Morphological units and relations
Shifting the basis of classification from units of analysis to the relations between units also determines a general taxonomy of morphological models. The contrast between what has been termed ‘constructive’ and ‘abstractive’ perspectives (Blevins 2006b) is a version of a split that reflects the the contemporary influence of two ancient Indo-European grammatical traditions.
The Eastern tradition is represented by constructive (or ‘atomistic’) frameworks, which can be traced, via Bloomfield, to the Sanskrit grammarians. These frameworks describe patterns in a morphological system in terms of relations between minimal units within individual forms. The main variants of this approach are the concatenative (‘item and arrangement’ or ‘IA’) models, which disassemble forms into sequences of minimal units and operational (‘item and process’ or ‘IP’) models, which encapsulate units of form in processes. There are nearcontemporary variants of both models. Lieber (1992) presents a IA account, while Steele (1995) outlines an IP approach.
The Western tradition is represented by abstractive frameworks, which describe the patterns within a morphological system in terms of relations abstracted from forms and associated properties. Different approaches abstract different types of patterns and relations. The realizational model described in Section 1.2 above abstracts individual associations between grammatical properties and form variants and expresses them via ‘spell out’ rules. The implicational models
Figure 1.1 General classification of morphological models
proposed or assumed within the grammatical tradition that includes Paul (1920), Saussure (1916) and Kurylowicz (1949) exhibit a network organization. The structure of a language is defined by an association between a system of contrasts at the level of grammatical meaning and a system of contrasts at the level of form, but there need not be any correspondence between individual form contrasts and meaning contrasts. This conception was not formalized until the modern period. An axiomatic WP model is proposed in Lieb (1992, 2005). The approach initially outlined in Ackerman et al. (2009) uses information theory to represent implicational structure and discriminative learning models to associate systems of contrasts. The development of these models is summarized in Chapters 7 and 8.
The contrasts outlined above determine the classification in Figure 1.1, in which the two abstractive models subdivide the ‘WP archetype’ from Hockett (1954). Realigning Hockett’s taxonomy so that it classifies models in terms of relations helps to avoid some of the confusions and oversimplifications that can arise on unit-based classifications. The fundamental problem is that there is often a degree of ambiguity between morphosyntactic and morphotatic interpretations of morphological units. This problem is clearly illustrated by the familiar and, in many ways, useful contrast between ‘morpheme-based’ and ‘word-based’ accounts. IA and IP models are both morphosyntactically ‘morpheme-based’, given that processes encapsulate the same grammatical information as morphemes. But unlike IA models, IP models are not morphotactically formative- or morph-based, since processes can apply to an input without any mediating segmental unit. Likewise, classical WP models and realization-based approaches are both mor- phosyntactically ‘word-based, since they treat the word as the smallest grammatically meaningful unit of a grammar. Classical models are also morphotactically word-based, in that they treat surface word forms as the basic form units of a system, and regard roots, stems and exponents as abstractions over a lexicon of full word forms. However, the morphotactic assumptions of realizational approaches pattern more with those of IA and IP models, in which surface word forms are assembled from smaller elements. Realization-based models assume a model of the lexicon in which open-class items are represented by minimal roots or stems, and surface word forms are defined in terms of these units through the application of realization rules. This places realizational approaches between classical WP models and constructive accounts, at least with respect to their ontological assumptions. Whereas realization-based approaches agree with classical WP approaches about the morphosyntactic status of words, they agree with constructive models in treating surface words as derived units.
Underlying the development of these morphological models (and theories) is the assumption that the word stock of a language is not an unstructured form inventory. Words are taken to exhibit various types of systematic patterns and interdependencies, reflected in their assignment to morphological ‘families’, ‘paradigms’, ‘classes’ and ‘series’. The models outlined in Figure 1.1 differ in what they identify as the locus of this organization, and in the strategies that they deploy to express patterns. It is instructive to isolate these strategies and accompanying notions of locality before examining the more technical form in which they are incorporated into contemporary accounts.
The information-theoretic approach developed within the implicational WP tradition offers a useful perspective on the common goals shared by these models, and clarifies how the strategies employed in different models serve to reduce the uncertainty involved in associating properties and forms. Abstractive models express the structuralist view that languages form “un systeme ou tout se tient” in which the correspondence between properties and forms is established in the context of a larger network of associations. The structuralist morpheme encapsulates a much more local conception, in which this correspondence is defined between individual properties and the smallest isolable units of form in a system, so that uncertainty can be fully resolved without reference to other formatives, words, or any type of paradigmatic structure.
These differences between abstractive and constructive perspectives are reflected in the ‘granularity’ of lexical units. As one moves from left to right in Figure 1.1, the morphotactic and morphosyntactic components of lexical entries become progressively larger. At one extreme lie the fundamentally ‘atomistic’ lexicons of constructive accounts, which associate minimal collections of lexical properties with (segmentally minimal) root morphemes. In concatenative accounts, minimal sets of grammatical features are also associated with grammatical morphemes, whereas in operational approaches these features are encapsulated within processes. The lexicon of an abstractive model is, by comparison, more ‘molecular’. Realizational models associate intrinsic lexical features with root (or stem) entries, but treat the paradigm cell—or ‘morphosyntactic representation’ (Anderson 1992)—as the smallest unit containing lexical and grammatical features. Hence realizational models are morphosyntactically molecular but morphotactically atomistic. Implicational models lie at the opposite extreme from concatenative accounts, as they assume a lexicon consisting mainly of (partially) instantiated paradigms.
Although the WP tradition is united by a common origin and a stock of core assumptions, individual WP models embody different and at least partly complementary conceptions of ‘language’, ‘grammar’ and ‘morphological analysis’. The exemplar-based format of classical WP models reflects their pedagogical heritage. Pedagogical goals or idealizations also underlie many of the properties of these models, from their appeal to effective but unspecified processes of analogy, to the completeness and uniformity of the exemplary paradigms and principal part inventories that they exhibit. The language described by a classical grammar corresponds to the set of forms jointly defined by the exemplary paradigms and principal parts of the grammar. Realizational and implicational models generalize this conception in characteristically different ways.
Realizational models adopt the perspective of formal language theory, in which a ‘language’ is a set of expressions or structures and a ‘grammar’ is a device for enumerating these sets. Realizational approaches thus extract the generalizations exhibited by the exemplary analyses of a classical WP model and encapsulate these patterns within an explicit system of rules and constraints. The network-based conception developed by implicational models expresses the patterns exhibited by systems of contrasts in terms of information theoretic measures (Shannon 1948) and discriminative learning models (Rescorla and Wagner 1972; Rescorla 1988). These formal differences reflect fundamentally divergent perspectives on the nature of ‘grammatical knowledge’. Whereas a realizational perspective expresses a speaker’s knowledge of a morphological system in terms of a symbolic rule system, an implicational perspective expresses this knowledge in terms of the state of a learning model.
These alternative conceptions of ‘grammar’ and ‘knowledge’ lend themselves to different types of implementation. The symbolic rule systems of realizational models are particularly amenable to NLP implementation. Models of Network Morphology (Corbett and Fraser 1993) adopt a general purpose, cognitively neutral, knowledge representation language, DATR, for the formulation and implementation of analyses. Work within the finite state morphological tradition has likewise shown that the descriptions assigned by models of Paradigm Function Morphology (Stump 2001) can be compiled into finite state transducers (Karttunen 2003). In contrast, network-based implicational models fall within a cognitive tradition that includes ‘emergentist’ (Bybee 1985, 2010), ‘usage-based’ (Tomasello 2003; Diessel 2015) and ‘construction-based’ (Goldberg 2005; Booij 2010) branches. This conception is more compatible with cognitively motivated models of comprehension, production and learning such as Baayen et al. (2011) and Ramscar et al. (2013a).
-  See Matthews (1994) and Cordona (1994) for surveys of these traditions.