Tom Lentz, assistant professor at the University of Amsterdam, is the guest speaker at the ACLC seminar on Friday 12 April 2019 at 16.15.
Machine learning as a tool for linguistic pattern comparisons: two applications in phonetics/phonology
The use of machine learning allows to identify patterns in complex data without strong a priori definition of such patterns. Hence, it is possible to test whether certain factors or conditions have an effect on such patterns, similar to how well-known statistic tests allow to test this for one-dimensional data, e.g., reaction time. I will illustrate this by discussing two cases. First, coarticulation patterns measured using electromagnetic articulography (EMA) are compared for German and Georgian. Such data is complex as there are at least six landmarks per consonant, each with four dimensions. Using Support Vector Machines (SVMs), these patterns can however be separated (allowing even limited insight in components of the patterns). The SVMs show that and how German participants are more capable of imitating the Georgian pattern (with lower overlap) than v.v. Second, pitch contours of three types of sentences, with overall (neutral) focus, contrastive focus on the last phrase or focus earlier in the sentence are separated using SVM and sequence learning networks. The sentences were elicited also to analyse articulatory data; SVMs allow a quick and robust check that participants at least produced three different patterns without having to define what these patters were, thus confirming that at least that part of the experiment was successful. In addition, sequence learning networks were useful to a limited extent to identify time points in the sentence where the pitch contour is informative.
More information can be found on the personal page of Tom Lentz
PC Hoofthuis, Room 5.02
1012 VB Amsterdam