Check out my blog, Random Words
I am interested in how children induce linguistic structure from data. There are two parts to this question: what assumptions about grammatical form do children implicitly make (i.e., what are their learning biases), and what is the evidence they rely on? I use structured probabilistic models to quantify how well, and in what ways, different kinds of evidence disambiguate various grammatical forms. In previous and current work, I have looked at how acoustic cues, such as word duration, might provide useful evidence for identifying syntactic structure. I plan to start using such models to evaluate how well different assumptions about morphosyntactic abstraction match child language comprehension and production. I am also interested in expanding my research program to include laboratory experiments that follow up on computational results.
For my dissertation ("Predictability effects in language acquisition"), which has been successfully defended and is under revision, I looked at how children might integrate different kinds of information to learn how to group words into syntactic phrases.
Specifically, I've been looking at two kinds of acoustic information. First, prosody (the rythm and intonation of speech) groups words together, such as words that would be written between commas or parentheses, and these groupings might serve as a sort of initial cue to syntactic groupings. Thus, acoustic cues to prosodic groupings might help with syntax acquisition. This is called the "Prosodic Bootstrapping" hypothesis and has been around as a proposal since at least the 1980's.
Second, it's been known for a long time that talkers tend to reduce words (i.e. pronounce more quickly and less distinctly) if they are highly predictable. More recently, it's been found that words are pronounced more quickly if they are in a low-probability syntactic structure. Thus, a reduced pronunciation provides a learner evidence that the hidden syntactic structure associated with that word is somehow unlikely. My dissertation introduces this as the "Predictability Bootstrapping" hypothesis.
My dissertation first provides some computational motivation for Predictability Bootstrapping as an especially easy kind of bootstrapping, shows that predictability effects exist in child-directed speech. Next, it provides computational modeling experiments that suggest both prosodic phrasing and predictability effects provide useful evidence for syntax acquisition.
Publications and Presentations2013
- John K Pate and Sharon Goldwater (2013). Unsupervised dependency parsing with acoustic cues. In Transactions of the Association for Computational Linguistics (TACL) 1. pp 63—74. [pdf]
- John K Pate and Sharon Goldwater (2011). Predictability effects in
adult-directed and infant-directed speech: Does the listener matter?
In Proceedings of the 33rd annual meeting of the Cognitive Science
- modelselect.R, model selection code for R based on code by Moreno Coco. This function does not work with factor predictors (i.e. predictors for which is.factor() returns true). If you want to use this function with categorical predictors, you'll need to do the contrast coding yourself. (updated September 19 2011).
- John K Pate and Sharon Goldwater (2011). Unsupervised syntactic chunking with acoustic cues: Computational models for prosodic bootstrapping. In Proceedings of the 2nd Workshop on Cognitive Modeling and Computational Linguistics. [Best student paper award] [pdf]
- John K Pate and Sharon Goldwater (2010). Learning Syntax Using Prosody. Poster presented at the SICSA PhD Conference 2010. Edinburgh, Scotland. [pdf (poster)]
- Laura Wagner, John Pate, and Cynthia Clopper (2009). Children’s Knowledge of Various Dialects of English. Poster presented at the Cognitive Development Society meeting, San Antonio, TX.
- Cynthia Clopper and John Pate (2009). Effects of talker and token variability on perceptual learning of dialect categories. In Proceedings of Meetings on Acoustics. [pdf] [preceding poster (pdf)]