Integration and Inference: Cross-Situational Word Learning Involves More than Simple Co-occurrences

Integration and Inference: Cross-situational Word Learning Involves More than Simple Co-occurrences Alexa R. Romberg and Chen Yu {aromberg, chenyu}@indiana.edu Department of Psychological and Brain Sciences, 1101 E. 10th Street Bloomington, IN 47405 USA Abstract Statistical word learning involves forming and aggregating associations between words and objects that co-occur across contexts (e.g., Vouloumanos & Werker, 2009; Smith & Yu, 2008; Yu & Smith, 2007). However, the mechanisms that support such learning are currently under debate, including the extent to which learners carry forward multiple ambiguous associations (e.g., Trueswell et al., 2013). The current study presented adults with a set of statistical word learning tasks designed to measure the statistical computations learners employ to build label-object mappings and to probe what information from past contexts is available to further process and integrate with new information. Results reveal that learners use the co-occurrence of label-object pairings to make inferences both about objects and labels currently present and those presented on previous trials. Further, the strength of learners’ memory for past contexts moderated their inferences, suggesting a role for a rich information structure in cross-situational word learning. Keywords: word learning; statistical learning; language acquisition; cross-situational learning Introduction Imagine an infant on a walk with his father. The father, like many parents, comments on the things they see together: “There’s a doggie and a kitty in the window!” and a few moments later: “Look, the man is walking the doggie!”. How might the father’s comments help the infant learn the meanings of words like doggie, kitty and man? Recent research has demonstrated that learners readily form label-object mappings by gathering co-occurrence statistics. Human infants (Smith & Yu, 2008; Vouloumanos & Werker, 2009), children (Scott & Fisher, 2011) and adults (Kachergis, Yu & Shiffrin, 2012; Suanda & Namy, 2012; Yu & Smith, 2007) are all capable of converting multiple individually ambiguous learning instances into specific knowledge as demonstrated by above-chance performance on a post-learning test or by an improvement in selection of the correct referent in a combined training and test procedure (Trueswell, Medina, Hafri & Gleitman, 2013). However, the precise ways in which learners resolve the local ambiguities have been relatively unexplored. Specifying how exactly learners use the information available is an important step to understanding the mechanisms contributing to success. When learners perform some computations but not others, this offers important constraints to any model of their learning and can inform discussion about the nature of the information stored. In the context of cross-situational word learning, two primary mechanisms, associative learning and hypothesis testing, have been proposed for how learners accrue information over time. These mechanisms differ largely in the amount of information stored and, consequently, in how prior information influences later learning (Yu & Smith, 2012). In particular, associative learning proposes that learners form multiple associations between the objects and labels present during each learning instance, storing a relatively rich information network. Hypothesis testing proposes that learners store only a single link between a label and possible referent, discarding other co-occurrence information. Distinguishing these possible mechanisms has been challenging thus far because of a lack of data regarding how learners process information on a trial-by-trial basis. Details about what information learners store and how they use it during cross-situational word learning is vital for advancing theories of this process. The typical cross- situational word learning experiment uses a fairly large novel vocabulary (up to 18 to-be-learned label-object mappings) and consists of a series of trials that each present a subset of the labels and objects. Thus, the learner is faced with the difficult task of tracking these many labels and objects across trials (typically between 27 and 60 trials) and using what co-occurrences they can glean to generate as many correct mappings as possible. While this experimental design is daunting for the participant, it is also daunting for the experimenter, as there are inevitably many possible paths to success. One cannot know definitively how participants arrived at a particular mapping over the course of statistical learning or whether the same types of computations were used for all learned mappings. The present study sought to alleviate these analytical ambiguities for the experimenter while maintaining the learning ambiguities for the participants. Rather than have participants view many trials across which to learn many mappings, learners were presented with a series of “miniature” cross-situational word learning tasks. These tasks consisted of only 2 or 3 trials and were constructed so that some, though not all, label-object mappings could (theoretically) be disambiguated, depending on which information learners stored and which inferences they made. The miniature tasks were constrained so that there was only one pathway to disambiguation, allowing us to infer the computations successful learners employed. We focused on three fundamental processes that could serve as building blocks for sophisticated statistical learning. The first was the tracking of co-occurrence information – noticing that some labels and some objects appear together across multiple trials. The simplest of