Computing Social Meaning

from Micro to Macro

Rob Voigt

Stanford Linguistics

robvoigt@stanford.edu

Computational Linguistics

and

• Operationalize rich linguistic cues computationally
• to study language in social context at scale
• accumulation across populations and time
• generalization and variation across contexts

Pushing Beyond Prediction

• A "pure" computational approach is not enough
  • e.g., build a classifier to
    predict gender/race from language
• Correlations \(\neq\) social meanings;
  meaning-making is an interpretive act
• To understand why rather than simply what,
  we need to grapple with the micro!

Today's Talk

Study 1: Quantifying Large-scale Disparity

Study 2: Exploring the Meanings of a Variable

Study 1

Computational Understanding
of Police Officer Respect

Voigt et al. (2017), Proceedings of the National Academy of Sciences

with Nicholas P. Camp, Vinodkumar Prabhakaran, William L. Hamilton,
Rebecca C. Hetey, Camilla M. Griffiths, David Jurgens,
Dan Jurafsky, and Jennifer L. Eberhardt

Police-Community Interaction

• Media focus on explosive incidents,
  research focus on outcomes
• But what's happening on the ground?
• 25% of adults interact with police during a year
  Eith and Durose (2011)

Procedural Justice

• A person treated with respect:
• has more trust in the officer’s fairness
• and the fairness of the institution
• and is more willing to support or
  cooperate with the police
  (Tyler 1990; Tyler and Huo 2002; Mazerolle et al. 2013)
• Black community members report:
  • more negative experiences with the police
  • and being treated with less respect
    (Huo and Tyler 2001; Peffley and Hurwitz 2010; Epp et al. 2014)

Study 1: Questions

How is respect instantiated linguistically in the police-community interactional context?

Are black community members treated with less respect by officers than is afforded white community members?

Previous Work on Procedural Fairness

• citizens’ recollection of past interactions
  (Epp et al 2014)
• researcher observation of officer behavior
  (Mastrofski et al 2009; Dai et al 2011;
  Jonathan-Zamir et al 2015; Mastrofski et al 2016)

• These methods are invaluable but indirect
• ... and the presence of a researcher may influence police behavior
  (Mastrofski and Parks, 1990)

This Work: Footage as Data

• Oakland PD has been wearing body cameras since 2010, usually used only as evidence

• ... but, a window into everyday behavior!

Study 1: Data

• Choose to focus on traffic stops,
  black and white community members
• 981 stops by 245 officers in April 2014
• 70% black drivers; 183 hours of footage
• Professionally transcribed and diarized:
  36,738 officer utterances, 350k+ words

"Thin-slice" Human Social Evaluations

• Formal, Friendly, Polite, Respectful, Impartial
• 414 utterances,
  10 raters each
• High consistency: Cronbach's \(\alpha\) 0.73-0.91

The Latent Space of Respect

 

 variance explained:

Formal

Friendly

Polite

Respectful

Impartial

Respect

 71%

0.27

0.47

0.49

0.47

0.50

Formality

 22%

0.91

-0.39

-0.04

0.03

-0.11

Computational Modeling of Respect

Use human judgments as training data
for a machine-learned model of respect
by operationalizing theories of politeness

• Linguistic theories of politeness from pragmatics focus on requests
  • Requesting that you do something is face-threatening
    (Goffman 1967; Lakoff 1973; Culpepper 1976; Brown and Levinson 1978)

Features: Operationalizing Politeness

Computational Model

• Linear regression predicting Respect
• 32 hand-engineered linguistic features
  lexicons, regular expressions, dependency features,
  more complex functions (e.g., "bald commands")
• 
  log-tranformed counts of features per utterance
• stepwise removal of uninformative features
• Reasonable R²: 0.258

What does the model learn?

Example	Score
Sorry to stop you. My name's Officer [name] with the Police Department.	0.84

Example	Score
There you go, ma'am. Drive safe, please.	1.21

Example	Score
It just says that, uh, you've fixed it. No problem. Thank you very much, sir.	2.07

Example	Score
Where are you guys coming from? Don't lie.	-0.57

Example	Score
So let me see that registration stuff, bro.	-1.03

Results Across the Entire Dataset

• Estimate respect for all 37k officer utterances
• Hierarchical mixed-effects model
  • predict respect from contextual and social variables
  • controlling for officer-level and stop-level variation
• Results - officers are more respectful:
  • with older community members
  • when a citation is issued
  • with white community members

Respect Through the Interaction

Controls

Replication(s)

• Original study participants
  were Stanford undergraduates
• Full replications:
  (Camp, Voigt, et al. in prep)
  • Same population, new stimuli
  • DMV participants, same stimuli

Study 1: Take-aways

Study 1: Ongoing Work

• Officer Prosody
• Officer-Community Discourse Structure
• Misunderstanding & Divergent Interpretations

Study 2

Multimodal Analysis of
Gesture and Embodiment

Voigt et al. (2016), Journal of Sociolinguistics

with Penelope Eckert, Robert J. Podesva, and Dan Jurafsky

Linguistic Approaches

to Gesture and Embodiment

• Long acknowledgement of the body's
  meaning-making capacity
  e.g., Hall's proxemics (1963, 1966); Birdwhistell's kinesics (1952, 1970)
• "Two parts of one system"
  (McNeill 1992, 2008; Kendon 1995, 2004)
• Largely observational and manual coding, e.g.:
  Mendoza‐Denton and Jannedy (2011), Loehr (2012),
  Mondada (2016), Pratt and D'Onofrio (2017)

Gestural Coding

• Complex and non-standardized annotation schemes
• Highly time and skill intensive
• Result:
  Detailed analyses,
  but small sample sizes

Kipp et al.'s (2007) 3D pose
annotation scheme for hand position

Computational Annotation

Overall Body Movement
and Prosodic Engagement

Voigt, Podesva, and Jurafsky (2014)

Smiling
and Vowel Fronting

Podesva, Callier, Voigt, and Jurafsky (2015)

This Work: Head Cant

Goffman (1979) in advertising
Costa et al (2001) in art

Study 2: Questions

Is head cant really about gender?

Is it used differently across interactional contexts?

What communicative purposes can it serve?

Study 2: Data

Computational Extraction of Head Cant

• Use a shape-fitting model to find key points on the face
  (Kazemi and Sullivan 2014)

Distribution by context and gender

• More cant in dialogues
• More cant when not speaking
• Men use it more!

Cant and Prosody

Related to prosodic engagement? (e.g. Jeon et al 2010; Schuller et al 2010; Wang and Hirschberg 2011)

F0

Intensity

Quantitatively-guided
Qualitative Analysis

• Extract phrases which match the trend,
  e.g., high cant + high F0 + low intensity

Framing of Shared Understanding

Joint variation:
high cant, high F0, increased
speech rate

Confirmation: Discourse Particles

• you know and I mean:
  shared understanding, "other orientation"
  Schiffrin (1987)
• um, uh, and like:
  fillers, hesitation, holding the floor
  Clark and Fox Tree (2002); Andersen (1998)
• mm, mmhm, and yeah:
  acknowledgement, engagement, taking the floor Jefferson (1984); Lambertz (2011)

Confirmation: Discourse Particles

Study 2: Take-aways

Discussion

and what's to come

Micro and Macro

• Expanding the scope of computational linguistics
• Central problem of interpretation
• Underexplored - much to learn at both scales!

Computational Fieldwork?

• Looking beyond the standard web sources
  e.g. RtGender, Voigt et al (2018)
• Laboratory and experimental data as corpora
• Long-term relationships with real-world actors
  • can be very time-consuming: fingerprinting, background checks, multi-year process
  • can be very messy: real-world data
    doesn't have an API

Computation and Impact: Policing

• Provides concrete strategies for officers
• Cooperation with Oakland to integrate results into procedural justice training
• ... and we can measure impact

Procedural Justice Elsewhere

• Language in AZ Eviction Courts
  with Daniel Bernal et al, Stanford Law
• RCT with informational mailers
• Linguistic questions:
  • Does more information help tenants be more assertive?
  • Are judges more dismissive of L2 English speakers?

Impact and Meaning in Society

• Many social problems are problems of meaning
  

  How do we interpret one another's behavior?
  ... recognize one another's intentions?
  ... validate one another's personhood?

  How do those interpretations accumulate?

• Linguists are uniquely positioned to contribute!

The common misconception is that language use has primarily to do with words and what they mean. It doesn't.
It has primarily to do with people and what they mean.
Clark and Schober (1992)