The Sociolinguistics Lab is pleased to host a talk by Dr. Dan Villareal, University of Pittsburgh. 

Sociolinguistic auto-coding: Applications and pitfalls

Researchers in sociophonetics and variationist sociolinguistics have increasingly turned to computational methods to automate time-consuming research tasks such as data extraction (e.g., Fromont & Hay 2012), phonetic alignment (e.g., McAuliffe et al. 2017), and accurate vowel measurement (e.g., Barreda 2021). In this talk, I discuss the advantages and challenges of using sociolinguistic auto-coding (SLAC), a method in which machine learning classifiers assign variants to variable data (Kendall et al. 2021; McLarty, Jones & Hall 2019; Villarreal et al. 2020; Villarreal under review).

Villarreal et al. (2020) trained random forest classifiers of two sociolinguistic variables of New Zealand English, non-prevocalic /r/ (varying between Present vs. Absent) and intervocalic medial /t/ (Voiced vs. Voiceless), using over 4,000 previously hand-coded tokens (per variable). Cross-validation revealed accuracy rates of 84.5% for /r/ and 91.8% for /t/. In addition to binary predictions, these auto-coders calculate classifier probabilities: the likelihood that a given /r/ token was Present, or a /t/ token was Voiced. In a listening experiment in which 11 phonetically trained listeners coded 60 /r/ tokens, we found a significant positive linear relationship between classifier probability and human judgments; this indicates that classifier probability successfully captures listeners’ perception of phonetically gradient rhoticity. Finally, auto-coders can report which features were most important in classification, helping to shed light on acoustically complex variables like /r/. In short, SLAC can be used for at least three specific functions: binary coding, gradient ‘coding’, and feature selection.

Like other machine learning (ML) methods, however, there are inherent concerns about SLAC’s fairness—that is, whether it generates equally valid predictions for different speaker groups  (e.g., Koenecke et al. 2020). First, given that there are multiple definitions of ML fairness that are mutually incompatible (Berk et al. 2018; Corbett-Davies et al. 2017; Kleinberg et al. 2017), fairness metrics must be decided upon within individual research domains; I argue for three fairness metrics relevant to the domain of sociolinguistic auto-coding. Second, I re-analyze Villarreal et al.’s (2020) /r/ auto-coder for fairness; I find poor performance on all three fairness metrics, with women’s tokens coded more accurately than men’s (88.8% vs. 81.4%). Third, to remedy these imbalances, I used the same data to test a variety of unfairness-mitigation strategies from the ML fairness literature; I find substantial improvement with respect to fairness, albeit at the expense of predictive performance.

Given these fairness issues, I reconsider SLAC under Markl’s (2022) premise that some speech and language technologies are too inherently flawed to use. I argue that while SLAC does not fit into this category, its potential users and consumers deserve a “warts and all” awareness of its drawbacks. To that end, I close with concrete recommendations for using SLAC in large-scale research projects.

Passcode: sociolab