Abstract
Recognizing speaker intent—identifying the "speech act"—is a fundamental challenge in natural language processing. Traditional approaches typically oscillate between rigid idiomatic templates and computationally expensive plan-based reasoning. This paper proposes a hybrid model that integrates linguistic conventions with plan-based conversational implicatures. By using grammatical features to constrain the set of possible interpretations and logic to filter them, the system achieves a balance between linguistic accuracy and computational efficiency.
Introduction
In human conversation, the literal meaning of a sentence often diverges from its intended function. A question like "Can you pass the salt?" is rarely an inquiry into the hearer's physical capabilities; it is almost always a request. Conversely, "Can you speak Spanish?" is usually a literal question. For artificial intelligence, distinguishing between these interpretations requires navigating the tension between what is said and what is meant.
Historically, AI research has split into two camps. The Idiom Approach treats indirect speech acts as fixed templates, which fails to account for literal interpretations. The Plan-Based Approach relies on pure logic to infer goals, which often ignores the specific linguistic conventions that govern how different languages express requests or assertions. This paper argues that neither approach is sufficient in isolation and proposes a two-stage model that utilizes both linguistic constraints and situational logic.
The Hybrid Architecture
The authors propose a system where linguistic form and plan reasoning act as dual constraints on ambiguity. The process is divided into two primary stages:
1. Linguistic Constraints
The first stage uses compositional rules to map linguistic features—sentence mood, verb form, and lexical markers—to a set of candidate speech acts. This is implemented through unification-based graph matching. For instance, the presence of the word "please" acts as a powerful structural constraint, immediately triggering a "directive-act" interpretation and eliminating purely interrogative possibilities. This stage defines the "input range" of what a sentence could mean based on how it is phrased.
2. Plan Reasoning as a Filter
Once a set of candidates is generated, a plan reasoner serves as a plausibility filter. It employs "conversational implicatures" to evaluate whether a specific speech act makes sense given the agent's knowledge and the current context. If the system knows the hearer is capable of an action, it may rule out a literal inquiry into ability in favor of a request. If no conventional interpretation fits the context, the system can fall back to "extended plan reasoning" to decode non-conventional hints (e.g., interpreting "It’s cold in here" as a request to close a window).
Empirical Basis and Implementation
To verify the linguistic rules, the authors examined the AP News Corpus, a dataset of approximately 43 million words. This corpus was specifically used to analyze the function of markers like "please" in directive utterances. The system was further evaluated using classic ambiguous cases, such as "Do you have a watch on?" and "Can you speak Spanish?"
Technically, the module was implemented using Common Lisp and the Rhetorical knowledge representation system. The implementation demonstrated that the system could successfully distinguish between literal and indirect meanings based on the interplay of grammatical cues and contextual beliefs.
Key Insights
- Convention vs. Logic: The paper highlights that linguistic forms are not merely hints for logic; they are formal constraints. Phrases like "Are you able to..." and "Can you..." share literal semantics but carry different "force" due to language-specific conventions.
- Efficiency through Constraint: By using linguistic patterns to narrow the candidate pool early, the system avoids the "state-space explosion" common in pure logic-based systems that attempt to derive every intention from first principles.
- Psycholinguistic Alignment: The model mirrors human processing patterns, where idiomatic or conventional meanings are often accessed as quickly as, or faster than, literal meanings.
- Embracing Ambiguity: The authors argue that a system does not always need to resolve a sentence to a single "correct" act. In many interactions, intentions are naturally ambiguous, and a robust system should be capable of acknowledging multiple valid interpretations.
Limitations and Future Work
While the hybrid model provides a robust framework, several challenges remain. The "projection problem"—how speech act interpretations function within complex, nested sentences—remains largely unsolved. Additionally, while the "please" rule was verified against a large corpus, the broader library of interpretation rules requires further scaling. Finally, the current text-based model does not yet account for spoken intonation, which serves as a critical prosodic clue for disambiguating speech acts in natural conversation.
Summary updated