Please note: This master’s thesis presentation will take place online.
Ruoxi Ning, Master’s candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Freda Shi
Experimental linguistics serves as a vital resource and evaluation criterion for natural language understanding in contemporary machine learning. While linguistically motivated experiments often use controlled settings to explore how models learn, they frequently overlook the role of training hyperparameters as a source of randomness.
In this study, by replicating three representative experiments (Chang and Bergen, 2022; Hu et al., 2020; Kuribayashi et al., 2024) with hyperparameters varied over a practical range, we show that modest hyperparameter changes can alter qualitative conclusions about models’ linguistic abilities or even reverse cross-model comparisons, while other findings remain stable across the explored space. These results suggest that some prior studies in this field may partly reflect optimization artifacts rather than the inductive bias of a model class. We argue that hyperparameter sensitivity is an under-examined factor that can meaningfully influence model behaviours, and thus call for future linguistically motivated experiments with neural models to attend to hyperparameter selection and the stability of the findings across the configuration space.