Title: A distributional approach to morphological relatedness
Olivier Bonami (Université de Paris)
Abstract:
Morphology can be seen as the study of the kinds of relations word entertains (see e.g. Haspelmath & Sims, chap. 1). Yet modern linguistics devotes strikingly little attention to the criteria that are used to decide whether two words are related, or whether two pairs of words are related in the same way: morphologists largely assume that the simple heuristics taught in introductory classes just need to be deployed at scale. However there are good reasons to doubt that this will work. As an example, consider the fact that English "go" and "went" are standardly assumed to be morphologically related, but "prison" and "carceral" are not. What is it about the difference between inflection and denominal adjective formation that justifies treating relatedness of meanings in the absence of relatedness of form differently? In another area, how do we decide whether two pairs of words stand in the same morphosemantic relation? Is "carceral" to "incarcerate" exactly what "criminal" is to "incriminate"?
Work in the tradition of Word and Paradigm has allowed some progress in this area, by diagnosing the absence of good segmentation heuristics (e.g. Spencer, 2012), highlighting the existence of morphological relations not reducible to affixation (e.g. Blevins, 2006), and developing systematic strategies for inferring such relations from unsegmented data at scale (Beniamine, 2021; Beniamine & Guzmán Naranjo, 2021, Bonami & Beniamine, 2021). Relatedness of meaning however still is largely uncharted territory, with most studies relying on qualitative examination of data samples that are small by necessity.
Distributional vector spaces (see e.g. Boleda 2020) are a promising way to approach this problem by allowing for a systematic assessment of the syntactic and semantic contrasts between pairs of related words. In this talk I will report on three recent case studies of French word formation that illustrate the fruitfulness and challenges involved. First I will show how contrasts between vectors can be used to assess the semantic similarity between derivational processes, and establish e.g. that "-ion" and "-ment" convey basically the same meaning (think of "incarceration" vs ."imprisonment"). Second I will present distributional evidence of the reality of paradigmatic relations in derivational morphology, i.e., systematic relations between derivatives despite the absence of a clear relation to a base (think of "optimist" vs. "optimism"). Finally, and time permitting, I will show that distributional vectors capture contrasts between native and learned morphology (think of "jailing" vs. "incarceration").
This talk is based on Joint work with Matías Guzman Naranjo (U. Tübingen) and Marine Wauquier (U. de Paris).
