Syntactic Structure Distillation Pretraining for Bidirectional Encoders

• Published in: Transactions of the Association for Computational Linguistics Vol. 8; pp. 776 - 794
• Main Authors: Kuncoro, Adhiguna, Kong, Lingpeng, Fried, Daniel, Yogatama, Dani, Rimell, Laura, Dyer, Chris, Blunsom, Phil
• Format: Journal Article
• Language: English
• Published: One Rogers Street, Cambridge, MA 02142-1209, USA MIT Press 01-01-2020; MIT Press Journals, The
• Subjects: Benchmarks; Bias; Context; Distillation; Error reduction; Language modeling; Natural language; Natural language processing; Questions; Representations; Syntactic structures; Syntax; Words (language)
• ISSN: 2307-387X; 2307-387X
• DOI: 10.1162/tacl_a_00345

Textual representation learners trained on large amounts of data have achieved notable success on downstream tasks; intriguingly, they have also performed well on challenging tests of syntactic competence. Hence, it remains an open question whether scalable learners like BERT can become fully proficient in the syntax of natural language by virtue of data scale alone, or whether they still benefit from more explicit
. To answer this question, we introduce a knowledge distillation strategy for injecting syntactic biases into BERT pretraining, by distilling the syntactically informative predictions of a hierarchical—albeit harder to scale—syntactic language model. Since BERT models masked words in bidirectional context, we propose to distill the approximate marginal distribution over words in context from the syntactic LM. Our approach reduces relative error by 2–21% on a diverse set of structured prediction tasks, although we obtain mixed results on the GLUE benchmark. Our findings demonstrate the benefits of syntactic biases, even for representation learners that exploit large amounts of data, and contribute to a better understanding of where syntactic biases are helpful in benchmarks of natural language understanding.