Different lexicons make different rivals

Analogy-based morphological theories assume that language users create novel words or word forms on the basis of similar existing forms in their Mental Lexicons (Skousen 1989; Daelemans & van den Bosch 2005; Baayen et al. 2011). Rivalry between affixation processes is, in this view, resolved on the basis of the structure of similarity distributions among existing words with competing affixes. Especially computationally implemented analogy-based models have been shown to be very successful in modelling pertinent rivalries empirically (cf., e.g., Chapman & Skousen 2005; Eddington 2006; Arndt-Lappe 2014). Interestingly, however, simulation studies of morphological rivalry have almost invariably adopted a rather abstractionist view of the Mental Lexicon, simulating its contents in terms of large collections of affixed words gleaned from dictionaries or corpora. The lexicon models used thus represent the word stock of the language, abstracting away from differences between individual speakers’ vocabulary knowledge. This is a problem because it precludes the possibility of testing a central prediction that analogical theories are making: if affixes are assigned on the fly on the basis of similar words in the lexicon, then speakers with different lexicons should make different choices in situations of competition. The present paper provides a proof-of-concept study addressing these issues for one testbed phenomenon: the form-based rivalry between the two English adjectival suffixes –ic and –ical, a case that is well-understood in the literature (esp. Lindsay & Aronoff 2013; Aronoff & Lindsay 2014) and hence provides an ideal test case for modeling. As a computational model, we use AML (Analogical Model of Language, Skousen, Stanford & Glenn 2013), in conjunction with a graphical user interface (TrAML, Transparent Analogical Model of Language, Arndt-Lappe et al. 2018) that makes detailed information about AML models available in format that is easily accessible for further statistical analysis. The paper provides a series of simulation studies in which we will explore how AML predictions for novel words change when assuming different input lexicons. Starting from a database of all –ic and ¬–ical adjectives contained in CELEX (863 word types), we define two different test lexicons based on register-specific profiles of –ic and –ical words in the British National Corpus (BNC). Thus, corpus analysis shows that most –ic and –ical words occur in academic registers only, with only a small proportion occurring in spoken language. On this basis, we compare predictions of the analogical model for the full lexicon (including academic words) and a lexicon of a hypothetical (group of) speaker(s) who do not know any academic words. Analysis of our AML simulations reveals that the two types of model share some basic properties, but make clear - and testable - predictions with regard to speaker differences in situations of competition between the two affixes.