Source-linked AI summary
A Computational Approach to Politeness with Application to Social Factors
Cristian Danescu-Niculescu-Mizil, Moritz Sudhof, Dan Jurafsky, Jure Leskovec, Christopher Potts
TL;DR
The paper addresses how linguistic politeness can be identified and how it relates to social context. It builds an annotated request corpus and theory-informed classifier, finding near-human performance and associations between politeness, power, status, gender, and community.
Problem
The paper investigates how linguistic aspects of politeness and their interactions with social factors can be identified computationally.
Method
The authors construct a politeness-annotated request corpus and train a domain-independent classifier using lexical, sentiment, dependency, and theory-based features.
Results
The classifier achieves near-human performance, while analyses associate politeness with election-based status, Stack Exchange reputation, gender, and community membership.
Takeaways & Limitations
The released corpus and classifier support further study of politeness and its relationships with social factors across online communities.
Takeaways & Limitations
The annotation uses an implicit binary politeness judgment, and limited human-labeled Stack Exchange data prevented controlling for the user's role in the reputation experiment.
Abstract
from arXiv · showhide
We propose a computational framework for identifying linguistic aspects of politeness. Our starting point is a new corpus of requests annotated for politeness, which we use to evaluate aspects of politeness theory and to uncover new interactions between politeness markers and context. These findings guide our construction of a classifier with domain-independent lexical and syntactic features operationalizing key components of politeness theory, such as indirection, deference, impersonalization and modality. Our classifier achieves close to human performance and is effective across domains. We use our framework to study the relationship between politeness and social power, showing that polite Wikipedia editors are more likely to achieve high status through elections, but, once elevated, they become less polite. We see a similar negative correlation between politeness and power on Stack Exchange, where users at the top of the reputation scale are less polite than those at the bottom. Finally, we apply our classifier to a preliminary analysis of politeness variation by gender and community.
1 Introduction
The paper develops a computational framework and annotated request corpus for studying politeness markers, their context-sensitive use, and their relationships with social power and group differences.
- Requests expose politeness strategies because they impose on addressees and often invite indirectness or apologies that minimize the imposition.
- The annotated corpus shows that politeness markers interact with morphosyntactic context, including the syntactic position and co-occurring markers of please.
- The classifier uses domain-independent lexical, sentiment, and dependency features to operationalize negative and positive politeness components.
- Near human-level accuracy across domains supports using the classifier to study politeness strategies in new data.
- Polite Wikipedia editors are more likely to gain high status through elections but become less polite after election, while high-reputation Stack Exchange users are less polite.
- Politeness varies across social groups and communities, including substantial differences among programming-language subcommunities on Stack Exchange.
2 Politeness data
The paper builds a large, continuously annotated corpus of online requests from Wikipedia and Stack Exchange, using controlled two-sentence prompts and five annotators per request. Agreement is strongest at the politeness extremes, while middle-quartile judgments are less reliable, motivating quartile-based classes and analyses.
- The corpus contains over 10,000 request utterances from Wikipedia and Stack Exchange, annotated for perceived politeness using Amazon Mechanical Turk.
- Requests contain exactly two sentences, with the second sentence forming the actual question-like request, and each request receives ratings from five annotators.Annotators judged politeness on a slider ranging from “very impolite” to “very polite.”
- Worker scores are z-score normalized before the request politeness score is computed as the average of five annotations.The resulting distribution has mean 0 and standard deviation 0.7 in both domains; positive scores indicate politeness.
- Inter-annotator labels are coherent and significantly differ from randomized scores (p < 0.0001).
- Full binary agreement is more common in the top and bottom politeness quartiles than in the middle quartiles.This indicates that somewhat polite or impolite requests have less reliable scores and motivates using the top quartile as a reference.
3 Politeness strategies
The analysis operationalizes politeness through linguistic strategies that mitigate imposition or build solidarity. It finds cumulative effects among markers and identifies specific interactions involving indirectness, modality, sentiment, gratitude, deference, and sentence-medial please.
- The study analyzes politeness markers in Wikipedia requests using dependency parses, regular expressions, and specialized lexicons.
- Positive politeness: Gratitude and deference signal positive politeness by imposing a social cost on the speaker and paying respect to the addressee.
- Positive politeness: Positive sentiment, greetings, and solidarity-oriented language build a positive relationship, whereas negative sentiment damages it.
- Negative politeness: Indirect requests, including shared first-person plural forms, are generally more polite than imperatives or statements about the addressee’s obligations.Indirectness can combine with hedges to give the addressee a face-saving way to deny the request.
- Modality: Irrealis forms would and could are more polite than the ability or future-oriented forms can and will.
- Marker interactions: Politeness markers often combine cumulatively, while sentence-medial please is polite because it can combine freely with other negative-politeness strategies.
4 Predicting politeness
The paper evaluates whether theory-inspired linguistic features can classify politeness robustly across domains. A linguistically informed classifier improves on a bag-of-words baseline and approaches human performance in-domain, while cross-domain performance leaves more room for improvement.
- Classifier construction: The classifier uses lexical, sentiment, dependency, and politeness-strategy features to automate judgments on large request datasets.The features operationalize both negative and positive politeness, including indirectness, gratitude, sentiment, solidarity, and inclusiveness.
- Experimental setup: Polite and impolite classes are defined as the top and bottom quartiles of politeness scores, producing balanced datasets of 1,089 Wikipedia and 1,651 Stack Exchange requests.The study compares a unigram SVM baseline, a linguistically informed SVM, and human reference performance.
- Classification results: 3-4% absolute improvement over the bag-of-words model is achieved by linguistically informed features in both domains and both evaluation settings.The evaluations include leave-one-out in-domain validation and cross-domain training and testing.
- Classification results: In-domain results fall within 3% of human performance, whereas cross-domain results show greater room for improvement.The authors interpret this pattern as evidence that the features are effective while motivating further study of domain-general politeness cues.
5 Relation to social factors
The framework is used to examine politeness alongside status, dependence, authority, and social-group variation. Politeness is associated with positive future status outcomes but declines after successful promotion, and similar negative relationships with power appear on Stack Exchange.
- Computational social analysis: Classifier-based analyses of roughly 400,000 requests reproduce qualitative results from human-labeled data, supporting computational study of social variables.The framework enables large-scale analysis with controls that would be difficult to obtain from human annotation alone.
- 5.1 Relation to social outcome: Editors who eventually become administrators are more polite than non-admins and failed candidates, with 30% of their requests in the top politeness quartile versus 25% for a random sample.Both comparisons are statistically significant at p<0.001, using 31k requests.
- 5.2 Politeness and power: After elections, successful editors become less polite while unsuccessful editors become more polite.Successful editors receive administrator privileges, whereas losing an election is described as having the opposite effect on politeness.
- 5.2 Politeness and power: On Stack Exchange, question-askers’ comments are more polite than comments from other users, consistent with their informational disadvantage.This dependence analysis covers 181k requests: 106k by question-askers and 75k by answer-givers.
- 5.2 Politeness and power: High-reputation Stack Exchange users are less polite than lower-reputation users, even after controlling for whether they ask or answer questions.The reputation analysis uses 25k question-asker requests.
- Prediction-based interactions: Preliminary analyses find politeness variation by geography, gender, and programming-language subcommunity.The paper reports that U.S. Midwest Wikipedians are most polite among census-defined regions and identifies substantial variation across Stack Exchange programming communities.
6 Related work
Prior work established politeness as a major concern in pragmatics and examined its variation across contexts, media, and social groups. This paper extends that work with domain-independent cues and connects politeness to social outcomes and power relations.
- Politeness research spans pragmatic theory, social meaning, cultural variation, and computational applications including generation, translation, dialogue, and subjectivity analysis.
- Online studies have documented politeness variation across contexts, media types, and social groups; this paper uses orders of magnitude more data to survey strategies more fully.
- Related power-and-status research often relies on domain-specific textual cues, whereas this paper seeks domain-independent politeness cues.
- The paper complements prior work by examining how politeness relates to social outcomes and power relations.
7 Conclusion
The paper releases a large annotated-request collection and a near-human politeness classifier, then applies them to social factors. It presents these resources as a basis for further study while noting that the investigation is only beginning.
- The authors construct and release a large collection of requests annotated for politeness to evaluate key aspects of politeness theory.
- The paper builds a near-human politeness classifier and uses it to study power, status, gender, and community membership.
- The publicly available annotated collection is intended to enable further research on politeness and its relation to social factors.