Implementing and Testing Theories of Linguistic Constituency I: English Syllable Structure

This paper proposes and tests an experimental method to evaluate models of linguistic constituency, including: 1) Connections within constituents are stronger than connections spanning constituent boundaries, 2) A constituent is more likely to be parsed out of the signal than a non-constituent (i.e., constituents are processing units), and 3) Both constituents and non-constituents are units, with constituents simply having higher frequency than non-constituents. The method, XOR learning, is designed to distinguish between associability of a whole and associability of its parts. Subjects learn to associate the whole with a different response than the response both of its parts have been associated with. We apply the method to the onset-rime organization of English CVC syllables with a lax vowel, showing that native English speakers can learn rime-affix associations but not body-affix associations. This difference in associability between bodies and rimes is observed in the absence of associability differences between onsets and codas, the only parts that bodies and rimes do not share. Competing theories of linguistic constituency are implemented in a Hebbian framework where parts of the syllable extracted from the signal become associated with affixes they co-occur with. Assuming automatic phonemic categorization, experimental results are explained only by models that assume that rimes and bodies differ in the level of activation they have during training. Applications of the experimental method and its variants to linguistic constituency in other domains are discussed. Introduction Theories of Constituency This paper introduces a method to distinguish between different theories of linguistic constituency. Thus, the question we would like to address is: what are constituents? What does it mean to say that in an English syllable consisting of an onset, a nucleus, and a coda, the nucleus forms a constituent with the coda and not with the onset? The traditional answer to this question in linguistic theory has been that the rime (nucleus+coda) is allocated a node in the tree structure while the body (onset+nucleus) is not (e.g., Fudge, 1987; Selkirk, 1982). A tree structure is a type of a network and, like in any network, it consists of nodes connected by links. By definition, then, a node is something that can be connected to/associated with something else. Thus, in the traditional view of linguistic constituency, constituents can be associated with other units, i.e., constituents are associable, while non-constituents are not. Thus, under this view, if the rime is a constituent while the body is not, rime-affix associations should be learnable while body-affix associations should not be. In order to associate a unit X with another unit Y, the two units must be extracted from the signal. Thus, things that are associable must be extracted from the signal. In other words, something that is associable must be a processing unit. Under the traditional view of constituency, then, constituents are processing units (cf. Cutler et al., 2001; Mehler, 1981). That is, at the very least, if the rime is a constituent and the body is not, the rime should be more likely to be extracted from the acoustic signal than the body is. An alternative to the tree-structural view of constituency is the dependency-based view, applied to syllabic constituency by Vennemann (1988) and Anderson and colleagues (e.g., Anderson & Ewen, 1987). Under this view, neither constituents not non-constituents are allocated nodes. Rather, connections between parts of a constituent are stronger than connections that cross constituent boundaries. CONSTITUENCY AND ASSOCIABILITY 243 Under this view, to say that the rime is a constituent while the body is not means to say that the nucleus is connected to the coda more strongly than to the onset. The dependency-based view does not straightforwardly predict a difference in associability between constituents and non-constituents. Any such difference would be an epiphenomenon, deriving from differences in associability between parts that the constituent and the non-constituent do not share. Thus, in the case of the body, the rime would be expected to be more associable than the body if and only if the coda is more associable than the onset. Finally, processing units may differ in how associable they are, depending on factors like frequency and the cumulative strength of associations they already have (e.g., Kamin, 1969; Moder, 1992). That is, nodes may differ in associability. The associative learning literature indicates that frequent stimuli are harder to associate than infrequent stimuli (the phenomenon known as preexposure, or desensitization effects, see Hall, 2003, for review). In the linguistic literature, Bybee and Brewer (1980) and Moder (1992) have argued that frequent words have weaker connections to similar words than infrequent ones. Thus, if a rime is more frequent than a body, the rime may be expected to be less associable than the body. Given this potential influence on associability, rimes and bodies may be equally likely to be parsed out of the signal (i.e., constituents and non-constituents may be equally salient) and still differ in associability. Equal salience is proposed within full-listing models in which all possible segment strings (up to a particular length) are parsed out of the signal (Skousen, 1989; see also Bod 1998 for an analogous approach to syntax in which all possible subtrees are parsed out). In this paper, we will compare associability of rimes to associability of bodies and then implement the various theories of syllable structure in a common framework to see which can account for the experimental data.