Introduction
Modern Large Language Models (LLMs) have achieved remarkable fluency, yet they often lack a fundamental human quality: pragmatics. While a human speaker adapts their language based on their goals and the specific needs of their audience, current AI systems primarily function by predicting the most probable sequence of words. This paper proposes a conceptual framework that shifts the paradigm of Natural Language Generation (NLG) from simple probabilistic modeling to a strategic decision-making process. By treating communication as a balancing act between communicative goals, production costs, and listener effort, we can move toward AI that "thinks" before it speaks.
The Pragmatic Gap in Current NLG
The core issue identified in contemporary NLG is the conflation of disparate communicative factors into a single probability value. In human conversation, we do not simply output the most likely tokens; we optimize for clarity and efficiency. For instance, the way a parent explains a concept to a child differs significantly from how a specialist explains the same concept to a peer. Current models, however, lack this nuanced "pragmatic sense."
This gap arises because models do not distinguish between the effort required to generate a sentence (production cost) and the effort required for a human to interpret it (comprehension cost). Without this distinction, AI remains a "stochastic parrot," failing to ground its output in the social and cognitive context of the listener.
A Framework for Strategic Communication
To bridge this gap, the proposed framework introduces three pillars: Communicative Goals, Costs, and Utility.
1. Communicative Goals ($G_s$)
A goal is defined as a function that transforms the current state of the world into an intended future state. For example, in a referential game, the goal is to ensure the listener successfully identifies a specific target image among distractors.
2. The Four Categories of Cost
The framework decomposes "effort" into four distinct metrics:
- Encoding Cost ($C^E$): The cognitive or computational effort required to translate a mental intent into a formal representation.
- Realisation Cost ($C^R$): The physical or computational resources spent producing the words (e.g., the length of the utterance).
- Processing Cost ($C^P$): The effort the listener spends parsing the words, often measured by linguistic "surprisal."
- Decoding Cost ($C^D$): The effort the listener spends grounding those parsed words into the real-world context or the specific task at hand.
3. Utility ($U_s$)
The speaker’s "profit" or Utility is maximized when the communicative goal is achieved while the total joint costs (for both speaker and listener) are minimized. This mathematical formulation allows NLG to be treated as an optimization problem.
Operationalizing Pragmatics
The paper suggests that these theoretical pillars can be operationalized using a combination of modern statistical methods:
- Rational Speech Act (RSA) Models: These models provide a Bayesian framework for optimizing the trade-off between informativeness and effort.
- Reinforcement Learning (RL): RL can be employed to allow models to learn optimal production strategies through experience and feedback.
- Gricean Maxims as Optimization: The framework provides a mathematical basis for Paul Grice’s Maxims of Conversation. For example, the Maxim of Manner (avoiding ambiguity) emerges naturally when a speaker attempts to minimize the listener's Processing and Decoding costs.
Case Studies
While theoretical, the paper applies this framework to two primary NLG tasks:
- Visually Grounded Referential Games: The speaker must describe one image out of a set. A pragmatic model would evaluate potential descriptions not just for accuracy, but for how easily a listener can distinguish the target from distractors.
- Abstractive Text Summarization: The goal is to condense information such that the audience can still answer factual questions about the source text. Here, the utility is defined by the listener's ability to reconstruct the core facts with minimal reading effort.
Conclusion and Future Directions
This position paper unifies disparate fields—formal pragmatics, information theory, and modern NLP—into a single cohesive strategy for NLG. By maintaining a "Theory of Mind" (a model of what the listener knows and how they process information), AI systems can exhibit more human-like linguistic behavior.
Future work must move from conceptualization to large-scale implementation. This includes developing new utility-based decoding algorithms and utilizing Reinforcement Learning to refine production strategies in multi-turn dialogues. Ultimately, the goal is to transition from models that merely predict text to agents that communicate with purpose and efficiency.
Summary updated