Surprisal Minimisation over Goal-directed Alternatives Predicts Production Choice in Dialogue

1. Problem Statement

The paper addresses a fundamental ambiguity in information-theoretic models of language production: how to interpret cost measures (such as surprisal, uniform information density, and length) as either speaker-oriented (production effort) or listener-oriented (comprehension effort).

Previous models often treat these costs implicitly or assume a fixed set of alternatives without distinguishing between:

• Goal-directed alternatives: Utterances that realize the same communicative intent as the speaker's actual choice.
• Goal-agnostic alternatives: Utterances that are contextually plausible but may realize different communicative goals.

The authors argue that the interpretation of a cost function depends entirely on the set of alternatives over which it is evaluated. Without explicitly defining these alternative spaces, it is impossible to determine whether a speaker is minimizing effort for themselves or optimizing for the listener's understanding.

2. Methodology

A. Theoretical Framework: Cost-Sensitive Choice

The authors formalize language production as a probabilistic choice over contextual alternatives. They define two distinct sets of alternatives for any given context $c$ and choice point:

1. Goal-Agnostic Set ($A_c$): Generated by sampling from a language model conditioned only on the context (and potentially dialogue history). These represent the listener's uncertainty about what the speaker might say next.
2. Goal-Directed Set ($A_{c,g}$): Generated by sampling from a language model conditioned on the context and a fixed communicative goal ($g$). These represent the speaker's uncertainty about how to realize a specific intent.

The production probability is modeled using a softmax rule over cost:
$$P_S(a | c, g) \propto \exp(-\alpha C(a; c))$$
Where $C$ is the cost and $\alpha$ is a sensitivity parameter.

• Speaker-oriented interpretation: Minimizing cost over $A_{c,g}$ (choosing the easiest way to say the same thing).
• Listener-oriented interpretation: Minimizing cost over $A_c$ (choosing an utterance that is easy for the listener to process given their uncertainty).

B. Experimental Setup

• Dataset: The Switchboard Dialogue Act Corpus (naturalistic spoken dialogue).
• Data Processing: 1,342 utterances were selected (10–30 words, statement/question acts). Each was split at the root verb (choice point) into a context ($c$) and a human continuation ($a^*$).
• Alternative Generation:
  • Goal-Agnostic: GPT-4o was prompted to complete sentences with varying amounts of dialogue history (no history, previous utterance, full history).
  • Goal-Directed: GPT-4o was prompted to generate 10 unique paraphrases of the human continuation, constrained to start with the same context.
  • Filtering: An LLM-as-a-judge (GPT-4o) verified that goal-directed alternatives were true paraphrases (98.75% accuracy).
• Cost Measures:
  1. Surprisal: $-\log p(a|c)$ (estimated via GPT-2 Small).
  2. Local Uniformity: Variance of surprisal between adjacent words.
  3. Global Uniformity: Variance of surprisal across the entire sentence.
  4. Length: Number of words in the continuation.
• Alignment: To prevent distributional confounds, the authors used stratified sampling to align the generated alternatives with human utterances regarding length and global uniformity.

C. Analysis Techniques

1. Deterministic Cost Minimisation: Calculating the rank of the human continuation among its alternatives. If humans minimize cost, the human choice should frequently be Rank 1 (lowest cost).
2. Probabilistic Cost Sensitivity: Using a pairwise logistic regression model to predict whether a human continuation is preferred over a generated alternative based on the cost difference ($\Delta C$). This tests for graded sensitivity rather than strict minimization.

3. Key Contributions

1. Explicit Alternative Spaces: The paper provides a principled method to operationalize "speaker" vs. "listener" pressures by explicitly generating goal-directed and goal-agnostic alternative sets using Large Language Models (LLMs).
2. Reinterpretation of Cost: It demonstrates that the same cost metric (e.g., surprisal) yields different predictive powers and interpretations depending on the alternative set used.
3. Scalable Framework: It introduces a scalable pipeline for studying production preferences in open-ended dialogue, moving beyond restricted syntactic variants to rich, contextually grounded alternatives.

4. Key Results

A. Deterministic Analysis (Rank-1 Outcomes)

• Surprisal (Goal-Directed): The strongest predictor. 53.4% of human continuations were the lowest-surprisal option among goal-directed alternatives (3.24x increase over chance). This supports a speaker-oriented view: speakers choose the most predictable realization of their intended goal.
• Surprisal (Goal-Agnostic): Weaker effect (15.2% Rank-1), suggesting speakers do not primarily minimize surprisal relative to the listener's broad uncertainty.
• Length: Showed the strongest relative effect in the goal-agnostic setting (3.69x over chance), suggesting length minimization is a listener-oriented pressure (reducing processing effort).
• Uniformity: Showed mixed results, with Rank-1 outcomes above chance but less consistent than surprisal.

B. Probabilistic Analysis (Logistic Regression)

• Surprisal: The coefficient for surprisal was significantly negative (indicating preference for lower surprisal) and ~7x stronger in the goal-directed condition than in the goal-agnostic condition. This confirms that surprisal functions primarily as a production-side constraint (speaker effort).
• Uniformity: Lower cost (higher uniformity) predicted choice in the goal-directed condition, but the relationship reversed in the goal-agnostic condition (higher cost/lower uniformity was sometimes preferred), indicating complex listener-oriented dynamics.
• Length: Predicted choice in the goal-agnostic condition but had no significant effect in the goal-directed condition, reinforcing the idea that length is a listener-oriented constraint.

5. Significance and Conclusion

The study concludes that surprisal minimization relative to goal-directed alternatives provides the most robust account of human production choices in dialogue. This supports the view that speakers act as rational agents minimizing their own production effort (cognitive control) when selecting among realizations of a fixed intent.

Conversely, length and uniformity appear to function more as listener-oriented pressures, where speakers adjust their output to facilitate comprehension when the communicative goal is not fixed or when considering the listener's broader uncertainty.

Broader Impact:

• Theoretical: Resolves the ambiguity of cost measures by linking them explicitly to the space of alternatives.
• Methodological: Establishes a framework for using LLMs to simulate "what could have been said" (counterfactuals) to test psycholinguistic theories at scale.
• Practical: Suggests that natural language generation systems should optimize for surprisal minimization over goal-directed alternatives to sound more human-like in terms of production efficiency.

Limitations: The study is limited to English, specific syntactic choice points (root verbs), and relies on LLMs for both cost estimation and alternative generation, which may not perfectly mirror human cognition or cover low-resource contexts.