The research presented in this PhD dissertation provides a computational perspective on Semantic Implicit Learning (SIL). It puts forward the idea that SIL does not depend on semantic knowledge as classically conceived but upon semantic-like knowledge gained through distributional analysis of massive linguistic input. Using methods borrowed from the machine learning and artificial intelligence literature, we construct computational models, which can simulate the performance observed during behavioural tasks of semantic implicit learning in a human-like way. We link this methodology to the current literature on implicit learning, arguing that this behaviour is a necessary by-product of efficient language processing.

Chapter 1 introduces the computational problem posed by implicit learning in general, and semantic implicit learning, in particular, as well as the computational framework, used to tackle them.

Chapter 2 introduces distributional semantics models as a way to learn semantic-like representations from exposure to linguistic input.

Chapter 3 reports two studies on large datasets of semantic priming which seek to identify the computational model of semantic knowledge that best fits the data under conditions that resemble SIL tasks. We find that a model which acquires semantic-like knowledge gained through distributional analysis of massive linguistic input provides the best fit to the data.

Chapter 4 generalises the results of the previous two studies by looking at the performance of the same models in languages other than English.

Chapter 5 applies the results of the two previous Chapters on eight datasets of semantic implicit learning. Crucially, these datasets use various semantic manipulations and speakers of different L1s enabling us to test the predictions of different models of semantics.

Chapter 6 examines more closely two assumptions which we have taken for granted throughout this thesis. Firstly, we test whether a simpler model based on phonological information can explain the generalisation patterns observed in the tasks. Secondly, we examine whether our definition of the computational problem in Chapter 5 is reasonable.

Chapter 7 summarises and discusses the implications for implicit language learning and computational models of cognition. Furthermore, we offer one more study that seeks to bridge the literature on distributional models of semantics to `deeper' models of semantics by learning semantic relations.

There are two main contributions of this dissertation to the general field of implicit learning research. Firstly, we highlight the superiority of distributional models of semantics in modelling unconscious semantic knowledge. Secondly, we question whether `deep' semantic knowledge is needed to achieve above chance performance in SIIL tasks. We show how a simple model that learns through distributional analysis of the patterns found in the linguistic input can match the behavioural results in different languages. Furthermore, we link these models to more general problems faced in psycholinguistics such as language processing and learning of semantic relations.