A statistical inferencing engine
This general research tradition raises two points of particular relevance to classical WP models. The first is that a range of current studies suggests that speakers are sensitive to frequency and distributional relations involving word-sized units. This information is almost completely absent from theoretical models of the lexicon, which are inexplicably preoccupied with redundancy avoidance. The second is that the processing of a word form appears to be influenced (whether facilitated or inhibited) by related forms. This effect conflicts with the constructivist idealization that individual forms are defined in isolation. Morphological family effects seem to indicate that related word forms are not only represented in a speaker’s mental lexicon but that they are ‘co-activated’ in the processing of a given form. These effects appear especially compatible with models that link families of word forms into networks of elements with shared formal, grammatical, and semantic properties.
In a contemporary setting, this type of network architecture was initially associated with the linguistic models proposed by Bybee (1985, 2001) and with ‘associ- ationist’ computational models. The general-purpose connectionist or neural net models (Rumelhart and McClelland 1986; Plunkett and Marchman 1993) are the most familiar of these approaches, though networks have been developed more recently to model aspects of complex morphological systems, as in Pirrelli et al.
(2011) and Marzi et al. (2014). A fundamentally similar conception underlies the sort of classical model outlined by Paul (1920). In a formal reconstruction of this model, the lexicon is represented not as a static collection of minimal units but as a statistical inferencing engine, in which frequent forms and patterns support deductions about the shape and properties of unencountered (or infrequently encountered) forms. Within inflectional systems in particular, clusters of cells will define expectations about missing forms of a paradigm, with confidence levels that correspond to the informativeness of the predictive cells. The patterns exhibited by families of derivational forms will also define a space of analogical extensions. There is recent evidence that family size and analogical pressures influence the stress patterns in noun compounds (Plag 2010), complementing earlier studies that found analogy to play a role in the interpretation of new compounds (van Jaarsveld
et al. 1994).
A network architecture is combined with an inferencing (or uncertainty reducing) mechanism in the naive discriminative model outlined in Chapter 8.
-  Other patterns, such as phonological similarity with phonological ‘neighbours’ may generateseparate and even opposing expectations, as discussed in Pertsova (2004).