Is inflectional irregularity dysfunctional to human processing?

Regularly inflected verb forms are classically associated with the formal transparency andpredictability of their internal constituents [ 1 , 2 , 3 ]. Transparency ensures that full forms can besegmented uniquely into their internal constituents: as in walk-s/walk-ed. Predictability allowsfor a speaker to fill in an empty paradigm cell, using information from other known forms of thesame lexical paradigm and its inflection macro-class. From this perspective, irregulars appear tobe dysfunctional to the human processing system, as they make it hard to infer - say - boughtfrom buy , or segment bought appropriately into its constituent parts. Likewise, an influentialpsycholinguistic tradition relegates irregulars to the lexical store, whereas regulars are segmentedby rules into their simpler constituents [ 4 , 5 ].Here, we offer a few reasons for questioning this view. First, transparency andpredictability are not dichotomous notions. Secondly, their influence on processing is notunidirectional. Unpredictable stems in irregularly inflected forms of complex inflectionalsystems provide a lot of processing information, by dynamically constraining the number ofpossible alternative endings during serial processing. Thirdly, acquisition of word inflection doesnot consist in associating co-occurring cues and outcomes, but in discriminating betweenmultiple cues that are constantly in competition for their predictive value for a given outcome.We present the results of a few computer simulations with Self-organising RecurrentNeural Networks (TSOMs, [ 8 , 9 ]) that learn how to inflect high-frequency verb paradigms in 6languages: English, German, Italian, Modern Greek, Modern Standard Arabic and Spanish. Aftertraining, each TSOM was tested on a word recognition (serial recoding) and a word production(serial recall) task, and results were analysed with generalised regression models. Processinguncertainty is differently apportioned on regulars and irregulars, depending on the nature of theprocessing task. While irregulars are harder to produce when they are unknown because theytypically have fewer neighbours than regulars have, they are readily accessed once they areacquired, for exactly the same reason.Our data are in line with psycholinguistic evidence [ 10 , 11 ] that lexical processing ispaced by two types of uniqueness point: Marslen-Wilson's Uniqueness Point (UP),distinguishing unrelated onset-overlapping words [ 12 ], and the Complex Uniqueness Point(CUP), distinguishing paradigmatically-related words [ 11 ]. Late UPs are inhibitory and elicitprolonged reaction times in acoustic word recognition, explaining an early delay in wordrecognition of irregular stems. Similarly, late CUPs are inhibitory, and this accounts for aslowdown in the processing advantage of regulars, compared to irregulars, after UP. Thesestructural factors interact in a variety of ways and concurrently affect human processing, to showthat irregularly-inflected forms may in fact reflect communicative and processing constraints ofthe word processor. They provide strong evidence against a processing architecture that assumescompartmentalized, independent processing routes for some specific combinations of thesefactors (e.g. a rule-based route for a combination of transparency and predictability, and amemory-based route for all other combinations). In addition, they seem incompatible withBayesian approaches to auditory word comprehension ignoring a word's internal structure [ 13 ].We suggest that a different design of the human language processor, based on a computationalarchitecture integrating memory and processing as two different dynamics of the sameunderlying mechanism, can shed light on the complexity of inflection, and vindicate the role ofirregular inflection in the system.