Learning Sentence-Level Representations with Predictive Coding

Abstract: ABSTRACT The importance of oral language skills in reading comprehension is widely recognized in contemporary models. Building on this foundation, we propose the Linguistic Pathways Model. In this model, we illuminate mechanistic and developmental detail by which individual components of oral language support reading comprehension and embrace the multiple dimensions across which reading development plays out. This is the level of theoretical detail needed to inform instruction in the classroom that is most likely to propel children on strong trajectories of reading development. We illustrate the value of this model by focusing on syntactic skills—the ability to understand and manipulate sentence structure. We hypothesize two core pathways by which syntactic skills impact reading comprehension. In the syntax‐to‐lexicon pathway, syntactic skills influence how readers construct lexical representations, ultimately impacting reading comprehension. In the syntax‐to‐sentence pathway, syntactic skills affect reading comprehension by shaping how readers parse sentences and generate predictions about upcoming information. In each, we elaborate on mechanisms of these influences. We also detail the nature of developmental effects, including changes in relative reliance on skills over time and the temporal order of effects, and the interactions between the two. This work provides a new theoretical model for understanding the precise pathways through which individual oral language skills contribute to reading comprehension development, making predictions that are testable in classrooms.

TL;DR: A new technique called t-SNE that visualizes high-dimensional data by giving each datapoint a location in a two or three-dimensional map, a variation of Stochastic Neighbor Embedding that is much easier to optimize, and produces significantly better visualizations by reducing the tendency to crowd points together in the center of the map.