Word learning in the first year of life

In the first part of this thesis, we ask whether 4-month-old infants can represent objects and movements after a short exposure in such a way that they recognize either a repeated object or a repeated movement when they are presented simultaneously with a new object or a new movement. If they do, we ask whether the way they observe the visual input is modified when auditory input is presented. We investigate whether infants react to the familiarization labels and to novel labels in the same manner. If the labels as well as the referents are matched for saliency, any difference should be due to processes that are not limited to sensorial perception. We hypothesize that infants will, if they map words to the objects or movements, change their looking behavior whenever they hear a familiar label, a novel label, or no label at all. In the second part of this thesis, we assess the problem of word learning from a different perspective. If infants reason about possible label-referent pairs and are able to make inferences about novel pairs, are the same processes involved in all intermodal learning? We compared the task of learning to associate auditory regularities to visual stimuli (reinforcers), and the word-learning task. We hypothesized that even if infants succeed in learning more than one label during one single event, learning the intermodal connection between auditory and visual regularities might present a more demanding task for them. The third part of this thesis addresses the role of associative learning in word learning. In the last decades, it was repeatedly suggested that co-occurrence probabilities can play an important role in word segmentation. However, the vast majority of studies test infants with artificial streams that do not resemble a natural input: most studies use words of equal length and with unambiguous syllable sequences within word, where the only point of variability is at the word boundaries (Aslin et al., 1998; Saffran, Johnson, Aslin, & Newport, 1999; Saffran et al., 1996; Thiessen et al., 2005; Thiessen & Saffran, 2003). Even if the input is modified to resemble the natural input more faithfully, the words with which infants are tested are always unambiguous – within words, each syllable predicts its adjacent syllable with the probability of 1.0 (Pelucchi, Hay, & Saffran, 2009; Thiessen et al., 2005). We therefore tested 6-month-old infants with such statistically ambiguous words. Before doing that, we also verified on a large sample of languages whether statistical information in the natural input, where the majority of the words are statistically ambiguous, is indeed useful for segmenting words. Our motivation was partly due to the fact that studies that modeled the segmentation process with a natural language input often yielded ambivalent results about the usefulness of such computation (Batchelder, 2002; Gambell & Yang, 2006; Swingley, 2005). We conclude this introduction with a small remark about the term word. It will be used throughout this thesis without questioning its descriptive value: the common-sense meaning of the term word is unambiguous enough, since all people know what are we referring to when we say or think of the term word. However, the term word is not unambiguous at all (Di Sciullo & Williams, 1987). To mention only some of the classical examples: (1) Do jump and jumped, or go and went, count as one word or as two? This example might seem all too trivial, especially in languages with weak overt morphology as English, but in some languages, each basic form of the word has tens of inflected variables. (2) A similar question arises with all the words that are morphological derivations of other words, such as evict and eviction, examine and reexamine, unhappy and happily, and so on. (3) And finally, each language contains many phrases and idioms: Does air conditioner and give up count as one word, or two? Statistical word segmentation studies in general neglect the issue of the definition of words, assuming that phrases and idioms have strong internal statistics and will therefore be selected as one word (Cutler, 2012). But because compounds or phrases are usually composed of smaller meaningful chunks, it is unclear how would infants extracts these smaller units of speech if they were using predominantly statistical information. We will address the problem of over-segmentations shortly in the third part of the thesis.