand a claim, an intelligent system must not only parse what was said, but who said it, where, and when.
Key Foundational Work
Problems in Logical Form (1981) — contributed 1 concept elements to this direction.
On the Use of Linguistic Quantifying Operators in the Logico-Semantic Structure Representation of Utterances (1969) — contributed 1 concept elements to this direction.