Events / ILST seminar: Student Mini-Talks

ILST seminar: Student Mini-Talks

April 21, 2023
1:30 PM - 3:00 PM

3401 Walnut Street, Room 401B, 3401 Walnut Street, Philadelphia, Pennsylvania 19104

Speaker: Héctor Javier
Title: A variational model of the loss of English OV
Abstract: The notoriously variable word order within VP in Old and Middle English has been analysed by Pintzuk & Taylor (2006), among others, as the competition between two mutually incompatible grammars, henceforth OV -> VO.  Whereas prior work has relied on having the rate of change s emerge from the data, we present a novel methodological contribution to the analysis by independently deriving s using the learning-based Variational Model (VM) of language change (Yang, 2000; Yang, 2002).  This novel application of the VM predicts that the OV -> VO change was inevitable once the homogeneity of the OV linguistic environment was broken in Old English.  Using only independently motivated and empirically determined parameters, the VM predicts the endpoints of the OV -> VO change closely following the empirical data in Pintzuk & Taylor (2006), and additionally predicts the rate of change s consistent with historical records.
Speaker: June Choe
Title: Fast cluster-based permutation test using mixed-effects models
Abstract: Cluster permutation analysis (CPA; Maris & Oostenveld, 2007) is a simulation-based statistical test of the difference between two (or more) time-series over a window of time. CPA is a popular choice of analysis in EEG, eyetracking, and other densely-sampled time data, with implementations in MATLAB (FieldTrip, permutest) and R (permutes, clusterperm, eyetrackingR). However, current implementations face two limitations. First, model size and complexity is a major bottleneck to performance; researchers must often sacrifice the resolution of the data (via binning and downsampling) and/or forgo modelling group-level differences (by dropping random effects). Second, the algorithmic steps of the CPA are hidden inside monolith functions; this makes the procedure inflexible and difficult to diagnose. I showcase how the R package jlmerclusterperm addresses these issues by using the fast MixedModels.jl Julia library as the backend, with a user-centric design that allows an interactive workflow over the individual components of the CPA.