Séminaire de Recherche en Linguistique

-Symbolic linguistic frameworks (including, for example, Minimalism, Principles and Parameters, Head Driven Phrase Structure Grammar, Lexical Functional Grammar, Construction Grammar, etc.) provide a refined portrayal of the grammatical deployment of phrasal types in languages across the world.    However, there is a phenomenon of historical language change, grammaticalization, which both vindicates and challenges the assumptions of these formalisms.    Grammaticalization is the term given to the tendency for content morphemes (e.g., nouns, verbs, adjectives) to spawn new functional elements (e.g., prepositions, auxiliarly verbs, complementizers, etc.)  as a language evolves across historical time.   Very often, in grammaticalization, a succession of slight changes leads, over time, to a  categorical  change of phrasal analysis.  For example, English “sort ” and “kind” were once simply nouns that could be modified by prepositional phrases (e.g., “We found a sort/kind of brick”) but now, following a protracted development that I document with historical corpora, “sort of” and “kind of” have acquired a new, and very high frequency role as degree modifiers (e.g., “It’s sort/kind of cloudy today”).     Symbolic formalism is good at making it clear that a structural change has taken place and at characterizing the grammatical nature of the endpoints of such historical developments.   What it is not as good at is explaining how speakers keep track of the location at which their competence resides along the gradual cline between the historical beginning point (Noun-Preposition sequence) and end point (Degree Modifier).  Here I propose that connected space models (e.g., neural networks) have a useful contribution to make.  I show how a learning neural network model predicts an observed correlation between changes in statistical distribution and changes in grammatical behavior.  I also identify a formal approach to neural network models which allows us to see beyond the “black box” character of these models, making the relationship between their claims and linguistic analyses clear in a way that is not usually done.