2 |
Human-like learning of syntactic islands by neural networks ...
|
|
|
|
Abstract:
Recently, computational linguists have shown that neural networks without any prior syntactic knowledge can induce human-like grammatical knowledge (Linzen & Baroni, 2021). This success has not been attested for all syntactic phenomena, however. Syntactic island constraints still receive mixed results. Wh-, complex NP, coordination, adjunct and left branch islands are, for example, successfully learned in most studies, but negative phrase, relative clause and (sentential) subject islands only partially or not at all (Chaves, 2020; Chowdhury & Zamparelli, 2018; Wilcox et al., 2018; Wilcox et al., 2019; Wilcox et al., 2021). Furthermore, the neural network can even learn subtle differences between these islands, for example whether an island type is experienced as strong or as weak. Coordination islands are, for instance, learned as very strong islands, while wh-islands are learned as less strong (Wilcox et al., 2021). While these island constraints have been tested with different computational models ...
|
|
Keyword:
Arts and Humanities; Computational Linguistics; Dutch Studies; FOS Languages and literature; Linguistics; Social and Behavioral Sciences
|
|
URL: https://osf.io/23teq/ https://dx.doi.org/10.17605/osf.io/23teq
|
|
BASE
|
|
Hide details
|
|
|
|