Statistical language learning

Theoretical arguments based on the "poverty of the stimulus" have denied a priori the possibility that abstract linguistic representations can be learned inductively from exposure to the environment, given that the linguistic input available to the child is both underdetermined and degenerate. I reassess such learnability arguments by exploring a) the type and amount of statistical information implicitly available in the input in the form of distributional and phonological cues; b) psychologically plausible inductive mechanisms for constraining the search space; c) the nature of linguistic representations, algebraic or statistical. To do so I use three methodologies: experimental procedures, linguistic analyses based on large corpora of naturally occurring speech and text, and computational models implemented in computer simulations. In Chapters 1,2, and 5, I argue that long-distance structural dependencies - traditionally hard to explain with simple distributional analyses based on ngram statistics - can indeed be learned associatively provided the amount of intervening material is highly variable or invariant (the Variability effect). In Chapter 3, I show that simple associative mechanisms instantiated in Simple Recurrent Networks can replicate the experimental findings under the same conditions of variability. Chapter 4 presents successes and limits of such results across perceptual modalities (visual vs. auditory) and perceptual presentation (temporal vs. sequential), as well as the impact of long and short training procedures. In Chapter 5, I show that generalisation to abstract categories from stimuli framed in non-adjacent dependencies is also modulated by the Variability effect. In Chapter 6, I show that the putative separation of algebraic and statistical styles of computation based on successful speech segmentation versus unsuccessful generalisation experiments (as published in a recent Science paper) is premature and is the effect of a preference for phonological properties of the input. In chapter 7 computer simulations of learning irregular constructions suggest that it is possible to learn from positive evidence alone, despite Gold's celebrated arguments on the unlearnability of natural languages. Evolutionary simulations in Chapter 8 show that irregularities in natural languages can emerge from full regularity and remain stable across generations of simulated agents. In Chapter 9 I conclude that the brain may endowed with a powerful statistical device for detecting structure, generalising, segmenting speech, and recovering from overgeneralisations. The experimental and computational evidence gathered here suggests that statistical language learning is more powerful than heretofore acknowledged by the current literature