Extraction from finite adjunct clauses: an investigation of relative clause dependencies in Norwegian

Abstract

Recent experiments have confirmed earlier informal evidence that finite adjuncts are not islands categorically. Specifically, it has been shown that adjuncts are not necessarily islands for all dependency types (Sprouse et al. 2016), and that the island status of an adjunct depends on the type of the adjunct clause in question (Kush et al. 2019; Müller 2019; Bondevik et al. 2021; Nyvad et al. 2022). The current study further explores these questions by testing three different adjunct clause types: Clauses introduced by om ‘if’, fordi ‘because’ and når ‘when’, in a relative clause (rc) dependency in Norwegian. We find that forming an rc-dependency into a finite adjunct in Norwegian overall causes island effects, but that there are fine-grained differences within the category ‘adjunct’. Specifically, we find that fordi ‘because’ and når ‘when’ yield large island effects, while om ‘if’, on a par with Kobzeva et al. (2022) and Nyvad et al. (2022), yields intermediate results. Rather than relying on binary distinctions only, we argue that any theory that is to explain the empirical landscape must be sufficiently fine-grained and allow for gradient distinctions.

Neural Networks Can Learn Patterns of Island-insensitivity in Norwegian

“…However, the average filler effects are not 0 in the adjunct island condition, suggesting that the models might not treat them as full islands. 2 Norwegian shows some variation in adjunct island effects, with extraction from conditional adjuncts rated higher than from temporal and reason-adjuncts (Bondevik et al, 2021;Bondevik and Lohndal, 2023). The result obtained here could be explained by the models' sensitivity to this variation (and potential overgeneralization).…”

“…Adjuncts are said to block FGD formation, which explains the unacceptability of ( 12): The filler what cannot be associated with the gap inside the adjunct when-clause. Norwegian, like English, does not al-low gaps inside temporal adjuncts (Bondevik et al, 2021;Bondevik and Lohndal, 2023). We created 30 items according to a 2 × 2 × 3 design that crossed FILLER, GAP, and CONDITION for each dependency type (360 test sentences per dependency).…”