Source-linked AI summary
Echoes of power: Language effects and power differences in social interaction
Cristian Danescu-Niculescu-Mizil, Lillian Lee, Bo Pang, Jon Kleinberg
TL;DR
The paper addresses the challenge of finding domain-independent language signals of social power in natural-language interaction. It proposes linguistic coordination as an analysis framework and finds consistent power-related coordination patterns across Wikipedia discussions and Supreme Court arguments.
Problem
Domain-independent language-based signals of roles, status, and power in natural-language group interactions have been difficult to identify.
Method
The paper uses language coordination in text content to analyze power differences across status-based and situational power relationships.
Results
Language coordination consistently reveals both types of power differences: lower-power people coordinate more, and people coordinate more with higher-power interlocutors.
Takeaways & Limitations
Linguistic style variations can provide information about power differences in consequential, goal-oriented discussions across diverse domains and settings.
Takeaways & Limitations
The two coordination hypotheses are not logically equivalent because the paper uses an asymmetric definition of coordination.
Abstract
from arXiv · showhide
Understanding social interaction within groups is key to analyzing online communities. Most current work focuses on structural properties: who talks to whom, and how such interactions form larger network structures. The interactions themselves, however, generally take place in the form of natural language --- either spoken or written --- and one could reasonably suppose that signals manifested in language might also provide information about roles, status, and other aspects of the group's dynamics. To date, however, finding such domain-independent language-based signals has been a challenge. Here, we show that in group discussions power differentials between participants are subtly revealed by how much one individual immediately echoes the linguistic style of the person they are responding to. Starting from this observation, we propose an analysis framework based on linguistic coordination that can be used to shed light on power relationships and that works consistently across multiple types of power --- including a more "static" form of power based on status differences, and a more "situational" form of power in which one individual experiences a type of dependence on another. Using this framework, we study how conversational behavior can reveal power relationships in two very different settings: discussions among Wikipedians and arguments before the U.S. Supreme Court.
1. INTRODUCTION
The paper addresses the challenge of inferring power differences from natural language rather than interaction structure or explicit status markers. It proposes linguistic coordination as a domain-independent signal and finds consistent relationships with status-based and situational power across Wikipedia and Supreme Court discussions.
- Language coordination: Linguistic coordination measures how speakers unconsciously echo interlocutors’ function-word choices, such as articles, prepositions, and pronouns.The approach focuses on language style rather than domain-specific substantive content.
- Power and linguistic style: Language coordination reveals power differences consistently across both datasets and across status-based and situational forms of power.The framework links differences in coordination to multiple forms of power in discussions among Wikipedians and before the U.S. Supreme Court.
- Power differences from language coordination: People with low power generally exhibit greater language coordination, while people coordinate more with higher-power interlocutors.These two directional findings describe coordination both by the speaker’s own power and by the respondent’s relative power.
- Power differences from language coordination: Coordination changes when a person’s status changes, including when a Wikipedian becomes an administrator, and also increases when someone tries to convince an opponent.The latter situation is interpreted as dependence and therefore a power deficit in exchange theory.
- Power differences from language coordination: The relationship between status level and language coordination transfers across domains and is a reliable cross-domain feature for status prediction.Using two substantially different corpora allows the study to test domain independence.
- Applications and novelty: The methods could identify power and status from text alone without domain-specific knowledge or explicit status markers and links.Function-word-based analysis can also operate after content words are redacted, with implications for compact representations and privacy.
- Applications and novelty: The work is presented as the first large-scale study connecting language coordination with social power across diverse individuals and domains.It also examines status change and situation-dependent power as novel applications of textual coordination analysis.
2. COORDINATION AND POWER
Communication accommodation theory motivates the principle that lower-power speakers coordinate more with higher-power targets. The section formulates target- and speaker-based hypotheses while distinguishing power imbalance from personal baseline coordination and noting competing prior evidence.
- Principle P states that coordination increases as the speaker’s power decreases relative to the target’s.
- The target hypothesis predicts greater coordination toward high-powered than low-powered people.
- The speaker hypothesis predicts that high-powered people coordinate less toward their targets than low-powered people.
- Because coordination is asymmetric, neither the target hypothesis nor the speaker hypothesis implies the other.
- Personal characteristics such as sociability may establish baseline coordination levels distinct from power imbalance.
- Prior evidence was intriguing but small-scale, and competing theories predict that high-status individuals may coordinate more to be understood.
3. POWER RELATIONS IN WIKIPEDIA AND SUPREME COURT DATA
The study compares power relations in large-scale Wikipedia discussions and Supreme Court oral arguments, combining status differences with dependence-based contrasts. The corpora differ in collaboration, modality, and domain, supporting tests of domain-independent relationships between power and linguistic coordination.
- Supreme Court arguments provide verbal, offline interactions between Justices and lawyers in a less collaborative setting with explicit opposing sides.
- Administrators have higher status than non-admins and can access restricted technical features such as protecting or deleting pages.
- The two corpora contain 240,436 Wikipedia conversational exchanges and 50,389 Supreme Court verbal exchanges for 204 cases.
- Wikipedia provides asynchronous, collaborative discussions with rich status and social-interaction metadata, including editor status changes.
- Dependence is operationalized by comparing coordination toward opposing versus aligned audiences in debates and toward unfavorable versus favorable Justices.
4. LINGUISTIC STYLE COORDINATION
The paper measures immediate linguistic-style coordination through function-word usage and extends an asymmetric person-to-person measure to groups. It uses multiple aggregation strategies to handle missing marker-level values and formulates power predictions over group coordination.
- Linguistic style coordination measures whether a person’s function-word usage immediately echoes the person they are replying to.
- The style markers comprise eight LIWC-derived function-word categories totaling 451 lexemes, excluding negation because it is sparse and semantic.
- For marker m, coordination measures how much the target’s marker usage increases the probability of that marker in the speaker’s next reply.
- The measure is asymmetric and depends on both participants’ behavior, matching the asymmetric power relations under study.
- Group coordination averages speakers’ coordination toward target groups, using a macro average so highly active speakers do not dominate.
- Three aggregated measures address missing marker values by restricting cases, smoothing with group averages, or averaging only defined markers.
- The formal predictions compare coordination toward high- versus low-powered groups and coordination by high- versus low-powered speakers.
5. EMPIRICAL INVESTIGATION
Across Wikipedia and Supreme Court discussions, linguistic coordination tracks power differences associated with status and dependence. Wikipedia analyses further separate status effects from personal characteristics and show coordination changes after promotion.
- Power from status: Wikipedia: Users coordinate more toward higher-powered admins than toward non-admins, supporting Ptarget.
- Power from status: Wikipedia: The Wikipedia status result was not mainly wholesale phrase repetition: conjunction trigram repeats occurred in 0.7% of exchanges and bigram repeats in 3.5%, while the difference remained significant after excluding them.
- Power from status: Wikipedia: Admins coordinate more than non-admins as speakers, contrary to the prediction that higher-powered people would coordinate less.
- Power from status: Wikipedia: Admins-to-be already coordinated more than persistent non-admins before promotion, indicating that personal characteristics contribute to the status-related difference.
- Power from status: Wikipedia: Among users who sought adminship, admins-to-be coordinated more than failed-to-be during their RfA discussions, separating success from shared ambition.
- Power from status: Wikipedia: After promotion, admins-to-be coordinated less as speakers but received more coordination as targets, with the largest change appearing in the second month.
- Power from status: Supreme Court: In Supreme Court arguments, lawyers coordinated more toward Justices, including significantly more toward the Chief Justice than toward Associate Justices.
- Power from dependence: Dependence also shaped coordination: lawyers coordinated more toward unfavorable Justices, while favorable Justices coordinated more toward lawyers.
6. CROSS-DOMAIN ANALYSIS AND INTERACTION AMONG HYPOTHESES
The analysis tests whether linguistic coordination generalizes across domains for predicting status, then examines how coordination interacts with communication volume and gender. Coordination features are the only feature set that generalizes reliably across the Wikipedia and Supreme Court datasets.
- 6.1 Coordination as a Cross-Domain Feature: The prediction task classifies which member of a different-status pair has higher status, using 50% accuracy as the random-guess baseline.Cross-domain experiments train on one domain and test on the other, whereas in-domain experiments use cross-validation within one domain.
- 6.1 Coordination as a Cross-Domain Feature: Coordination features encode whether one participant coordinates more than the other across stylistic markers and an aggregated feature.The classifier also receives separate stylistic-frequency and bag-of-words feature sets for comparison.
- 6.1 Coordination as a Cross-Domain Feature: Coordination features are the only ones significantly better than random guessing in cross-domain status prediction.Other classifiers learn domain-specific cues and fail to generalize; bag-of-words performs worse than random guessing when transferred across domains.
- 6.1 Coordination as a Cross-Domain Feature: Coordination features are the only feature set significantly better than random guessing on both in-domain datasets.The datasets compare Wikipedia admin/non-admin pairs with Supreme Court Justice/lawyer pairs.
- 6.2 Interactions among Hypotheses P and B: On Wikipedia, higher communication volume is associated with greater coordination, while the Supreme Court data show the inverse relation.The Wikipedia association also holds when comparing users with the same status, including admins and non-admins.
- 6.2 Interactions among Hypotheses P and B: Figure 6 reports that female lawyers coordinate more toward Justices, while Justices coordinate more toward male lawyers.The passage identifies gender-based status or communication differences as competing interpretations, making the distinction between Hypotheses P and B challenging.
7. FURTHER RELATED WORK
Related work spans structural approaches to power and status, linguistic coordination studies, and research on collaborative online communities. Earlier coordination studies were generally small-scale or domain-specific, while Wikipedia has served as a major testbed for online interaction and promotion research.
- Related work: The paper builds on earlier work using text content to analyze online networks and on research concerning language coordination and power.The related-work discussion situates these topics alongside exchange-theoretic notions of status and dependence.
- Power and structural features: Prior work inferred network importance and online status primarily from structural features rather than language.The cited literature includes social-network and Web studies of importance, plus online social-network status inference.
- Power and language: Earlier studies of linguistic coordination and status examined small datasets such as 15 Watergate transcripts, 40 courtroom cases, or one simulated trial.A large-scale Twitter study found only a weak correlation between coordination and follower counts.
- Collaborative communities: Wikipedia has been used extensively to study large-scale user interaction and promotion processes in collaborative communities.Related work also examines reviewer behavior and participation incentives on commercial review sites.