Models don’t explain. Theoretical and empirical challenges to modelling syntactic alternations
Ludovic De Cuypere, Hilde De Vaere, Thomas Belligh, Klaas Willems · Ghent University Academic Bibliography (Ghent University) · 2021
The probabilistic grammar approach to syntactic alternations is arguably one of the most successful linguistic paradigms of the last two decades. Fitting a multifactorial statistical model to a randomized set of corpus sentences, optionally triangulated with an acceptability judgement task to evaluate the corpus model is now a standard approach in syntactic research. Apart from the fact that the empirical steps of the approach are rather straightforward (i.e., retrieve-annotate-model), part of its appeal is also that the approach is theoretically rather agnostic. However, the excellent predictive quality of our models does not relieve one from theoretical reflection. In this talk we will not argue that we need more complex modelling tools (i.e., deep learning), nor will we suggest improvements on how to measure certain variables (e.g., different collocational measurements). Inspired by the workshop abstract, we tackle two deeper questions. (1) What are the variants of the alternation? (2) Are alternation factors cognitively real? With respect to question (1), we are particularly interested in the constructional status of the variants associated with the German ditransitive alternation. The latter involves the alternation between the Indirect Object Construction with an ADDRESSEE expressed in the dative case and the Prepositional Object Construction where the same role is introduced by a prepositional object. Combining the view that coded meanings ought to be distinguished from inferred information (Levinson 2000, Carston 2012, Coene & Willems 2006) with the view that distinguishes between constructemes and allostructions (Cappelle 2006 and Perek 2015), we make the case that IOC and POC are two constructional (i.e., semantically similar) variants (allostructions) related to one general three place GOAL-construction. We additionally discuss how this bears on the relationship between lexical and constructional semantics as well as the valency vs. constructionist approaches to syntactic alternation. With respect to the second question, one key observation is that all mathematical models are a simplification of reality. This is not necessarily problematic. However, a model can also be fundamentally wrong. A logistic regression model, for instance, can be interpreted as a cause-effect structure, with the choice of construction as the effect and the input variables as the cause. We make the case that this interpretation is linguistically/cognitively wrong. For one, a speaker does not make a “choice” based on a set of input variables, as this would imply that the latter precede the former. References Cappelle, B. (2006). Particle placement and the case for "allostructions”. Constructions SV1,7. 1-7. Carston, R. (2012). Word meaning and concept expressed. The Linguistic Review 29(4). 607-623. Coene, Ann & K. Willems. 2006. Konstruktionelle Bedeutungen. Kritische Anmerkungen zu Adele Goldbergs konstruktionsgram¬ma¬ti¬scher Bedeutungstheorie. Sprachtheorie und germanistische Linguistik 16. 1-35. Perek, F. (2015). Argument Structure in Usage Based Construction Grammar. Amsterdam, Benjamins. Levinson, S. C. (2000). Presumptive meanings. The theory of generalised conversational implicature. Cambridge, MIT Press.