Abstract
Conversational implicatures are usually described as being licensed by the disobeying or flouting of a Principle of Cooperation. However, the specification of this principle has proved computationally elusive. In this paper we suggest that a more useful concept is rationality. Such a concept can be specified explicitly in planning terms and we argue that speakers perform utterances as part of the optimal plan for their particular communicative goals. Such an assumption can be used by the hearer to infer conversational implicatures implicit in the speaker's utterance.
Introduction
In the study of linguistics and artificial intelligence, "conversational implicature" refers to the meaning conveyed by a speaker beyond their literal words. Since H.P. Grice's seminal work in 1975, the "Principle of Cooperation" and its associated maxims (Quality, Quantity, Relation, and Manner) have served as the standard framework for understanding these nuances. For instance, if a user asks, "Can I switch off the computer?" and an expert responds, "You will damage the hard drive," the listener intuitively understands the answer as "No."
However, Grice’s maxims are notoriously "computationally elusive." They lack the formal specification required for a dialogue system to determine what constitutes being "relevant" or "brief" in a given context. This paper, authored by Mark Lee, proposes replacing these fuzzy maxims with a concrete framework based on rationality and optimal planning.
A Planning-Based Approach to Implicature
The core thesis is that conversational implicatures can be derived by treating speakers as rational agents. Instead of relying on social maxims, the system evaluates utterances based on three rationality heuristics: Correctness (grounded in true beliefs), Relevance (achieving the maximum set of goals), and Efficiency (achieving goals at minimum cost).
When a speaker’s utterance appears "inefficient" for a single, obvious goal, a rational listener assumes the speaker is pursuing additional goals. This process, termed "Goal Ascription," allows the system to infer hidden intentions that make the utterance part of an optimal plan.
As shown in Figure 1, the system identifies two primary types of additional goals:
- Conjunctive Goals: The speaker aims to achieve multiple objectives simultaneously (e.g., answering a question while educating the listener).
- Avoidance Goals: The speaker intends to avoid a specific state or topic.
Implementation in ViewGen
The author implements this theory within ViewGen, a system designed to model the beliefs and intentions of multiple agents. The system uses a partial order clausal link planner to generate and compare potential plans. The process begins with Plan Recognition, where the system identifies the literal speech act. If the speaker's actual utterance is longer or more complex than the theoretical "optimal" plan for the primary goal, the system performs a "Plan Repair." It ascribes new goals to the speaker until their behavior appears rational and optimal again.
Figure 2 illustrates this in the context of the computer expert example. If the expert's only goal was to say "No," a simple "No" would be the most efficient plan. By providing information about hard drive damage, the expert is seen as pursuing a conjunctive goal: informing the user of a consequence to prevent future damage (an "educate_goal"). The system then updates its internal belief model to reflect that the speaker intended to communicate a "No" through the side effects of their statement.
Insights and Future Directions
This framework offers several key insights. First, it treats implicature as a form of "Plan Repair"—we don't assume a speaker is irrational; rather, we assume our model of their goals is incomplete. Second, it highlights the natural tension between relevance (doing more) and efficiency (doing it simply), suggesting that implicatures reside in this gap.
While the paper provides a formal proof-of-concept using illustrative examples (e.g., hard drive safety, swimming in strong waves), it acknowledges several limitations. It remains an open question whether formal rationality can capture the full spectrum of Gricean implicatures. Furthermore, the computational cost of comparing multiple potential plans increases with the number of possible goals. Future work will require empirical validation through corpus analysis to test the approach against real-world human dialogue.
Summary updated