A variety of representational architectures are proposed for allomorphy. I distil these into six types (see Figure 3.2), which result from varying two properties: (i) shared versus unshared represent- ations for allomorphs of a morpheme (or both shared and unshared), and (ii) one-level versus two-levels of lexical structure. There is not a one-to-one correspondence between models of mor- phological processing and representational architecture for allomorphy: for example, within a given model of morphological processing, the representational architecture of allomorphy can be dependent on the type of allomorphy: predictable, morpho-phonological or suppletive.
The schematisations in Figure 3.2 represent two separate stimuli which are allomorphs of a morpheme (e.g.,singandsang). time-course is not represented. In (a), there is one representation for two allomorphs of a morpheme which includes both phonological and syntactic/semantic features. In contrast, in (b), there are two representations for two allomorphs of a morpheme. In (c), two phonological representations mediate access from two allomorphs to a single syntactic representation. In (d), two phonological representations mediate access to two separate syntactic representations for two allomorphs. In (e), one phonological representation mediates access to one syntactic representation for two allomorphs. In (f), two allomorphs are represented with separate phonological representations that link to a shared phonological representation (underlying form) which in turn links to one syntactic representation.
For models (b) and (d), the allomorphs of a morpheme do not share a lexical representation corresponding to a morpheme. Morphological relatedness must be handled through correspond- ences outwith the lexical system. For connectionist models, for example, this is handled through separate networks of semantic and phonological relatedness. The model sketched out in Chapter 2.1.7 incorporates hypotheses from morphological theory into a model of processing. It argues 27This is in line with accounts of morphological processing put forth by Rastle and Davis (2008) whereby there is an
early morpho-orthographic stage of segmentation. Phonological overlap seems to be a greater determinant of facilitation in masked priming than overt priming.
for a separation between phonological and syntactic representations. Furthermore, the evidence from previous experimental work which investigates allomorphy indicates that allomorphs share a syntactic representation. As such, we are left with architectures from (c), (e) and (f).
(a) Syntactic Stimulus Phonological (b) Syntactic Stimulus Phonological (c) Syntactic Stimulus Phonological (d) Syntactic Stimulus Phonological (e) Syntactic Stimulus Phonological (f) Syntactic Stimulus Phonological
Figure 3.2: Schematisation of allomorphy architectures
Marslen-Wilson and Zhou (1999) proposed the “direct access model” for processing of morpho- phonological allomorphy, in which allomorphs are linked to a single representation, a “lexical entry”, which specifies syntactic, semantic, and phonological information. This model is repres-
ented in (a). For suppletion (which they term “irregular morpho-phonological alternations”), they state that stored representations of the allomorphic forms would be required, the architecture for which is represented in (b).
Allen and Badecker (2002) argue for two-levels of lexical structure to account for evidence for syntactic morphological effects and competition effects between allomorphs of a morpheme. In (c), two phonological representations mediate access from two allomorphs to a single syntactic repres- entation. This corresponds to the model of Allen and Badecker (2002) (for whom the phonological level is termed theLemmaand the syntactic level is termed theLexeme).
Stockall and Marantz (2006) argue for an architecture of allomorphy which influences the model developed here. Also incorporating hypotheses from “Distributed Morphology”, they argue for an architecture in which both regular and irregular forms activate syntactic representations corres- ponding to the stem at an early stage of processing. Their model differs from the one developed here in that it proposes that the default allomorph as well as any applicable morpho-phonological rules are stored with the syntactic representation. As such, their model also provides an account of processing asymmetries between default allomorphs and derived allomorphs, which I leave aside here. Their model best matches the architecture in (e).
taught
√teach
/titʃ/ *
129[past]
*
13Figure 3.3: “Schematic representation of initial stage of root activation. Processing the past tense form activates the root TEACH and the functional morpheme [PAST]. The *s indicate specific morpho-phonological rules. Rule 129, for example, would generate the irregular taught form in the past tense.” (Stockall and Marantz 2006: 89)
The morphological theory adopted in this dissertation is generative and derivational. As such, it presents some inherent questions for accounting for comprehension in addition to production. A generative theory does not necessarily present a problem: “Analysis by synthesis” (see Bever and Poeppel 2010) holds that both bottom-up and top-down mechanisms are involved in language processing, as do many other accounts. However, a particularly interesting question does arise for morpho-phonological allomorphy. Section 3.2.1 explains that morpho-phonological allomorphy
is analysed as derived from a single phonological underlying form. This raises the question of how the allomorph is comprehended, i.e., how is the derived form recognised as the output of a morpho-phonological rule? One approach to this question, provided by Stockall and Marantz (2006), is that the language user is sensitive to phonological patterns associated with both regular and irregular allomorphy, and therefore form-based decomposition occurs in both cases. This involves storage of the phonological pattern associated with the derived allomorph but not a phonological representation of the allomorph per se. As discussed, a suppletive analysis is also consistent with the distributional evidence, which would hold that phonological representations of the allomorphs are stored. A hybrid account is also possible, in which phonological representations of the allomorph are stored but also associated with an underlying phonological form. Ultimately, the work presented here does not disambiguate between these potential accounts, but they are important considerations for future work concerning the representation of allomorphy. The ar- chitecture in (c) is what “Distributed Morphology” would predict for suppletive allomorphy. The architecture in (e) is what “Distributed Morphology” might predict for morpho-phonological or predictable allomorphy. The architecture presented in (f) is the hybrid alternative discussed.