Use this URL to cite or link to this record in EThOS:
Title: Predictability effects in language acquisition
Author: Pate, John Kenton
Awarding Body: University of Edinburgh
Current Institution: University of Edinburgh
Date of Award: 2013
Availability of Full Text:
Access through EThOS:
Full text unavailable from EThOS. Please try the link below.
Access through Institution:
Human language has two fundamental requirements: it must allow competent speakers to exchange messages efficiently, and it must be readily learned by children. Recent work has examined effects of language predictability on language production, with many researchers arguing that so-called “predictability effects” function towards the efficiency requirement. Specifically, recent work has found that talkers tend to reduce linguistic forms that are more probable more heavily. This dissertation proposes the “Predictability Bootstrapping Hypothesis” that predictability effects also make language more learnable. There is a great deal of evidence that the adult grammars have substantial statistical components. Since predictability effects result in heavier reduction for more probable words and hidden structure, they provide infants with direct cues to the statistical components of the grammars they are trying to learn. The corpus studies and computational modeling experiments in this dissertation show that predictability effects could be a substantial source of information to language-learning infants, focusing on the potential utility of phonetic reduction in terms of word duration for syntax acquisition. First, corpora of spontaneous adult-directed and child-directed speech (ADS and CDS, respectively) are compared to verify that predictability effects actually exist in CDS. While revealing some differences, mixed effects regressions on those corpora indicate that predictability effects in CDS are largely similar (in kind and magnitude) to predictability effects in ADS. This result indicates that predictability effects are available to infants, however useful they may be. Second, this dissertation builds probabilistic, unsupervised, and lexicalized models for learning about syntax from words and durational cues. One series of models is based on Hidden Markov Models and learns shallow constituency structure, while the other series is based on the Dependency Model with Valence and learns dependency structure. These models are then used to measure how useful durational cues are for syntax acquisition, and to what extent their utility in this task can be attributed to effects of syntactic predictability on word duration. As part of this investigation, these models are also used to explore the venerable “Prosodic Bootstrapping Hypothesis” that prosodic structure, which is cued in part by word duration, may be useful for syntax acquisition. The empirical evaluations of these models provide evidence that effects of syntactic predictability on word duration are easier to discover and exploit than effects of prosodic structure, and that even gold-standard annotations of prosodic structure provide at most a relatively small improvement in parsing performance over raw word duration. Taken together, this work indicates that predictability effects provide useful information about syntax to infants, showing that the Predictability Bootstrapping Hypothesis for syntax acquisition is computationally plausible and motivating future behavioural investigation. Additionally, as talkers consider the probability of many different aspects of linguistic structure when reducing according to predictability effects, this result also motivates investigation of Predictability Bootstrapping of other aspects of linguistic knowledge.
Supervisor: Goldwater, Sharon; King, Simon Sponsor: Scottish Overseas Research Student Awards Scheme (SORSAS) ; Scottish Informatics and Computer Science Alliance (SICSA)
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available
Keywords: language ; predictability effects ; language-learning ; adult-directed speech ; child-directed speech