Constructions are influenced in their realization by the degree of activation of horizontally linked constructions: using BERT sentence embeddings to test contamination between constructions

Abstract Pijpops and Van de Velde (2016. Constructional contamination: How does it work and how do we measure it? Folia Linguistica 50(2). 543–581) demonstrated that an alternation within a construction may be influenced by a different construction that is coincidentally formally similar through the process of constructional contamination. This mechanism has been explained as a consequence of the storage of unanalyzed chunks produced by a contaminating construction, which are later recycled in a target construction. Consequently, constructional contamination has hitherto been measured by calculating the frequency of individual lexemes (which are assumed to be part of exemplar chunks eligible for recycling) in the contaminating construction. In this paper, we propose a different measurement for constructional contamination. Using the large language model RobBERT, we define “prototype embeddings” for both the target construction and the contaminating construction involved in constructional contamination. For each instance of the target construction, we calculate its similarity to these prototypes. We then show that these similarity variables are significantly correlated with the alternation within the target construction: the more an instance of the target construction is semantically similar to the contaminating construction, the more likely it is to appear with the alternant that is exclusively used in that contaminating construction. Our results are in line with a scenario in which horizontal links, which hold between formally similar constructions, can directly trigger constructional contamination. In addition, we find that the more semantically similar an instance is to the prototype of the target construction itself, the more likely it is to appear with the alternant that is not found in the contaminating construction, supporting a view of the constructional network in which constructions are themselves conceptualized as links rather than nodes.