Probabilistic linguistic representations: between learning and processing
Committee: Alec Marantz (chair), Gillian Gallagher, Maria Gouskova, Liina Pylkkänen and Florian Jaeger (University of Rochester)
Our interactions with language often lead us to consider multiple hypotheses simultaneously. As we read a sentence, we anticipate upcoming structure; there are normally many such continuations. When we acquire a language, our input is typically compatible with many hypotheses about the rules of the language. This dissertation explores the consequences of maintaining uncertainty over representations in language comprehension and learning.
The first part of the dissertation investigates how readers’ or listeners’ uncertainty about linguistic structures affects comprehension difficulty. Following a theoretical discussion of uncertainty in language comprehension, three empirical studies are presented. A self-paced reading study found that words that reduce the reader’s uncertainty about the syntactic structure of the sentence,as estimated from a corpus-based probabilistic grammar, can cause a slowdown in reading. An electromagnetic (MEG) study of single word recognition found that the processing of words with higher syntactic uncertainty was associated with decreased neural activity in anterior temporal regions. Finally, an MEG experiment is reported that applied the framework proposed in the preceding chapters to the recognition of morphologically complex spoken words such as builder (build+er).
The second part takes up a central question in language learning: Can general regularities be acquired before specific ones, or do learners only posit a generalization once they have learned several specific instances of it (conservative generalization)? In nonprobabilistic models, conservative generalization can avert commitment to overly broad generalizations that are difficult to retreat from. Three artificial language experiments on the learning of phonotactics (regularities in how sounds combine to form words) found that learners were not conservative: they generalized from a single item and were able to learn general rules without first learning their specific instances. A Bayesian model of phonotactic learning is proposed to account for these results. It considers both specific and general hypotheses simultaneously, but incorporates a parsimony bias: when less is known about the language, a single general pattern is acquired; with additional data, the bias is overcome and multiple specific patterns are learned. The model illustrates that conservative generalization is unnecessary in a probabilistic learner that maintains uncertainty about the rules of the language.