Paper

BERT's Insights Into the English Dative and Genitive Alternations

---

Abstract

We investigate the English dative alternations using BERT by constructing two models that encode varying degrees of context to predict noun phrase order from their embeddings. The models can successfully predict dative alternations, even from context-absent embeddings, and behave similarly to traditional logistic regression approaches. The models also consider features like animacy, definiteness, and pronominality as more likely to appear first, aligning with established associations. This suggests that BERT embeddings encode relevant information in determining dative alternations. To assess whether BERT-based models possess knowledge of more general word order preferences, we consider the zero-shot transfer to predict genitive alternations, which share similarities with dative alternations both in terms of semantics and predictive modeling. Our models are reasonably successful in the transfer task, indicating some understanding of the shared factors that shape the two alternations. However, the transfer results show varying degrees of agreement with language production principles and known associations between features and alternations. This suggests that the determinants of alternations are not encoded in the BERT embeddings in an entirely compatible manner across constructions. These findings provide insights into the extent to which BERT exhibits human-like word order preferences and demonstrate the potential application of large language models in replacing hand-annotated features for corpus-based studies of syntactic knowledge.

Keywords: dative alternation, genitive alternation, syntactic alternation, BERT, language production principles

How to Cite:

Yao, Q. & Todd, S., (2024) “BERT's Insights Into the English Dative and Genitive Alternations”, Society for Computation in Linguistics 7(1), 52–62. doi: https://doi.org/10.7275/scil.2130