Data-Rich Approaches to English Morphology: From corpora and experiments to theory and back

(in alphabetical order of author’s last name) Measuring the unobservable: quantifying paradigm gaps and morphological retreat Adam Albright, MIT Observing the non-occurrence of ‘expected’ forms is a well-known problem for language acquisition and corpus studies (the problem of negative evidence). In morphology, expected forms may be missing for a variety of reasons, including paradigms gaps, selectional restrictions, and decreasing productivity of an affix. For example, for many speakers of American English, the verb ‘stride’ has a paradigm gap in the past participle (‘He has *stridden/*stroden/*strode’). Synchronically, this fact is reflected in the low token frequency of (etymologically expected) ‘stridden’ relative to other forms (e.g., present ‘stride’, past ‘strode’); however, merely observing low token frequency does not prove that stridden is rarer than expected, let alone explain why it is rare. In this talk, I consider how diachronic data can help document the existence of paradigm gaps, and offer insight into how gaps and selectional restrictions arise. I show that we can observe at least two different sources of restrictions: (1) failure to generalize an affix beyond its original domain of application, and (2) ‘retreat’ of an affix to a smaller phonological domain. I argue that both patterns are predicted by a conservative inductive learning model, such as the Minimal Generalization Learner (Albright and Hayes 2003). The first goal of the talk is empirical: using data from the Corpus of Historical American English (COHA; Davies 2011) and the Google n-grams corpus (Michel et al. 2011), I show that it is possible to observe the gradual erosion of previously attested forms such as ‘stridden’. A basic obstacle to observing underattestation is that we must estimate the expected frequency of affixed forms. In English, the token frequency of a verb’s past participle is related to the verb’s lemma frequency, but it depends on many additional factors as well (argument structure, collocations, real world factors), some of which are difficult to estimate. However, these factors appear to be relatively stable across time, such that in general, the relative frequency of the participle of a given verb tends to remain fairly constant. Fig 1 shows this for the proportion of ‘sung’ (etc.) tokens relative to present/infinitive ‘sing’ in the American English Google n-grams corpus. Not all verbs show such constancy, however. In particular, many of the verbs in the ‘stride’ class show a gradual decrease in the ratio of participle forms over the past 100 years, shown in Fig. 2 (normalized for lemma frequency). A loglinear model confirms that participles of this class have been waning over time, affecting especially (but not only) low frequency verbs like ‘stride’ and ‘strive’.