The speed of detection vs. segmentation from continuous sequences: Evidence for an anticipation mechanism for detection through a computational model

Abstract

To understand the latent structure of a language, one of the first steps in language learning is word segmentation. The rapid speed is an important feature of statistical segmentation, and exact quantifications would help us understand the underlying mechanism. In this study, we probe the speed of learning by using a novel experimental paradigm and compare them to results obtained through the traditional word segmentation paradigm. Using a novel target detection paradigm, we replicated and extended a study on when participants start to show learning effects. We successfully replicated a facilitation effect showing rapid learning, which showed that learners obtained statistical information following a single exposure. However, we also found a similar facilitation effect when the syllable sequence contained words that were uniform or mixed in length. Importantly, this contrasts with results from traditional word segmentation paradigms, where learning is significantly better in uniform-length sequences than in mixed-length sequences. Thus, even though the target detection paradigm showed robust effects, it may have required mechanisms different from those in word segmentation. To understand these mechanisms, we proposed both theoretical analyses and a computational model to simulate results from the target detection paradigm. We found that an anticipation mechanism could explain the data from target detection, and crucially, the anticipation mechanism can produce facilitation effects without performing segmentation. We discuss both the theoretical and empirical reasons why the target detection and word segmentation paradigm might engage different processes, and how these findings contribute to our understanding of statistical word segmentation.