Abstract
This paper introduces a formal framework for AI systems to distinguish between "means clauses" (describing how an action is performed) and "rationale clauses" (describing why an action is performed). By mapping these linguistic structures to a logic of beliefs and intentions (BDI), the model enables machines to infer an agent's mental state, including intentionality and the perceived success of actions.
Introduction
Natural language often combines two actions into a single utterance, yet the choice of connective—such as "by" versus "to"—drastically alters the implied mental state of the agent. Consider the sentences:
- "John dirtied the carpet by walking across the room." (Means Clause)
- "John walked across the room to dirty the carpet." (Rationale Clause)
To a human reader, the first suggests the outcome (dirty carpet) definitely occurred, though perhaps unintentionally. The second suggests a clear intention, though the outcome remains uncertain. For computational systems to achieve true natural language understanding, they must move beyond identifying actions to inferring these underlying mental states. This paper addresses the challenge of formalizing how prepositions like "by" and "to" modulate beliefs about intentions and causal success.
Methodology
The author proposes an interpretation model grounded in formal logic, utilizing several key components:
- Action Reification: Actions are treated as first-order individuals, following Davidsonian semantics, allowing them to be quantified and described by predicates.
- Logical Forms (LF): Sentences are translated into logical expressions where "by" and "to" act as two-place predicates linking action variables.
- BDI Modeling: The framework employs formal predicates such as
BEL(agent, time, proposition)andINT(agent, time, act-type, time)to represent the agent's beliefs and intentions. - Axiomatization: The model adopts standard logical axioms and introduces a specific "Intention Axiom" (I1). This axiom posits that if an agent intends to perform action $A$ to generate $B$, they must intend both $A$ and $B$ and believe that $A$ will cause $B$.
- Interpretation Rules: A set of rules ($LF1, LF2, LF3$) bridges linguistic structures and mental states, deriving specific BDI profiles from "by" or "to" predicates.
The model was validated using task-oriented dialogues from collaborative human tasks and descriptive text from Associated Press news stories.
Insights into Linguistic Structures
A key insight of the model is the "potentiality" inherent in rationale clauses. In the sentence "John walked across the room to dirty the carpet," the speaker does not assert the carpet is dirty, but rather asserts their belief about John's belief. This nested belief structure allows the model to represent potential actions that may not materialize in reality.
The paper also distinguishes between information derived from verb tense and that derived from sentence structure. While both means and rationale clauses describe a "generation" relation where one action results in another, they prioritize different aspects of the agent's state. Means clauses emphasize the result and the causal link, whereas rationale clauses emphasize the agent's intent.
Results and Formalization
The differences between these constructions are formalized into a truth table that maps utterances to inferred beliefs.
As shown in Table 1, the model yields distinct profiles:
- Means Clauses (MC): The system derives that the speaker believes both actions occurred and a causal link exists, but remains neutral ("?") regarding the agent's original intention.
- Rationale Clauses (RC): The system derives that the agent intended both the outcome and the method, while the speaker remains neutral ("?") regarding whether the outcome actually occurred or if the causal link was valid.
By applying these rules, AI systems can automatically populate knowledge bases with accurate BDI profiles for agents described in text.
Conclusion
This work provides a robust formalization for simultaneous actions in rationale and means clauses. Future research will focus on extending the model to sequential actions (e.g., "going to the store to buy milk"), handling complex negation and modals, and integrating these utterance-level rules into broader discourse models to track evolving beliefs in conversation.
Summary updated