A fundamental question in word learning is how, given onlyevidence about what objects a word has previously referred to,children are able to generalize the total class (Smith & Medin,1981; Xu & Tenenbaum, 2007). E.g. how a child ends upknowing that ‘poodle’ only picks out a specific subset of dogsrather than the whole class and vice versa. The Na ̈ıve Gen-eralization Model (NGM) presented in this paper offers an ex-planation of word learning phenomena grounded in categoryformation (Smith & Medin, 1981) The NGM captures a rangeof relevant experimental findings (Xu & Tenenbaum, 2007;Spencer, Perone, Smith, & Samuelson, 2011), including thosewhich are in conflict with a Bayesian inference theory (Xu &Tenenbaum, 2007).

Vowel harmony denotes a class of phonotactic constraintswhich limit which vowels can co-occur in words. The charac-teristics of harmony systems have been well-researched fromtheoretical, typological, and developmental perspectives. Chil-dren are sensitive to harmony very early in their development,as young as seven months, so the mechanisms responsible forharmony acquisition must be able to identify its presence aswell as the specifics of individual vowel harmony systemswith little input. Prior computational work has sought eitherto detect the presence of harmony without describing the spe-cific implementation or to describe a specific implementationwhen the general details are known beforehand. We presenta new computational acquisition approach inspired by phono-logical notions of restrictiveness which succeeds in automat-ically detecting harmony in some language and describes thegross characteristics of the underlying harmony grammar with-out prior knowledge about the type of system to expect.

Recent studies have shown that a majority of word learning input in the natural environment is highly ambiguous, with only a small handful of super-informative instances. This characterization, though, may obscure details of what that “ambiguity” comprises, and how it nonetheless supports learning. We advance this idea by distinguishing a typology of three classes of word learning contexts: highly-informative (H) cases, ambiguous (A) cases which result in many possible hypotheses for the underlying meaning, and misleading (M) cases which generate largely just a single response, but one that is incorrect. We present two experiments on verb learning: first, we used a Human Simulation Paradigm (HSP) to categorize videos extracted from a parent-infant toy-play video corpus into these H/A/M classes. Second, we used the output of the HSP to select stimuli for a new cross-situational verb learning experiment—utilizing different combinations of H/A/M videos to assess their cumulative impact of learning.

Vowel harmony is a class of phonotactic restrictions in which vowels in a language are divided into two or moresubclasses, and words must contain only vowels from only one such subclass regardless of intervening consonants. Languagesworldwide (Turkish, Finnish, Mongolian, Warlpiri, but not English) exhibit vowel harmony. The opacity of such potentiallylong distance alternations poses a challenge for the learner. Nevertheless, infants are sensitive to vowel harmony alternationsat as young as seven months. We present a computational model for vowel harmony acquisition. By normalizing transitionalprobabilities over the vowel tier, and making minimal assumptions about the phonology, we successfully determine which testlanguages have harmony processes and correctly categorize their vowels into harmonizing classes. Using universal typologicalpatterns to inform the search space, we find that phenomena which appear opaque can be captured by simple distributionallearning.

What happens to the acoustic signal after it enters the mind of a listener during real-time speech processing? Sinceprocessing involves extracting linguistic evidence from multiple, temporally distinct sources of information, successfulcommunication relies on a listeners ability to combine these potentially disparate signals. Previous work has shown thatlisteners are able to maintain, and rationally update, some type of intermediate representations over time. However, exactlywhat type of information is being maintainedbe it acoustic-phonetic or rather a probability distribution over phonemeshasbeen underspecified. In this paper we present a perception experiment aimed at identifying the internal contents of in-termediate representations in speech processing. Using an accent-adaptation paradigm, we find that listeners adapt tomodulated acoustic signal when the corresponding orthography is provided before the audio, but not when audio followsthe orthography. This supports the position that intermediate representations are uncertainty-distributions over discreteunits (e.g. phonemes) and that, by default, speech processing involves no maintenance of the acoustic-phonetic signal.

In order to understand early word learning we need to better understand and quantify properties of the input that young children receive. We extended the human simulation paradigm (HSP) using egocentric videos taken from infant head-mounted cameras. The videos were further annotated with gaze information indicating in-the-moment visual attention from the infant. Our new HSP prompted participants for two types of responses, thus differentiating referential from semantic ambiguity in the learning input. Consistent with findings on visual attention in word learning, we find a strongly bimodal distribution over HSP accuracy. Even in this open-ended task, most videos only lead to a small handful of common responses. What's more, referential ambiguity was the key bottleneck to performance: participants can nearly always recover the exact word that was said if they identify the correct referent. Finally, analysis shows that adult learners relied on particular, multimodal behavioral cues to infer those target referents.