📄 Generating Politeness in Task Based Interaction: An Evaluation of the Effect of Linguistic Form and Culture

Summary

The paper describes POLLy (Politeness for Language Learning), which generates task-oriented dialogue with levels of politeness that follow Brown & Levinson's (B&L) theory. The long-term goal is a virtual environment where learners of English as a second language can practise politeness with agents. A GraphPlan-based AI planner builds a plan for a cooking task, such as making pasta. The plan steps are split between two agents, and these agents talk about them through speech acts such as Requests, Offers and Informs. A spoken language generator then realises each act with one of four B&L strategy families: Direct, Approval, Autonomy and Indirect. It does this by transforming RealPro DSyntS trees. In the evaluation, British and Indian native speakers of English rated the generated utterances. The ratings largely confirmed B&L: speech to strangers needs more politeness than speech to friends. However, the off-record Indirect forms, which theory ranks as politest, were judged the rudest. Indian raters found utterances more polite overall than British raters did. This suggests that politeness norms differ even among native speakers of the same language.

Key points

• B&L model the face threat Θ as the sum of three factors: the hearer's power (P), social distance (D) and the imposition rank of the act (R). The strategy is chosen by escalating levels of Θ, following the four-way scheme of Walker et al. (1997).
• The architecture has two parts: a STRIPS-style GraphPlan planner, and a generator with content planning, sentence planning and surface realisation. RealPro realises unordered, lexicalised DSyntS dependency trees.
• Strategies are named variants for Request and Inform acts, such as RD1Imperative, RAp3AsPossible, RAu2ApologizeQModAbility and RI1AsNegation. They are built by adding four kinds of politeness formula to the tree: address forms ("mate"), subject abstraction ("someone should…"), softeners ("please", "possibly") and additives (apologies, "must", "for me").
• Forms with the insisting "must" were rated least polite (df=3, F=185.6, p<0.001).
• Indian raters were more lenient than British raters, especially when a friend was speaking (F=6.0) and for Requests (F=6.37). Ratings for strangers were nearly the same in both groups. This contrasts with Johnson et al. (2005), who found American and German perceptions to be identical.

Interesting ideas

• I1. The planner's structure determines content: direct forms come from plan steps, while indirect forms come from preconditions or negated effects ("X is not done yet"). This links each surface strategy to a specific piece of the action model.
• I2. Politeness is added as small, named transformations on a base dependency tree (inserting softeners, address terms or an abstract subject), not by writing separate templates.
• I3. Face threat is scored additively from power, distance and imposition, and the score selects a strategy tier. This gives a transparent numeric link between social context and linguistic form.
• I4. The evaluation compares theory-predicted politeness with human ratings and splits raters by culture. This exposed both a false prediction (indirect forms seen as rude) and differences between groups.
• I5. Each speech-act type has a default implicit subject: "I" for offers and informs, "you" for requests, "we" for suggestions. This keeps the generator's input minimal.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): The same mapping can run in reverse. Mova could recognise that an assertion of an unmet effect or precondition works as an indirect request, and explain the speech act by pointing to the action frame involved.
• I2 (back): Mova's generator could apply register and politeness as composable tree edits on the meaning graph's realisation. Each edit would be logged, so the style of any output stays explainable.
• I3 (frames): The prag crate could annotate speech acts with an estimated P/D/R threat and the observed strategy tier. Mismatches between the two would flag likely rudeness, irony or unusual deference.
• I4 (learn): Evaluations of pragmatic judgements should break results down by annotator population and test explicit theoretical predictions. That way, systematic theory failures show up instead of being averaged away.
• I5 (back): Mova's generation from frames could use per-act subject defaults like these to produce fluent dialogue turns without needing a full agent specification.