Refurbishing the structuralist foundations
Although a learning-based perspective is neutral about unit size, it applies most fruitfully to the units that are well discriminated within a system of contrasts. As discussed in earlier chapters, word-sized units appear to provide an optimal combination of discriminability and predictiveness. Both the determinacy of individual analyses and the implicational structure that they define are disrupted by the disassembly of word-sized forms into smaller units. As in WP models in general, the distinguished status of words within a learning-based approach reflects the striking contrast between how well languages facilitate the deduction of whole systems from subsets of forms and how poorly they facilitate the disassembly and reassembly of individual forms.
To a significant extent, this point is acknowledged by nearly all contemporary approaches. What Hockett (1987) termed ‘The Decade of the Morpheme’ closed with the acknowledgement in Hockett (1954) that the ‘problems of morphemic analysis’ articulated in descriptivist studies were not ultimately susceptible to solution in terms that preserved the initial morphemic conception. Morphological models developed within the generative school retained a commitment to ‘morphemes’ (apart from the brief paradigmatic sketch of Chomsky 1965: §4.2.2) but without solving the problems identified by descriptivist accounts. Instead, these problems became moot as a consequence of essentially terminological shifts, as morphemic models came to operate with notions that had at most a historical connection to the morphemes defined in the models of Harris (1942) and Hockett (1947). The break from any morpheme-as-sign conception is complete when models of Distributed Morphology propose that “a morpheme is an abstract syntactic unit” (Marantz 2013:905).
From a learning-based perspective, morphemic models are not just technically deficient, but more fundamentally misconceived. The ultimate motivation for the entire morphemic tradition can be seen to rest on the two pillars of Post- Bloomfieldian approaches to the analysis of language in general. These are the assumption that (i) recurrence entails redundancy and (ii) that structure implies decomposed (morphemic) representations. The goal of morphological analysis in this tradition involves a distillation of variation into general symbolic statements, schemas or rules that describe the distribution and interpretation of isolable units of form. Descriptions that exhibit recurrent patterns are regarded as deficient on the grounds that they ‘miss linguistically significant generalizations’ or “fall short of scientific compactness”, to again repeat the terms that Bloomfield (1933:238) uses in the quotation on p. 70 above. A similar intuition underlies the use of inheritance hierarchies and other strategies for providing compact representations of lexical and grammatical patterns.
The first of these assumptions operates with notions of ‘identity’ and ‘redundancy’ that are defined primarily in terms of orthographic or phonemic transcriptions. Transcriptions may have been regarded as providing sufficient precision in Bloomfield’s time, but their limitations are by now familiar. For a contemporary description that aims to model the variation relevant to human speakers, it is not sufficient that an orthographic or phonemic system fails to distinguish separate occurrences of an element; the item must also be produced and interpreted alike by speakers. The usefulness of phonemic transcriptions for capturing the notion of ‘identity’ relevant for speakers has been challenged by a range of studies that probe sub-phonemic contrasts. Careful acoustic and psychoacoustic investigation of units ranging in size from words to single-segment affixes have shown that speakers consistently produce and comprehend durational differences and other types of phonetic variation that do not determine phonemic contrasts. At the word level, Gahl (2008) found systematic differences in duration between ostensibly homophonous items such as English time and thyme. Drager (2011) reported similar variation for English like in its different functions. At the segment level, Plag et al. (2016) likewise found “significant differences in acoustic duration between some morphemic /s/’s and /z/’s and non-morphemic /s/ and /z/, respectively”.
Morphological units exhibit similar contrasts. In the domain of word formation, Davis et al. (2002) found differences in duration and fundamental frequency between a word like captain and a morphologically unrelated onset word such as cap. Of more direct relevance to the issue of “scientific compactness” are studies of inflectional formations. The exploratory study of Baayen et al. (2003) found that a sample of speakers produced Dutch nouns with a longer mean duration when they occurred as singulars than when they occurred as the stem of the corresponding plural. In a follow-up study, Kemps et al. (2005) tested speakers’ sensitivity to prosodic differences, and concluded that “acoustic differences exist between uninflected and inflected forms and that listeners are sensitive to them” (Kemps et al. 2005:441). That is, from the perspective of a Dutch speaker, singular forms like rat ‘rat’ and geit ‘goat’ do not recur in the corresponding plurals ratten ‘rats’ andgeiten ‘goats’ but instead have a distinctive prosodic profile that speakers are sensitive to. As Kemps et al. (2005:441) note in their conclusion, this effect calls into question a fundamental assumption of combinatoric approaches, namely that complex units are constructed via the assembly of simple units drawn from a fixed ‘alphabet’:
The prosodic mismatch effect documented in this study has important consequences for our understanding of the morphological structure of complex words. The way words are written in languages such as Dutch and English suggests that they consist of stems and affixes that are strung together as beads on a string. Phonemic transcriptions convey the same impression. Our experiments show that this impression is wrong. Plurals are not just singulars with an additional suffix. The precise acoustic realization of the stem provides crucial information to the listener about the morphological context in which the stem appears.
These sub-phonemic effects demonstrate that it is not sufficient for a morphemic model to solve recalcitrant segmentation and association problems; even success in segmentation and association is no guarantee of a correct analysis. A form like geiten can be segmented into a stemgeit and suffix-en, and the stem canbe assigned lexical properties while the suffix assigned grammatical properties. However, as Kemps et al. (2005) show, it is an error of analysis to identify the plural stem geit with the singular form geit or to associate plurality solely with the suffix -en. The morphotactic split between geit and -en does not correlate with a division in grammatical meaning between ‘caprine’ and ‘plurality’. Instead, the plural stem is tuned to its morphological environment. From a discriminative learning perspective, this suggests that the function of the affix is not characterizable just in terms of the grammatical meaning that it conveys but also involves the ‘stem-tuning’ context it provides for learners.
Taken together, the kinds of sub-phonemic effects described above suggest that the analytic assumptions of a morphemic model are motivated by an appeal to a conception of economy founded on imprecise orthographic and phonemic notions of ‘identity’. If seemingly recurrent units are not genuinely identical, then there is no redundancy in a description that contains their distributional variants, and consequently no economy is achieved by a description that collapses variants into a single item. In this case, morphemic analyses are merely ‘lossy’ descriptions that achieve no gain in “scientific compactness”.
Parallel remarks apply to ‘redundancy-free’ models of the lexicon, irrespective of whether the redundancy is eliminated by disassembling forms into separate inventories of minimal bases and exponents or by introducing recurrent forms within an inheritance hierarchy. No economy is achieved by ensuring that “an identical feature ... is noted only once” (Bloomfield 1933: 238) if the identity is an artifact of an imprecise encoding of contrasts.
-  As suggested in Section 8.6 below, it is ultimately contrasts rather than units that are significantin a morphological system from a learning-based perspective (i.e., phonemic contrasts do not implyunitary phonemes). As the contrasts discriminated in a system change as a function of learning, thederivative ‘units’ are expected to vary too.
-  See Anderson (2015) and Blevins (2016) for further discussion of the various senses of‘morpheme’proposed or assumed in different traditions.