Cognitive Model of Polite Speech

• Cognitive models of polite speech formalize how agents, human or artificial, generate and interpret language to manage social goals like politeness while also being informative.
• These models leverage linguistic theories, especially Brown and Levinson's politeness theory focusing on face management, and are empirically grounded through computational methods like annotation, feature engineering, and machine learning.
• Politeness deployment is strategic and context-dependent, dynamically adapting based on factors like social status, community norms, and inferred communicative intentions, as shown in studies on Wikipedia and Stack Exchange users.

A cognitive model of polite speech formalizes how communicative agents—human or artificial—generate and interpret utterances to achieve social goals while balancing other objectives such as informativity and self-presentation. Such models draw on foundational theories in linguistics and pragmatics, particularly the framework by Brown and Levinson (1987), and have achieved empirical grounding and computational instantiation through extensive annotation, feature engineering, and machine learning. Politeness, in this context, is not a fixed property but emerges from the interplay of linguistic strategies, intentions, social context, and dynamic adaptation to hierarchical and community norms.

1. Foundations: Politeness Theory and Core Concepts

Politeness theory, most prominently explicated by Brown and Levinson (1987), conceptualizes politeness as strategies that manage "face"—the public self-image individuals claim. Face is divided into two universal dimensions:

• Positive face: The desire for one's wants, values, and identity to be appreciated and accepted by others.
• Negative face: The desire to be free from imposition and to act unimpeded.

Speech acts may threaten (Face-Threatening Acts/FTAs) or support ("face-enhancing" acts) positive or negative face, for either the speaker (S) or hearer (H), giving rise to a typology (e.g., HNEG− for impositions, HPOS+ for agreement, SNEG+ for asserting autonomy).

Politeness strategies mitigate FTAs by using indirectness, hedging, gratitude, apology, impersonalization, or positive forms such as praise or inclusive language. The deployment of these strategies is sensitive to the social context, including status, distance, gender, and community-specific norms.

2. Computational Frameworks: Annotation, Feature Engineering, and Classifier Design

Large, annotated corpora of requests (Wikipedia and Stack Exchange) enable empirical study. Human annotators rate the politeness of requests using continuous scales, z-score normalized for consistency and used to partition examples into quartiles—top and bottom forming "polite" and "impolite" classes (Danescu-Niculescu-Mizil et al., 2013).

Key linguistic features operationalize politeness theory:

• Indirection: Indirect prompts, hedges (e.g., "suggest", "wonder"), modal verbs in irrealis ("could you", "would you"), and positional cues for markers such as "please".
• Deference: Expressions of gratitude ("thank you") and explicit acknowledgments.
• Impersonalization: Avoidance of direct "you", favoring passive constructions or first-person plurals ("we").
• Modality: Modal verbs differentiating strength of imposition, with "could"/"would" being more polite than "can"/"will".

Features are extracted using dependency parsing, regular expressions, and lexica (e.g., Hyland's hedges, Liu's sentiment lexicon). A machine learning pipeline using an SVM with both bag-of-words and linguistically informed features calibrates outputs to probabilistic politeness scores via logistic regression. This approach achieves near-human performance and robust domain transfer, evidencing that theoretical constructs are computationally realizable with high fidelity.

3. Mapping Politeness to Intentions and Discourse Function

Every face act is underpinned by a communicative intention, frequently proxied by dialog acts (e.g., request, inform, question, directive) (Soubki et al., 2024). The mapping between dialog acts and face acts reveals that the nature of an intended action—whether a request or opinion—determines which face components are threatened or supported, and thus which politeness strategies are optimal.

Computational models integrating dialog act annotations improve the identification of rare or complex face acts (macro F1: 0.60→0.63; micro F1: 0.69→0.73). This reflects the importance of modeling not just surface cues but also the underlying communicative goals that motivate polite or impolite action. Such joint modeling is most beneficial for disambiguating subtle intentions in multi-strategy and multi-label settings.

4. Social Hierarchies, Power, and Strategic Deployment

Empirical evidence demonstrates that politeness is dynamically adapted to social power and hierarchical context. For example, Wikipedia editors who later achieve administrative status are significantly more polite pre-promotion, decreasing politeness when promoted—consistent with Brown & Levinson's prediction that higher power reduces the incentive for politeness (Danescu-Niculescu-Mizil et al., 2013). Stack Exchange users with higher reputation also exhibit less politeness, after controlling for interaction role.

Gender and community are salient variables: women are generally more polite; community-based and regional norms yield systematic differences in politeness strategies. Therefore, any comprehensive model must account for individual, group, and status-based priors that influence the frequency and type of politeness deployed.

5. From Linguistic Cues to Cognitive Modeling

The operational computational models provide concrete support for a view of polite speech as both strategic and contextually adaptive resource. Classification features map directly to theoretical components:

Theory Component	Feature(s)	Empirical Effect
Indirection	"Could you", "Would you", hedges	Higher politeness, correlates with lower power
Deference	Gratitude, positive lexicon	Associated with upward mobility in status
Impersonalization	Avoidance of "you", passives	Reduces face-threat, aligns with positive politeness
Modality	Counterfactual modals	Strong, consistent modulation of politeness
Social Factors	Status, gender, community	Systematic variation and adaptation across settings

Furthermore, cognitive models should implement reasoning about cost-benefit trade-offs: agents deploy politeness strategies when the social benefit justifies the increased linguistic or cognitive effort, and adapt policies post-status change or across social boundaries. These mechanisms echo both rational pragmatics and utility-maximizing frameworks.

6. Towards Generalizable, Contextual Cognitive Models

The cumulative evidence supports a model of polite speech comprising the following modules:

1. Status and Context Perception: Encoding of interlocutor roles and expectations, including dynamic contextual updates post-status changes or in community entry.
2. Intention Attribution: Inferring and selecting underlying communicative goals, which direct face management strategy.
3. Strategy Selection and Realization: Mapping intentions and social context to optimal linguistic forms, balancing solidarity (positive face) and non-imposition (negative face).
4. Cost–Benefit Reasoning: Politeness is deployed as a resource—its use is predicted by hierarchical incentives, reputation-seeking, and adaptation to group norms.
5. Adaptive Learning: Variation by gender, subgroup, and region is explained via learned priors, incorporated into both production and interpretation.

These elements together yield a comprehensive cognitive account of polite speech, uniting classic linguistic theory with empirical, algorithmic, and statistical perspectives. The practical realization of such models has advanced the capacity to classify, generate, and adapt polite language in real-world, socially complex environments, furnishing a platform for further research in both cognitive science and computational social interaction.