Contrast, minimal pairs and feature specification

Stage Features

1 [labial], [voice], [dorsal], [nasal], [liquid]

2 [labial], [voice], [dorsal], [nasal], [liquid], [continuant], [approximant]

3 [labial], [voice], [dorsal], [nasal], [liquid], [continuant], [approximant]

4 [labial], [voice], [dorsal], [nasal], [liquid], [continuant], [approximant]

5 [labial], [voice], [dorsal], [nasal], [liquid], [continuant], [approximant]

6 [labial], [voice], [dorsal], [nasal], [liquid], [continuant], [approximant]

7 [labial], [voice], [dorsal], [nasal], [liquid], [continuant], [approximant]

8 [labial], [voice], [dorsal], [nasal], [liquid], [continuant], [approximant]

9 [labial], [voice], [dorsal], [nasal], [liquid], [continuant], [approximant]

10 [labial], [voice], [dorsal], [nasal], [liquid], [continuant], [approximant]

Table 2.2: Development of features in word-initial onsets in Catootjes actual productions

to the subject. It is both descriptively and explanatorily adequate, and at the conceptual level, comes with the boon of cross-modular uniformity: vowel harmony is parallel to Agree in syntax, and the search algorithm works the same in both modules. In other words, syntax and phonology are, to a degree, procedurally equivalent. Nevins shows that the model is able to account for a wide variety of vowel harmony cases – as long as we know which vowels are needy, and what they are needy for. In Nevins (2010) proposal, segments are defined as bundles of binary features, and all vowels contain the same features (and thus differ only in the values of these features). Any vowel lacking a feature (or a value for a feature) is thus needy.

As mentioned above, the Search algorithm is parametrised for a number of dimensions. One of those dimensions, contrastiveness, is an important param-eter in dparam-etermining what counts as a potential source, for example in trans-parency in vowel harmony. Taking Finnish as an example, Nevins adapts the Search algorithm in such a way that it only considers contrastive occurrences of [±back] as potential sources. In other words, locality is relativised to con-trastiveness.

This is not a trivial step, because it implies that the Search algorithm knows or is able to know when and where a feature is contrastive. The point is that contrastiveness must be marked somehow. Contrastiveness is not a property of features, but of segments: two segments contrast if they differ in the spec-ification of at least one feature. Going back to the example of Finnish, the Search algorithm must know for each instance of [-back] it encounters whether it is contrastive, or not. The same feature (value) can be both contrastive and non-contrastive in the same language; contrastiveness becomes a property of the feature.

A possible solution to this is that the Search algorithm evaluates the entire inventory in each search pass, or that features or feature bundles are some-how marked for contrastiveness. The latter option would add a new element (sub-features? non-phonological features?) to the phonological alphabet (as per minimalist tradition), the former dramatically increases the computational load of the Search algorithm.

Underspecification

In order to get around the problem of marking contrastiveness, feature theories have often made use of underspecification. By specifying only those features (or values) that are contrastive, no reference need to be made to contrastive-ness as independent part of the phonological grammar: contrast has become epiphenomenal, emerging from specification (but see below on Dresher, 2009’s critique of practices of underspecification). Contrastive underspecification is not an option for Nevins, however, as the formal mechanism of underspecification is already used to denote neediness. To underspecify neither needy features, nor redundant features would lead to hyper-harmony: all segments would be needy for whatever feature values they are not contrastive for, initiating the Search

algorithm. In conclusion, in the Needy Vowel approach to vowel harmony, dia-critics are unavoidable to mark either contrastiveness, or neediness. It appears that any theory of full specification needs to either not refer to contrastiveness as a functional element of grammar, or mark it diacritically.

Some elements in phonology display a behaviour that is different from what one would expect on the basis of their surface form. For example, in a process like vowel harmony (see above), some vowels are ‘transparent’. Although at the surface they appear to form a natural class with the triggering vowels (or their complement), they have no effect on the harmony process whatsoever. Consider Finnish, which has a [±back] harmonising requirement:

(17) Transparent vowels in Finnish [±back] harmony (taken from Hall (2011, example 6))

a. [grøtsi+næ] ‘porridge + essive’

[tsaari+na] ‘tsar + essive’

b. [syyte+ttnæ] ‘action + abessive’

[suure+na] ‘entry + abessive’

c. [væitel+lyt ‘dispute + past part’

[ajatel+lut ‘think + past part’

d. [værttinæ+llæ+ni+hæn] ‘with spinning wheel, as you know’

[palttina+lla+ni+han] ‘with linen cloth, as you know’

Situations like these lead to the idea that not all surface features are ‘visible’

to phonology; in fact, that some segments are (lexically) underspecified. The question of what may be left unspecified and what not has been answered in a number of ways. In Radical Underspecification (Kiparsky, 1982; Archangeli, 1984), for all features [±F], only one value, either + or –, can be stored. Ev-ery language has one ‘default’ vowel (for instance, the vowel that shows up in epenthesis), and it is assumed that this vowel is underspecified for all its features, and that these features are filled in in the course of derivation. Be-cause the fill-in rules are there by necessity, it becomes possible to erase from the lexicon all feature specifications that the default vowel will receive even-tually. That is, if the default vowel is [+high], [+high] is underspecified in the lexicon. Other patterns of phonological activity can also be used to motivate which value of a feature can be left unspecified, but the general theme is that somehow predictable feature values are unspecified in the lexicon and filled-in derivationally.

A serious problem for Radical Underspecification is not underspecification, but the point at which a given feature becomes specified. This point, at which value α is assigned to F, is defined as ‘. . . automatically ordered prior to the first rule referring to [αF]’ (Archangeli, 1984, cited in Dresher, 2009), thus de-fying any falsification. In Contrastive Underspecification (Steriade, 1987) the criterion for specification is not only unpredictability but rather contrastive-ness: only feature values that are contrastive in the language are specified. The

problem is that Contrastive Underspecification does not supply a principled way of determining contrasts, and the algorithm that became identified with it (Dresher, 2009, p. 203), the Pairwise Algorithm (or minimal pairs, see above) is inadequate, as we shall see in section 2.1.2. A third possibility to achieve underspecification is simply to identify one feature and label it ‘underspecified’

universally. This is the route taken by proponents of the Featurally Under-specified Lexicon (FUL, see Lahiri & Reetz, 2002, for example; see also section 2.3.2), who claim that [Coronal] is always underspecified, a proposal which we shall adopt in chapters 3 and 4.

In section 1.3.2 above, we have discussed a number of theories that approach the acquisition of the inventory as a logical problem. In other words, they posit an abstract, hypothetical learner to which certain capacities are ascribed, and demonstrate that the model they propose is learnable, or even emerges from a learning mechanism. One of the most prominent of such theories, as we have seen, is known as the ‘Modified Contrastive Hierarchy’ (Dresher, 2009; Hall, 2007, 2011, among others). As the name indicates, the Modified Contrastive Hierarchy (henceforth, MCH) puts a high burden on the contrastive status of features. In fact, the ‘Contrastivist Hypothesis’ holds that only contrastive features can be phonologically active in a language (Hall, 2007). Hence, MCH represents a special instance of underspecification theory. What is more, the MCH represents an alternative to the minimal pair hypothesis, and Dresher (2009) effectively shows how the minimal pair hypothesis faces a learnability problem if it is used to derive underspecification.

Minimal pairs in acquisition

Interestingly, criticism of the minimal pair hypothesis has come from both the perspective of the ‘logical problem of language acquisition’ (Dresher, 2009), and the ‘developmental problem of language acquisition’ (see Ingram (1989) and section 1.3 for an exploration of these terms).

Much of the literature in first language acquisition reports research on a remarkable transition that children go through in their first year. Infants start out being able to discriminate between all possible speech sounds, and gradu-ally lose this ability while improving the recognition of sounds relevant to their own language. Many models have been proposed and/or adapted to interpret this transition (e.g., Kuhl (1991); Best (1995)), but generally, this develop-ment is seen to be due to a ‘reorganisation in perceptual biases’ (Werker &

Pegg, 1992). The process of acquiring native language speech sound categories proceeds from universal yet categorical discrimination at a very young age (Eimas, Siqueland, Jusczyk, & Vigorito, 1971), to an adult-like perception of native vowels (Kuhl, 1991) at around six months, and finally to a manner of discriminating consonants similar to that of adults before the end of the first year (Werker & Tees, 1984).

In the developmental perspective, the Minimal Pair hypothesis assumes a top-down developmental process, in which the child first stores whole word

forms, and only later, as the lexicon is large enough, begins to analyse these forms in smaller units of representation (see, for instance, MacKain (1982);

Jusczyk (1985); Werker and Pegg (1992); Best (1995)). Crucially, the Minimal Pair hypothesis relies on the assumption that the acquisition of a phonology is based on linguistic factors (more specifically, some part of the chain between sound and meaning), not simply on bare statistical generalisations (e.g., Werker

& Pegg, 1992:p. 299).

Maye, Werker, and Gerken (2002) argue that the Minimal Pair hypothesis cannot provide a model of phonetic categorisation, as it depends on the occur-rence of minimal pairs in the lexicon. We will come back to Maye et al. (2002) in more detail below. However, from earlier studies (see, e.g., Eimas et al., 1971;

Werker & Tees, 1984; Kuhl, 1991)⁶, we know that native phonemic categories are formed robustly before a reliably large lexicon has been acquired that is, native phonemic categories are largely in place before the end of the first year of life. What is more problematic, hardly any minimal pairs occur in the child’s receptive lexicon when it contains of 50 words (Maye et al., 2002). For these reasons, Maye et al. (2002) state that the child’s lexicon cannot serve as the basis for phonemic development as it is not large enough to contain the required number of minimal pairs. We may have to relativise their claim somewhat, as it seems that their conception of the lexicon is the lexicon in its final, adult-like state. The acquisition of the adult lexicon has two major precursors: retention of word forms, and formation of concepts (the third ingredient of an adult-like lexicon being syntactic information). These elements are then integrated in the adult lexicon. Thus, the child must have storage of phonological forms before the onset of a meaningful lexicon. That this is so, at least at eight months of age, has been shown in Jusczyk and Hohne (1997). Again, this is not taken to mean that indeed, children use minimal pairs in their (proto-)lexicon to form phonological categories, but it does serve to indicate that the characterisation of the minimal pair hypothesis sketched in Maye et al. (2002) may be overly restrictive (see also Swingley, 2008). Even so, the Minimal Pair hypothesis is a poor candidate as a model of acquisition – it seems to suffer from a (fatally) severe case of poverty of the stimulus.⁷

6Although Werker and Tees (1984) note that “It is probably no accident that this decline, or tuning, occurs at about the age that the child is beginning to understand and possibly produce sounds appropriate to his/her native language.”

7In contrast to the Minimal Pair hypothesis, many studies propose that children have direct access to a great deal of fine-grained statistical information present their linguistic environment, and that they are able to exploit this information in order to construct phono-logical categories. According to these studies, acquiring the features of the native languages proceeds through distributional learning (see, for example, Maye et al. (2002)). However, one problem for a Distributional Learning hypothesis is, as Peperkamp (2003) notes, that distri-butional data is notoriously unreliable in itself: one phoneme may have multiple allophones that overlap in phonetic space, and, thus, one surface phoneme may belong to a number of underlying segments. Furthermore, it is not a priori clear what distributions the child should be sensitive to: type or token distributions, word-level, phoneme-level or feature level? This is a serious problem for distributional learning studies, because a) post-hoc correlations do not carry much meaning, and b), while many distributional learning studies assume a causal

2.1.3 Holisticity, contrast and specification in Feature