Acquiring New Speech Sounds by Clustering

Acquiring a second language phonetic contrast involves learning new sound categories from examples. Consequently, research in category learning should have important implications for research examining how a foreign sound contrast is acquired. The current project is an initial attempt to apply work in category learning to work in second language sound acquisition. Here, I advance that a clustering account developed in category learning research can account for not only various empirical phenomena in the category learning literature but also some critical aspects of acquisition of English /l/– /r/ distinction by adults who are native speakers of Japanese surprisingly well. The clustering account suggests new training methods that should aid adults’ acquisition of new speech sounds. In learning to perceive and produce a second language phonetic contrast, learners must acquire new sound categories from examples. Thus, acquiring a foreign sound contrast is an act of category learning. Then, one would expect that advances in our understanding of category learning would play a critical role in understanding how humans acquire new speech categories. Unfortunately, there has not been a strong link between research in category learning and research in foreign speech sound acquisition. The current project is an initial attempt to make an explicit connection between these two fields by applying what is known in one field to the other. Here, I apply a clustering account developed in category learning research to various aspects of acquisition of English /l/–/r/ distinction by adults who are native speakers of Japanese. Category Learning by Clustering Research in category learning seeks to better understand how humans encode, organize, and use knowledge. The clustering account of category learning (e.g., Love, Medin, & Gureckis, 2004) posits that humans represent categories by one or more clusters. The clustering model starts with one cluster centered upon the first item encountered, and an item is assigned to the cluster it is most similar to (e.g., closest in multi-dimensional space). A critical mechanism of the clustering model is its formation of new clusters in response to surprising events or expectation violation (cf. Rescorla & Wagner, 1972). Surprising events can take two forms (e.g., Gureckis & Love, 2003): (a) when people predict an incorrect category (e.g., a child predicts that a bat belongs to the bird cluster, makes a mistake, and develops a new cluster to encode the bat) and (b) when people encounter an item that is not similar to any existing cluster (e.g., one might create a new cluster for penguins, which look quite different from typical birds). Both of these two forms of surprising events may operate in supervised learning, in which corrective feedback is provided (e.g., you are told bats are mammals and not birds). In contrast, in unsupervised learning, in which there is no corrective feedback, only encountering an item that is highly dissimilar to any existing cluster results in a surprising event and the development of a new cluster (Sakamoto & Love, 2004). Using this mechanism, the clustering model can account for various empirical phenomena (Love et al., 2004). A strong support for the cluster recruitment mechanism is that people better remember items that deviate their expectation (e.g., bats) than items that follow their expectation (von Restorff, 1933). This robust memory phenomenon has been established in various forms under both supervised and unsupervised learning. For example, given a list of items to memorize, people show a memory advantage for an item that differs from others in some way, such as an American city (Austin) in a list of Canadian cities (Vancouver, Toronto, Montréal). Likewise, deviant faces, behaviors, and category members result in enhanced memory (Sakamoto & Love, in press). In fact, the memory advantage of deviant items is problematic for category learning models that do not accord special status to surprising events (e.g., Sakamoto, Matsuka, & Love, 2004). For example, exemplar models (e.g., Medin & Schaffer, 1978) store individual items without considering what other items are already stored in memory. Prototype models (e.g., Posner & Keele, 1968) average all members of the same category, including items that deviate one’s expectation. These models are insensitive to expectation violation and cannot predict enhanced memory for deviant items. In contrast, the clustering model can predict enhanced memory for deviant items by averaging expectation-following items into the same clusters but storing deviant items in their own clusters (Sakamoto & Love, 2004). Acquiring New Speech Categories According to the clustering account, learners’ ability to acquire a phonetic contrast of a second language is determined by the existing clusters encoding the sounds of their native language. For example, it will be difficult to form new sound clusters if a cluster already exists that encodes a native sound that is highly similar to the newly experienced sounds. Consistent with this idea, a dominant explanation for the difficulty distinguishing between the English /l/ and /r/ by adults who are native speakers of Japanese (e.g., Logan, Lively, & Pisoni, 1991) is that Japanese speakers map both English /l/ and /r/ to the Japanese flap (e.g., Guion, Flege, Akahane-Yamada,