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

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 semantic 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. Followup 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 contextual constraint effects when studying word prediction in these models, and highlight possible parallels with human processing.

Sorting through the noise: Testing robustness of information processing in pre-trained language models

“…Some scattered work has explored more semantic types of attractors for testing LMs-in particular, there is work looking at whether presence of certain context words will prime corresponding targets in context. Such work has experimented with contextual factors like distance between prime and target (Kassner and Schütze, 2020), as well as contextual constraint (Misra et al, 2020). We build on this existing work with a more systematic exploration of impacts of different types of attractors, and with a more targeted goal of testing models' robustness in processing new facts from context.…”

John praised Mary because _he_? Implicit Causality Bias and Its Interaction with Explicit Cues in LMs

“…The analysis of semantic capabilities in LMs includes studies on negative polarity in LSTM LMs (Marvin and Linzen, 2018;Jumelet and Hupkes, 2018), reasoning based on higher-order linguistic skill (Talmor et al, 2019), arithmetic and compositional semantics (Staliūnaitė and Iacobacci, 2020), stereotypic tacit assumptions and lexical priming (Misra et al, 2020;Weir et al, 2020). Many of these studies look at recent PLMs and draw mixed conclusions about the level of semantics encoded by these models.…”

Uncovering Constraint-Based Behavior in Neural Models via Targeted Fine-Tuning

“…We made use of synthetically modified language data that accentuated, or weakened, evidence for certain linguistic processes. The goal of such modification in our work is quite similar both to work which attempts to remove targeted linguistic knowledge in model representations (e.g., Ravfogel et al, 2020;Elazar et al, 2021) investigates the representational space of models via priming (Prasad et al, 2019;Misra et al, 2020).…”