The discriminative perspective
This general orientation places a learning-based model within a larger tradition that includes overlapping ‘emergentist’ (Bybee 1985, 2010), ‘usage-based’ (Tomasello 2003; Diessel 2015) and ‘construction-based’ (Goldberg 2005; Booij 2010) branches, along with other broadly ‘cognitivist’ frameworks. In the context of this tradition, a learning-based morphological approach could be described as ‘a usage-based approach with a discriminative learning model’.
Sections 8.3 and 8.4 suggest that the classic WP model (as well as the model presented in Bloomfield 1933) are best interpreted in discriminative terms. Thus, for example, the idea that form variation serves to discriminate larger forms, rather than to express individually meaningful contrasts, seems implicit in Matthews’s description of the classical WP model:
But there is an alternative method, whose sources lie in the work of the ancient grammarians of Greek and Latin. This is simply to relate words as wholes. (Matthews 1991:186)
Yet the implications of this perspective remain implicit through the whole WP tradition, as reflected in Robins’s various criticisms of the failure of classical models to develop “a theory of the morpheme” (Robins 1997:31):
It was certainly a weakness in the classical grammarians, and in many later writers who followed their example, that they barely recognized any grammatical unit below the level of the word, and certainly never set out with any rigour the establishment of the morphemes of a language. (Robins 1959:119)
The point of departure for a discriminative perspective on language is the learning rule of Rescorla and Wagner (1972) and Rescorla (1988). This rule provides the basis for the discriminative models developed over the past decade in a series of papers by Ramscar and associates (Ramscar and Yarlett 2007; Ramscar and Dye
2010; Ramscar et al. 2010, 2013a; Ramscar 2013). The claim that form variation serves a discriminative function is first explicitly stated in Ramscar and Yarlett (2007:931) and elaborated by Ramscar and Dye (2010) below:
While the approach of Haskell et al. (2003) is broadly comparable to our own, we would suggest that a main function of phonology is to discriminate between semantic alternatives (Ramscar & Dye, 2009; Ramscar & Yarlett, 2007; Ramscar et al., 2010). For this reason, we expect that phonological forms that are more discriminable in any given context will be more useful (i.e., informative) in this regard. Given that the presence or absence of a final sibilant is used to discriminate between the plural and singular forms of most nouns, and that there are conventions that apply to plurals in compounds, we would expect that when the phonological forms of singular nouns are easily distinguished from forms with final sibilants, this will be more informative—both about plurality, and about the conventional status of a given compound—than when they aren’t (i.e., rats may be a more informative plural form than both mice and houses). (Ramscar and Dye 2010:28)
There are also antecedents for a broadly discriminative conception of meaning and communication. However, Ramscar et al. (2010) provide the first statement of this perspective in an explicitly discriminative model:
If symbolic communication involves predicting symbols from meanings (and context)—and we have outlined many reasons for assuming that it does—then meaning is something that a speaker elicits in a listener simply by engaging the listener in a game of prediction. In this game, symbols are not used to convey meaning, but rather are used to reduce a listener’s uncertainty about a speaker intended message (Shannon, 1948). In order for a listener to predict a speaker, the listener has to activate the same semantic cues to symbolic form as the speaker, such that the listener comes to understand an utterance by thinking about that utterance in a way that converges on that of the speaker. This proposal has much in common with the idea that language is a form of joint action (see e.g., Altmann & Mirkovic, 2009; Clark, 1993; Garrod & Pickering, 2009; Gennari & MacDonald, 2009; Pickering & Garrod, 2007; Tanenhaus & Brown-Schmidt, 2008); it differs in that it is explicitly nonreferential. (Ramscar et al. 2010:35)
The conception of communication as a process of cooperative ‘uncertainty reduction’ also clarifies the relation between a discriminate perspective and the information-theoretic measures discussed in Section 8.2.2. Discriminative learning involves the reduction of uncertainty in the information-theoretic sense. Hence the types of entropy measures discussed in Chapter 7 can be seen as providing a global estimation of the uncertainty encountered by a discriminative learning network. The linkage between entropy measures and discriminative learning emerges perhaps most clearly in models that use the Rescorla-Wagner learning equations for comprehension and production.
-  “Understanding language in terms of learning—and without underspecified appeals to reference—involves a reassessment of what human communication involves, requiring revised theories of languageand its role in culture (Quine, i960; Tomasello, 1999, 2003; Wittgenstein, 1953; see also Fodor, 2000).”(Ramscar etal. 2010: 34)
-  For example, in current implementations of Implicit Morphology (Baayen et al. югг, югб),the forms abstracted from the speech stream are encoded by networks of n-phone (and n-graph)units, associated to a system of units encoding lexically and grammatically contrastive properties,with connection weights between units estimated from corpora using the Rescorla-Wagner learningequations.