Weak representational bias and the discovery of linguistic categories from speech waveforms

How does a child start to learn the sound patterns of her native language? There are two diametrically opposed approaches to this problem. According to one approach, the learner always looks for a symbolic representation in input, and such representation is based on a set of universal phonetic features (Chomsky and Halle, 1968). Since phonetic segments – or more intuitively “speech sounds” – are treated as bundles of these features, the universal feature-based representation also implicitly assumes a temporally discrete, segmental input to the learner. In this paper, this perspective will be referred to as strong representational bias. In contrast, the connectionist approach (Elman and Zipser, 1989) claims that the basis of phonological acquisition is the ability to extract statistical information from speech signals. At the level of categories, networks have been trained to simulate the perceptual change experienced by infants after exposure to input distributions (Guenther and Gjaja, 1996). Some believe that phonetic segments are an “emergent” property (Plaut and Kello, 1999) in the learning system, and that universal features offer little insight regarding how phonological acquisition can proceed from acoustic signals. Hence, connectionism illustrates a class of approaches that consider no representation bias as necessary for the learner to accomplish the task. This paper takes an intermediate standpoint between the two views, which we have named weak representational bias. Intuitively, the current approach treats phonological acquisition as a problem of “breaking the speech code”, where the input is raw speech signals, and the “code” is made of abstract representations similar to those of adult language. The emphasis on the “code”, or a highly constrained hypothesis space, is shared by the weak and the strong