Token frequency as a determinant of morphological change

Abstract

This paper demonstrates that morphological change tends to involve the replacement of low frequency forms in inflectional paradigms by innovative forms based on high frequency forms, using Greek data involving the diachronic reorganisation of verbal inflection classes. A computational procedure is outlined for generating a possibility space of morphological changes which can be represented as analogical proportions, on the basis of synchronic paradigms in ancient Greek. I then show how supplementing analogical proportions with token frequency information can help to predict whether a hypothetical change actually took place in the language’s subsequent development. Because of the crucial role of inflected surface forms serving as analogical bases in this model, I argue that the results support theories in which inflected forms can be stored whole in the lexicon.

Analogy in Inflection

“…The frequency with which speakers encounter forms is another significant factor. Sims-Williams (2022) shows that in the history of Greek, the textual token frequencies of the cells in a paradigm are correlated negatively with the likelihood of a cell undergoing change and positively with its likelihood of serving as an analogical model for change. Boyé & Schalchli (2019) argue that models of inflectional analogy are unrealistic if entropy is computed from complete systems containing a wordform for every paradigm cell; rather, a more faithful model should be able to learn from partial data that contain gaps, similar to what a speaker may have been exposed to.…”

OSU at SigMorphon 2022: Analogical Inflection With Rule Features

“…Many theories of inflection production propose a central role for memorized word forms in shaping the outcomes for unknown or weakly represented words (Bybee, 1995). In such memory-based models, speakers retrieve exemplar forms A from memory for which the outcomes B are known and use them to predict the outcome for a word C via a process of analogical reasoning: exemplar source A : exemplar target B :: source C : target D. This type of analogical reasoning is detectable in historical changes (Sims-Williams, 2021) and in experiments with nonce-words (D ąbrowska, 2008), and underlies some influential computational models of inflection (Albright and Hayes, 2003;Daelemans, 2002). Recently, Elsner (2021) and Liu and Hulden (2020) show that transformer models for inflection prediction can also benefit from access to exemplars.…”

Proceedings of the 19th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology

“…Many theories of inflection production propose a central role for memorized word forms in shaping the outcomes for unknown or weakly represented words (Bybee, 1995). In such memory-based models, speakers retrieve exemplar forms A from memory for which the outcomes B are known and use them to predict the outcome for a word C via a process of analogical reasoning: exemplar source A : exemplar target B :: source C : target D. This type of analogical reasoning is detectable in historical changes (Sims- Williams, 2021) and in experiments with nonce-words (D ąbrowska, 2008), and underlies some influential computational models of inflection (Albright and Hayes, 2003;Daelemans, 2002). Recently, Elsner (2021) and show that transformer models for inflection prediction can also benefit from access to exemplars.…”