Models of Language Acquisition in Phonology and Morphology
My primary interest is in computational models of language learning and acquisition. I am especially interested in learning problems that are traditionally viewed as difficult but that we know children and other language learners are able to overcome. For instance, much of language learning happens in the absence of any explicit negative data; parents rarely provide children with ungrammatical sentences, and children are generally not told that their utterances are ungrammatical (and tend to ignore or misunderstand the corrections they do receive).
These difficult learning situations (under the umbrella term "poverty of the stimulus") have led many researchers to propose mental scaffoldings, innate structures or knowledge that a language learner can use to provide hints of or limitations to the structures that they are learning. While it is possible that such scaffolds are necessary, advances in learning theory and modeling suggest that they are not so prominent or necessary as they once seemed.
My dissertation research examines the ability of cognitively-general learning mechanisms and biases, such as Bayesian inference with simplicity priors, to learn linguistic structure from realistic data with little reliance on innate knowledge. I focus on Bayesian models of learning phonological constraints and word segmentation [Doyle Levy 2013].
Learning from Multiple Information Types/Sources
For many learning problems, an optimal solution requires the intelligent combination of multiple types or sources of information. In learning to effectively segment the words of their language, infants use such disparate cues as phonemic wordforms, stress patterns, syntactic structure, and eye-gaze of the speaker. Learners put different emphasis on these cues at different ages, and may even overcompensate in adopting a new primary cue. How can they learn which cues are informative, as well as how to adjudicate between cues when they disagree? Such problems are not limited to humans; machine learning algorithms also often benefit from bringing together correlated pieces of information.
My research in these problems include the word segmentation work above [see also Doyle Levy 2011 for a maximum entropy version] as well as applying multimodal learning to classify texts and images using links between the two [Costa Pereira et al 2013, Rasiwasia et al 2010].
I also am interested in issues of speaker choice, how speakers choose between different possible ways of saying something, often in a matter of milliseconds.
What information does a speaker have available when composing the rest of their sentence? What factors affect the speaker's decision making? When is the speaker more concerned
with alleviating production difficulty, and when with improving ease of comprehension? I look primarily at syntactic choice, specfically in the alternations needs to be done/needs doing
[Doyle Levy 2008] and the choice of that or who as a relative pronoun [Doyle Levy 2010].
In addition to my modelling work on langauge acquisition, I also work extensively with topic models, computational models that automatically identify words that occur in similar contexts.
This work includes an extended topic model that accounts for word burstiness [DCMLDA], the application of topic models to financial data [Doyle Elkan 2009],
and mapping texts and images multimedia learning problems [Rasiwasia et al 2010].
[My complete list of publications is available here.]