Theories of why humans forget have been challenged by the newly discovered list-length/output-interference paradox, in which--under certain testing conditions--learning is not harmed by the amount of verbal material studied, whereas retrieval of that material becomes more difficult with increases in the number of items tested. The latter finding is known as output interference, and the results of the experiment reported here indicate that a release from output interference is obtained when the nature of the items is changed during testing. Specifically, when participants are asked to recognize items from two categories, output interference is minimized when items from each category are tested separately in large blocks. This finding supports models of forgetting that assume interference arises from information about the to-be-learned material that is stored in memory; in contrast, this finding is difficult to explain using models that assume forgetting is the result only of changing context.

Simultaneous Cross-situational Learning of Category and Object Names Tarun Gangwani, George Kachergis, and Chen Yu {tgangwan, gkacherg, chenyu}@indiana.edu Department of Psychological and Brain Science, and Cognitive Science Program 1101 East 10 th Street Bloomington, IN 47405 USA Abstract Previous research shows that people can acquire an impressive number of word-referent pairs after viewing a series of ambiguous trials by accumulating co- occurrence statistics (e.g., Yu & Smith, 2007). The present study extends the cross-situational word learning paradigm, which has primarily been used to investigate the acquisition of 1-to-1 word-referent mappings, and shows that humans can concurrently acquire both 1-to-1 and 1-to-many mappings (i.e., a category relation), even when the many referents of a single word have no unifying perceptual features. Thus, humans demonstrate an impressive ability to simultaneously apprehend hierarchical regularities in their environment . Keywords: statistical learning; cross-situational learning; category learning; mutual exclusivity; language acquisition Introduction In order to make sense of their world, human infants must learn relationships between words and referents in their environment. Infants simultaneously come into contact with many diverse, novel objects and equally diverse words that name them. Thus, there is much potential for acquiring erroneous word-referent mappings, given only a single situation. Despite this, both infants and adults have a remarkable ability to learn many novel word-referent associations quickly and accurately. Cross-situational word learning (CSWL) studies give us insight into how people are capable of learning multiple word-referent associations from individually ambiguous situations. Previous CSWL studies have shown that both infants and adults are able to learn simple 1-word to 1-referent mappings with astonishing speed (Kachergis, Yu, & Shiffrin, 2009; Smith & Yu, 2008; Klein, Yu, & Shiffrin 2008; Yu & Smith, 2007). In adult studies, participants are typically instructed to learn which words go with which referents, and are then presented with a few consistently co-occurring objects and spoken pseudowords on each of a series of training trials. On every trial, each pseudoword corresponds to a particular on-screen object, but the intended referent is never indicated. In a typical cross-situational training block, participants attempt to learn 18 word-referent pairs from 27 twelve-second trials consisting of four spoken words and four displayed objects (i.e. a 4x4 design). On average, participants in this condition managed to learn half of the 18 pairs by relying on cross- situational statistics (Yu & Smith, 2007). Further studies have shown that human learning often reflects statistics manipulated during training such as pair frequency, contextual diversity (the diversity of other pairs each pair appears with over time), and within-trial ambiguity (the number of co-occurring words and referents per trial) (Kachergis, et al., 2009). However, simple 1-to-1 mappings are only a subset of the types of word-referent relations that exist in natural languages. 1-to-many mappings include referents that have one common label shared among them, such as a category or concept label. For example, both an apple and a banana may be labeled ‘fruit.’ Learners must learn to map both the superordinate label (‘fruit’) and each basic level name (‘banana’ ‘apple’) to the appropriate referent. Even in a learning paradigm like the 4x4 cross-situational learning condition discussed above, which is simpler than the real world, it is difficult to imagine that learners consider all 16 possible pairings, as might be necessary to learn higher- order relations. Constraints such as mutual exclusivity (ME) can drastically reduce the complexity of such ambiguous situations by limiting the possible pairings to a single word for each object (and vice-versa). Consider Markman and Wachtel’s study (1988), in which a child was placed in front of a learned object (ball) and an unlearned object (gyroscope) and was prompted to retrieve the ‘toma.’ While ‘toma’ could be another name for the ball, the child moves to the unlearned object, exhibiting ME. However, despite its power to speed learning, the strict use of ME as a constraint in cross-situational learning would also make it impossible to learn non-1-to-1 mappings. To determine whether learners use the ME constraint when learning names for previously unknown objects, Yurovsky and Yu (2008) presented learners with ME- violating mappings in the CSWL paradigm. An ME- violating mapping is a word (or object) that is consistently paired with more than one object (or word). Participants were trained on 12 words and 18 referents, where 6 words were paired with 12 referents (i.e., 2 referents per word), known as double words, and the other 6 words were paired 1-to-1 with the remaining 6 referents, known as single words. Participants had to decide how to manage two names that co-occur with one referent in the same set of trials. The results showed that participants had equal performance in learning both single and double words if each double word’s two referents were interleaved rather than temporally separated (i.e. one referent was shown in the first half only

Previous research shows that people can acquire an impressive number of word-referent pairs after viewing a series of ambiguous trials by accumulating co-occurrence statistics (e.g., Yu & Smith, 2007). The present study extends the cross-situational word learning paradigm, which has primarily been used to investigate the acquisition of 1-to-1 word-referent mappings, and shows that humans can concurrently acquire both 1-to-1 and 1-to-many mappings (i.e., a category relation), even when the many referents of a single word have no unifying perceptual features. Thus, humans demonstrate an impressive ability to simultaneously apprehend hierarchical regularities in their environment.

Previous research shows that people can acquire an impressive number of word-referent pairs after viewing a series of ambiguous trials by accumulating co-occurrence statistics (e.g., Yu & Smith, 2006). The present study extends the cross-situational word learning paradigm, which has previously dealt only with noun acquisition, and shows that humans can concurrently acquire nouns and adjectives (i.e., a natural category with a distinctive, unifying feature). Furthermore, participants are able to learn ad hoc categories of referents consistently cooccurring with a label, while simultaneously learning instance labels. Thus, humans demonstrate an impressive ability to simultaneously apprehend regularities at multiple levels in their environment.