possible coincidence between phonetically natural, phonologically natural and phonologically active classes, the nature of language change prohibits that that goal will never be reached (see again Hyman, 1975). The fact that the infants in the Saffran and Thiessen (2003) study were unable to induce generalisa-tions based on random groups of segments indicates that contrary to Mielke’s predictions, ‘crazy classes’ are difficult to learn.
gradient phonetics into the grammar (Contra DT and PBPP), and within what can be presumed to be the limits of linguistic competence (contra DT). Whereas the shape of the inventory is the result of phonetic enhancement of contrast, the structure of the inventory, the contrasting elements in relation to each other, is derived by means of the Successive Division Algorithm (Hall, 2007), an extension of the Modified Contrastive Hierarchy model (Dresher, 2009). The contrast and enhancement theory of dispersedness does not come without its own issues, though; first of all, it is unclear what the ontological status of the Contrastive Hierarchy is. By relying so much on the system-level property of contrast, it appears that MCH is or needs to be holistic in its attitude towards the segment inventory. This is also expressed in the learning algorithm: it only works for a highly idealised one-pass-does-all type of learner.
In the last section, we discussed the innateness of features, based on a definition of innateness proposed by Elizabeth Spelke. We then proceeded to investigate whether distinctive features reach the criterium set by this defini-tion, differentiating between features in their contrastive/categorising, lexical and phonotactic uses. From each of these three perspectives, the concept of innate features cannot be discarded and often, the evidence supports it.
In the present thesis, we propose a model of the segment inventory in which many of these issues are avoided. First of all, the ontological status of the inven-tory is clear: it has none. Or, to put it more nuanced, it is entirely epiphenome-nal. The means by which it is derived, furthermore, are entirely implementable within existing theories of phonology; in fact, in one of the dominant models of the phonological inventory (Optimality Theory), feature co-occurrence con-straints have been dormantly present since its inception. Hence, learning is not thought to be categorically different from learning of other language-specific properties of phonology. The theory will be explored in detail in chapter 3, and a demonstration of how it functions in acquisition is provided in chapter 4.
The Final State: Feature Co-occurrence Constraints in the adult grammar
3.1 Introduction
The general claim of this thesis is that the acquisition of the segment inven-tory can be accurately described with a system of monovalent features and co-occurrence restrictions on those features. It aims to provide insights in phono-logical acquisition, in feature theory and – to some extent – constraint theory.
For any theory of phonological acquisition, it is imperative that it is able to de-scribe the end point of the acquisition process: the adult grammar, or, in more technical terms, the Final State. For our present purposes, the Final State is the segment inventory of Dutch, and in this chapter, we will see how the Fea-ture Co-occurrence Constraint theory accurately and adequately describes the Dutch segment inventory, and what we can learn from that concerning the two major ingredients of the theory: features and constraints.
The theory developed here takes as its central apparatus the Feature Co-occurrence constraint. An FCC is structurally very simple: it refers to no more and no less then two features, and comes in either of two types: c-constraints ban a combination of the two features within one and the same segment, whereas i-constraints demand a feature to be present, given the presence of another feature. Definitions of the two are given below, in standard OT fashion for convenience:
(24) a. *[F, G]
assign a violation mark for every segment Σ iff [F] is dominated by
Σ and [G] is dominated by Σ (c-constraint ) b. [F→G]
assign a violation mark for every segment Σ iff [F] is dominated by Σ and [G] is not in Σ(i-constraint )
Feature Co-occurrence Constraints in this thesis are defined as segment-based, even though the segment is seen as no more than the coincidence of features under a root node. As we will see later, constraints can be seen as functions that take as their input a phonological representation and output a number:
0 or 1 in non-OT type frameworks (that is to say, violated or not violated), and an integer equal to the number of violations in OT. There are different metrics to decide the outcome of such a constraint-function. In this thesis, violations are assigned to segments that contain an illicit structure. This is in line with the proposal in McCarthy (2004). The alternative would be to assign violation marks for every feature that co-occurs with an illegal partner. The practical difference between these metrics is negligible for our present purposes;
it is manifested mostly in cases where the autosegmental behaviour of features and segments come into play. In the current thesis, we are not concerned with spreading, deletion and epenthesis. Hence, we will assume the segment-based definition as given above.
In this chapter, we will explore the formal and linguistic properties of FCCs and the implications thereof for feature theory, beginning with an illustration of an FCC account of the segment inventory of Dutch. In the next chapter, we will see how Feature Co-occurrence Constraints provide insights in phonological acquisition.
One important aspect of the FCC-approach is that is entails a division of labour in defining the shape and content of the inventory between representa-tional and derivarepresenta-tional (grammatical) devices. Both are obligatory ingredients in any phonological theory (Scheer, 2011). The representational side of FCC theory is formalised as the set of (possible) segments, the derivational side is the set of FCCs and their filtering effects. As in OT, this division takes the following shape: the representational device proposes, the derivational device disposes.
Many authors have proposed – or tacitly assumed – that the members of the segment inventory are determined by a set of feature co-occurrence constraints (Hayes, 1999; Kager, 1999). Whether, for the current proposal, Optimality The-ory is the most suitable framework to be adopted will be discussed in chapter 5. In other frameworks too, the ungrammaticality of certain feature combina-tions has been formalised by means of feature co-occurrence constraints. In principle, this holds for every theory that adopts a generator-and-filter model of phonology; a model where possible structures are generated by one module, and filtered by another (constraints). The generator is most explicitly defined in Optimality Theory, where we know it as GEN. In many other models, the generator function is performed by rules, where constraints may be active to filter unwanted rule output. The type of generator we assume in the current model is much like GEN, in that once a feature is acquired, it can in principle
combine with any other feature. Whether co-occurrence restrictions hold only at the surface level or also underlyingly is a question that pertains directly to the framework in which one chooses to implement the theory. We will come back to this issue in chapter 5.
For the current discussion, suffice it to say that constraints on feature com-binations are an effective way to delimit the segment inventory in a number of frameworks, but that a precise discussion of their shape is often lacking (no-table exceptions being Hayes & Wilson, 2008; Hayes, 1999), and that, prior to Levelt and van Oostendorp (2007), the possibility of using FCCs in a theory of acquisition was never investigated.
3.1.1 Overview of the chapter
In section 3.2 we will provide a descriptive picture of the Dutch segment in-ventory. Here, we will be introduced to the feature set that was arrived at to complete the analysis in terms of Feature Co-occurrence Constraints presented in section 3.2.3. The assumptions and consequences involved in defining the constraints are motivated in section 3.3. Next, section 3.4 explores the assump-tions and consequences of the chosen feature set.
Before we continue, however, it is important to note that in a theory which characterises acquisition as the simultaneous development of representational and derivational devices, the characteristics of these devices are interlocked.
That is to say, the feature set is not trivial. An important benefit that comes from approaching a single subject (the segment inventory) from both the repre-sentational (features) side and the derivational (constraints) side, is that we can learn something not just about both, but also how choices on one side have con-sequences for the other. As an example (this will be discussed in much greater detail below), major class features and feature geometric grouping nodes are not easily combined with Feature Co-occurrence Constraints.