Towards Pragmatic Production Strategies for Natural Language Generation Tasks

Mario Giulianelli

Abstract

This position paper proposes a conceptual framework for the design of Natural Language Generation (NLG) systems that follow efficient and effective production strategies in order to achieve complex communicative goals. In this general framework, efficiency is characterised as the parsimonious regulation of production and comprehension costs while effectiveness is measured with respect to task-oriented and contextually grounded communicative goals. We provide concrete suggestions for the estimation of goals, costs, and utility via modern statistical methods, demonstrating applications of our framework to the classic pragmatic task of visually grounded referential games and to abstractive text summarisation, two popular generation tasks with real-world applications. In sum, we advocate for the development of NLG systems that learn to make pragmatic production decisions from experience, by reasoning about goals, costs, and utility in a human-like way.

Introduction: The Pragmatic Gap in NLG

Most modern Natural Language Generation (NLG) systems, including state-of-the-art Large Language Models, are trained primarily to maximize the probability of the next word in a sequence. While this approach produces remarkably fluent text, it often ignores the pragmatics of human communication. Humans do not merely generate likely sequences; we adapt our language to achieve specific goals while minimizing the effort required for both ourselves and our listeners.

For instance, a teacher explains a concept differently to a child than to a colleague because they are accounting for the "cost" of comprehension. Current models often conflate these nuances into a single probability value, leading to outputs that are pragmatically "deaf" to the listener's needs or the specific context. This paper addresses the lack of a unified framework to design systems that can reason about these trade-offs explicitly.

A Formal Framework for Strategic Communication

The proposed framework treats communication as a strategic balancing act based on three core pillars:

  1. Communicative Goals ($G_s$): This is defined as a function that transforms the current state of the world ($w$) into an intended future state ($w^*$). This might involve changing a listener's belief or shifting their attention to a specific object in a visual field.
  2. Costs: The framework categorizes effort into four distinct types:
  3. Encoding Cost ($C^E$): The speaker's mental effort to map a goal to a representation.
  4. Realisation Cost ($C^R$): The physical or computational effort of producing the utterance (e.g., length or syntactic complexity).
  5. Processing Cost ($C^P$): The listener's effort to parse the words, often quantified as "surprisal."
    • Decoding Cost ($C^D$): The listener's effort to ground those words in the real world to understand the speaker's intent.
  6. Utility ($U_s$): A measure of success where utility is directly proportional to how well the goal is achieved and inversely proportional to the total costs incurred by both parties.

By plugging in modern statistical methods, such as neural language models and computer vision, we can estimate these abstract costs in real-world scenarios.

Communication as Optimization

This framework frames language production as a decision-making problem. A "good" utterance is one that maximizes utility by balancing effectiveness (getting the job done) with efficiency (not wasting resources). This approach allows us to map Paul Grice’s classic "Maxims of Conversation" to a mathematical optimization problem. For example, the Maxim of Quantity—being as informative as required—is viewed as the optimization of realization and processing costs.

A core argument of the paper is the "Theory of Mind" gap. For an NLG system to be truly pragmatic, it must maintain a mental model of the listener to estimate the listener's comprehension costs ($C^P$ and $C^D$). Current models like GPT-3 often conflate all costs into a single probability; the author suggests that future research should utilize Reinforcement Learning (RL) to help models learn strategies for weighing these costs over multiple turns of dialogue.

Case Studies: Reference Games and Summarization

To demonstrate the framework's versatility, the paper explores two distinct applications:

  • Reference Games: In tasks where a speaker must help a listener identify a specific image, costs can be quantified using image segmentation uncertainty (encoding) and cumulative surprisal (processing). Success is measured by the listener's ability to correctly identify the target.
  • Abstractive Summarization: In headline generation or text summarization, utility can be measured by how accurately a listener (or another model) can answer questions about the original text after reading only the summary. This moves the metric of success from simple n-gram overlap (like ROUGE) to functional utility.

Conclusion and Future Directions

As a position paper, this work provides a theoretical roadmap rather than empirical results. It advocates for the development of more sophisticated decoding algorithms, such as Minimum Bayes Risk (MBR) decoding, which can explicitly weigh utilities and costs rather than simply following the highest probability path.

While defining the "state of the world" remains difficult for open-ended tasks, the author suggests that combining Reinforcement Learning with the Rational Speech Act (RSA) model is the most promising path forward. This would allow agents to learn pragmatic strategies from experience and observation, moving NLG closer to the efficiency and purpose of human speech.

Summary updated

Abstract

This position paper proposes a conceptual framework for the design of Natural Language Generation (NLG) systems that follow efficient and effective production strategies in order to achieve complex communicative goals. In this general framework, efficiency is characterised as the parsimonious regulation of production and comprehension costs while effectiveness is measured with respect to task-oriented and contextually grounded communicative goals. We provide concrete suggestions for the estimation of goals, costs, and utility via modern statistical methods, demonstrating applications of our framework to the classic pragmatic task of visually grounded referential games and to abstractive text summarisation, two popular generation tasks with real-world applications. In sum, we advocate for the development of NLG systems that learn to make pragmatic production decisions from experience, by reasoning about goals, costs, and utility in a human-like way.
BibTeX
@inproceedings{giulianelli2022towards,
  author = {Mario Giulianelli},
  title = {{Towards Pragmatic Production Strategies for Natural Language Generation Tasks}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2022},
  url = {https://openreview.net/forum?id=6J2ZHETTp1},
}
Figure