The Making and Breaking of Classification Models in Linguistics

The book provides a methodological blueprint for the study of constructional alternations using corpus-linguistic methods in combination with different types of experimental data. The book looks at a case study from Estonian. This morphologically rich language is typologically different from Indo-European languages such as English. Corpus-based studies allow us to detect patterns in the data and determine what is typical in the language. Experiments are needed to determine the upper and lower limits of human classification behaviour. They give us an idea of what is possible in a language and show how human classification behaviour is susceptible to more variation than corpus-based models lead us to believe. Corpora and forced choice data tell us that when we produce language, we prefer one construction. Acceptability judgement data tell us that when we comprehend language, we judge both constructions as acceptable. The book makes a theoretical contribution to the what, why, and how of constructional alternations.

Autorentext
Jane Klavan, University of Tartu, Estonia.

Klappentext

The book offers a cognitively realistic account of morphosyntactic alternations through a combination of corpus-based analysis and (psycho)linguistic experiments.

The central thesis of the book is that the choice between competing constructions is driven by semantic, syntactic, morphological, phonological, discourse-related, and lectal features. The study outlines ways to assess the relative importance and cognitive plausibility of the various factors by advocating methodological pluralism - one of the pillars of a cognitively realistic linguistic theory. A deeper understanding of the cognitive grounding of the probabilistic structure of language is developed as reflected in the patterns and structures we see across different types of language data. The book outlines some of the different modelling techniques available for capturing the multivariate and probabilistic nature of linguistic knowledge quantitatively.

It is argued that for making cognitively realistic statistical models, we first need to break these models by pitting different modelling techniques against each other and against human classification behaviour as manifested in a range of (psycho)linguistic experiments.