Exploring language learning as uncertainty reduction using artificial language learning
Maša Vujović · UCL Discovery (University College London) · 2020
How do children learn language in a way that allows generalization -- producing and comprehending utterances that have never been heard before? A prominent view is that this is achieved through learning the statistical distributions of linguistic forms in the input. There is extensive evidence that humans are sensitive to the statistics of the input, but the exact nature of the learning mechanism that underpins it is unclear. In this thesis, a discriminative approach to learning is taken, whereby language learning is a process of reducing uncertainty about the form and the meaning of the message by discriminating between informative and uninformative cues in the environment and in the utterance itself. This process is driven by key principles of learning theory -- prediction error and cue competition, which are available to differing degrees in different learning contexts. Specifically, (1) learning suffixes provides greater cue competition than prefixing, which facilitates generalization via discriminative learning; (2) learning pre fixes, on the other hand, facilitates the processing of upcoming parts of the utterance, because the pre fix smoothes information content over the whole utterance, which promotes better learning of the utterance itself compared to suffixing. This thesis tests these predictions in a series of artifi cial language learning experiments (with adult native speakers of English), which are either "suffixing" or "prefixing". Support for (1) was found across multiple experiments, but no consistent evidence for (2) was found. On the whole, the thesis demonstrates that discriminative learning provides a coherent theoretical basis for testing specific predictions about language, and identifies avenues for future work to address (2) more appropriately.