Have Attention Heads in BERT Learned Constituency Grammar?
This work addresses the interpretability of pre-trained language models for researchers in NLP, but it is incremental as it builds on existing analysis methods.
The study analyzed whether attention heads in BERT and RoBERTa learn constituency grammar, finding that some heads induce grammar types better than baselines, with fine-tuning on sentence meaning similarity tasks decreasing this ability by an average amount, while natural language inference tasks increased it.
With the success of pre-trained language models in recent years, more and more researchers focus on opening the "black box" of these models. Following this interest, we carry out a qualitative and quantitative analysis of constituency grammar in attention heads of BERT and RoBERTa. We employ the syntactic distance method to extract implicit constituency grammar from the attention weights of each head. Our results show that there exist heads that can induce some grammar types much better than baselines, suggesting that some heads act as a proxy for constituency grammar. We also analyze how attention heads' constituency grammar inducing (CGI) ability changes after fine-tuning with two kinds of tasks, including sentence meaning similarity (SMS) tasks and natural language inference (NLI) tasks. Our results suggest that SMS tasks decrease the average CGI ability of upper layers, while NLI tasks increase it. Lastly, we investigate the connections between CGI ability and natural language understanding ability on QQP and MNLI tasks.