Exploring BERT’s Sensitivity to Lexical Cues using Tests from Semantic Priming

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4625–4635 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 4625 Exploring BERT’s Sensitivity to Lexical Cues using Tests from Semantic Priming Kanishka Misra1 Allyson Ettinger2 Julia Taylor Rayz1 1Department of Computer and Information Technology, Purdue University 2Department of Linguistics, University of Chicago 1{kmisra, jtaylor1}@purdue.edu, 2aettinger@uchicago.edu Abstract Models trained to estimate word probabilities in context have become ubiquitous in natural language processing. How do these models use lexical cues in context to inform their word probabilities? To answer this question, we present a case study analyzing the pre-trained BERT model with tests informed by seman- tic priming. Using English lexical stimuli that show priming in humans, we find that BERT too shows “priming,” predicting a word with greater probability when the context includes a related word versus an unrelated one. This effect decreases as the amount of information provided by the context increases. Follow- up analysis shows BERT to be increasingly distracted by related prime words as context becomes more informative, assigning lower probabilities to related words. Our findings highlight the importance of considering con- textual constraint effects when studying word prediction in these models, and highlight pos- sible parallels with human processing. 1 Introduction The field of natural language processing (NLP) has recently seen a dramatic shift toward the use of lan- guage model (LM)-based pre-training (Howard and Ruder, 2018; Peters et al., 2018)—training based on estimating word probabilities in context—as a foundation for learning of a wide range of tasks. Leading this charge was the BERT model (Devlin et al., 2019), which is optimized in part to use con- text information to predict masked words. Because of the impressively strong performance of BERT and its successors (Yang et al., 2019; Liu et al., 2019; Clark et al., 2020), there has been increasing need for understanding how these types of mod- els work, and what linguistic properties LM-based pre-training confers upon them. In this paper, we focus on the question of how BERT uses individual lexical relations to inform word probabilities in context. For example, if a word like airplane is prepended to (1a), to what extent does this increase the model’s probability for the word pilot in the blank position in (1b)? (1) a. I want to become a . b. airplane. I want to become a . This question is particularly relevant because hu- man brains show a robust phenomenon of semantic priming (McNamara, 2005), in which the presence of a word such as “airplane” will give rise to faster reactions to a related word like “pilot”. We explore whether the same lexical relations that show prim- ing in humans will also be utilized by BERT to influence word predictions in context. Our analysis includes three experiments. First, we test BERT’s sensitivity to single-word lexical cues for word prediction in context, using word pairs that show priming in humans, and testing for influence of contextual constraint. We find clear priming in BERT, but this effect is …

Content truncated to 500 words