Assigning priority to relations within a system, rather than to inventories of elements, is more characteristic of ‘European Structuralism’. On one reading of
Saussure (1916), the units of a linguistic system are defined in terms of their opposition to other units and have no value outside that system:
Applique a l’unite, le principe de differenciation peut se formuler ainsi: les caracteres de l’unite se confondent avec l’unite elle-meme. Dans la langue, comme dans tout systeme semiologique, ce qui distingue un signe, viola tout ce que le constitue. Cest la difference qui fait le caractere, comme elle fait la valeur et l’unite.
Parmi les oppositions quelle comprend, il y en a qui sont plus significatives que d’autres; mais unite et fait de grammaire ne sont que des noms differents pour designer des aspects divers d’un meme fait general: le jeu des oppositions linguistiques. (Saussure 1916: i67f., emphasis in original)
This perspective again tends to be implicitly discriminative, as reflected particularly in the claim that units are defined by oppositions with contrasting units (“ce qui distingue un signe, viola tout ce que le constitue”). The primacy of relations is stressed even more explicitly by Hjelmslev (1948) when he asserts in the passage below that “the real units oflanguage are the relata”:
the real units oflanguage are not sounds, or written characters, or meanings: the real units of language are the relata which these sounds, characters, and meanings represent. The main thing is not the sounds, characters, and meanings as such, but their mutual relations within the chain of speech and within the paradigms of grammar. These relations make up the system of a language, and it is this interior system which is characteristic of one language as opposed to other languages, whereas the representation by sounds, characters, and meanings is irrelevant to the system and may be changed without affecting the system. (Hjelmslev 1948:27)
Identifying the unifying properties of morphological systems as the relations between elements helps to clarify why unit-based typologies lead to agnosticism (Matthews 1972), diachrony (Anderson 1992) or despair (Aronoff 1998). A learning- based approach suggests a further refinement in which the primacy of discriminative and implicational relations reflects their role in reducing uncertainty and thereby aiding learning and communication.
As acknowledged at the beginning of this chapter, the reconstruction of the classical WP model developed in previous chapters can be separated from its reconceptualization in terms of a learning-based approach. But that is to say that it is possible to consider How a WP model can be applied to the analysis of morphological systems without considering why it applies so generally. A discriminative learning perspective offers a cohesive frame of reference in which to interpret the properties of WP models and the systems that they model.
At the system level, the analogical generalizations expressed (explicitly or implicitly) within classical WP models exploit the implicational dependencies exhibited by morphological systems. But these models provide no insight into why the dependencies exist in the first place. Information theory provides measures that are useful for quantifying dependencies in terms of uncertainty and uncertainty reduction. But the applicability of these measures does not clarify why they, like classical models in general, work as well as they do.
At the level of individual analyses, the advantages of WP approaches largely derive from the that fact that what have subsequently been labelled ‘abstractive’ analyses (Blevins 2006b) can describe the composition and structure of forms that cannot be ‘constructed’ from their minimal parts (or at least not without the aid of diacritic features that encode ‘assembly instructions’). Yet apart from the general observation that word-sized units appear more stable and informative than smaller units, the basis for the descriptive advantages of abstractive analyses of morphological systems remains mostly unexamined.
A discriminative learning approach offers answers to both of these questions. Implicational dependencies exist because they are a precondition for learning from sparse, biased input. The descriptive success of abstractive analyses derives from the fact that the form variation in the systems they are describing serves a discriminative function. These answers raise further questions, many of which are currently under investigation in the literature cited in this chapter. However, by shifting the debate from ‘how’ to ‘why’ questions, this perspective holds out the prospect of explanations for the organization of morphological systems, grounded in factors external to those systems.
-  Applied to units, the principle of differentiation can be stated in this way: the characteristicsof the unit blend with the unit itself. In language, as in any semiological system, whateverdistinguishes one sign from the others constitutes it. Difference makes character just as it makes valueand the unit. Some of its oppositions are more significant than others; but units and grammatical facts are onlydifferent names for designating diverse aspects of the same general fact: the functioning of linguisticoppositions. (Saussure 1959: inf.)