Discovering words and rules from speech input: an investigation into early morphosyntactic acquisition mechanisms

To acquire language proficiently, learners have to segment fluent speech into units – that is, words -, and to discover the structural regularities underlying word structure. Yet, these problems are not independent: in varying degrees, all natural languages express syntax as relations between nonadjacent word subparts. This thesis explores how developing infants come to successfully solve both tasks. The experimental work contained in the thesis approaches this issue from two complementary directions: investigating the computational abilities of infants, and assessing the distributional properties of the linguistic input directed to children. To study the nature of the computational mechanisms infants use to segment the speech stream into words, and to discover the structural regularities underlying words, I conducted seventeen artificial grammar studies. Along these experiments, I test the hypothesis that infants may use different mechanisms to learn words and word-internal rules. These mechanisms are supposed to be triggered by different signal properties, and possibly they become available at different stages of development. One mechanism is assumed to compute the distributional properties of the speech input. The other mechanism is hypothesized to be non-statistical in nature, and to project structural regularities without relying on the distributional properties of the speech input. Infants at different ages (namely, 7, 12 and 18 months) are tested in their abilities to detect statistically defined patterns, and to generalize structural regularities appearing inside word-like units. Results show that 18-month-old infants can both extract statistically defined sequences from a continuous stream (Experiment 12), and find internal-word rules only if the familiarization stream is segmented (Experiments 13 and 14). Twelve-month-olds can also segment words from a continuous stream (Experiment 5), but they cannot detect wordstraddling sequences even if they are statistically informative (Experiments 15 and 16). In contrast, they readily generalize word-internal regularities to novel instances after exposure to a segmented stream (Experiments 1-3 and 17), but not after exposure to a continuous stream (Experiment 4). Instead, 7-month-olds do not compute either statistics (Experiments 10 and 11) or within-word relations (Experiments 6 and 7), regardless of input properties. Overall, the results suggest that word segmentation and structural generalization rely on distinct mechanisms, requiring different signal properties to be activated --that is, the presence of segmentation cues is mandatory for the discovery of structural properties, while a continuous stream supports the extraction of statistically occurring patterns. Importantly, the two mechanisms have different developmental trajectories: generalizations became readily available from 12 months, while statistical computations remain rather limited along the first year. To understand how the computational selectivities and the limits of the computational mechanisms match up with the limitations and the properties of natural language, I evaluate the distributional properties of speech directed to children. These analyses aim at assessing with quantitative and qualitative measures whether the input children listen to may offer a reliable basis for the acquisition of morphosyntactic rules. I choose to examine Italian, a language with a rich and complex morphology, evaluating whether the word forms used in speech directed to children would provide sufficient evidence of the morphosyntactic rules of this language. Results show that the speech directed to children is highly systematic and consistent. The most frequently used word forms are also morphologically well-formed words in Italian: thus, frequency information correlates with structural information -- such as the morphological structure of words. While a statistical analysis of the speech input may provide a small set of words occurring with high frequency, how learners come to extract structural properties from them is another problem. In accord with the results of the infant studies, I propose that structural generalizations are projected on a different basis than statistical computations. Overall, the results of both the artificial grammar studies an the corpus analysis are compatible with the hypothesis that the tasks of segmenting words from fluent speech, and that of learning structural regularities underlying word structure rely on statistical and non-statistical cues respectively, placing constraints on computational mechanisms having different nature and selectivities in early development