How to Probe Linguistic Knowledge and Bias

Abstract

While most animals have communication systems, few exhibit such high-level of complexityas human languages. One central question of linguistics and cognitive science is to explore what human cognitive underpinnings and learning mechanisms are necessary to master such a complex system. One influential approach is Chomskyan generativism which premises a universal set of linguistic parameters and principles that delimits a range of possible variations (UG) as human inductive biases (Chomsky, 1980). However, the language learnability issue (e.g., poverty of stimuli) is made without a specified theory of learning (Pater, 2019). The recent development in typological work (e.g., Dunn et al., 2011), psycholinguistic studies (e.g., Culbertson and Kirby, 2016), and modern neural networks (e.g., Manning et al., 2020a) start to challenge the proposal of a rich linguistic innate endowment. In this dissertation, I present studies that utilized different experimental paradigms to explore the two questions: 1. How to verify whether a typologically common linguistic pattern reflects a human cognitive bias or results from cognitive-external factors? 2. How to verify whether abstract biases or linguistic knowledge can emerge from the input data or needs to be pre-specified to generalize to unseen data? To address these questions, the current dissertation explores two lines of research.