The MBB Lunch Series is free and open to the Harvard community.

What Do Neural Networks Learn About the Structure of Natural Language? The Filler---Gap Dependency as a Test Case.
Ethan Wilcox
Graduate Student, Linguistics

Recurrent Neural Networks (RNNs) are one type of neural model that has been able to achieve state-of-the-art scores on a variety of natural language tasks, including translation and language modeling (which is used in, for example, text prediction). However, the nature of the representations that these 'black boxes' learn is poorly understood, raising problems for accountability and controllability of many NLP systems. In this work, I assess what neural networks trained on a language modeling objective learn about syntax, or the structure of natural language. I use one particular natural language dependency -- the filler---gap dependency --- as a test case, providing evidence that the models tested are sensitive to the hierarchical relationships implicated in this dependency. Next, I turn to "island effects", or structural configurations that block the filler---gap dependency, which have been theorized to be unlearnable. I demonstrate that RNNs are able to learn some of the "island" constraints, although their behavior remains un-humanlike for others.

Neural Correlates of Component Reading Processes Among Adolescents with Diverse Language and Reading Backgrounds
Laura Mesite
Graduate Student, Education

In U.S. schools, both English Learners (EL) and children with Reading Disabilities (RD) often struggle with reading, the main channel to access learning. Although there is a strong body of behavioral and neuroimaging research on reading difficulties among monolingual children, no study to date has explored the neural correlates of L2 reading processes among ELs, who are in the process of becoming bilingual. The present study addresses this gap by exploring the neural correlates of processes underlying word reading in English among adolescents from diverse language and literacy backgrounds. Forty-seven adolescents aged 11-15 years (M_age = 13 years, 26 females) participated in the study and belong to one of these groups: 1) typically-developing (TD) English monolinguals (n=14), 2) English-speaking monolinguals with RD (n=13), 3) Spanish-English proficient bilinguals (BL, n=12), or 4) Spanish-English emergent bilinguals (EL, n=8). All participants completed a battery of cognitive and reading assessments as well as eye-tracking tasks during the behavioral session. Then, they returned for a fMRI session, which included a passive word-reading task adapted from Malins et al. (2016) designed to identify brain regions associated with word reading, lexicality, spelling-sound consistency, semantic knowledge, and spoken language processing in English. TD and BL groups performed similarly on behavioral measures of word reading in English, on average, and outperformed their EL and RD peers. Different behavioral profiles of reading difficulty emerged between the EL and RD groups, such that ELs exhibited relative weaknesses in phonological awareness, vocabulary, and reading comprehension, while participants with RD displayed challenges with pseudoword reading, rapid naming, and phonological memory. Overall, the groups exhibited similar patterns of neural activation to each of the component reading processes with few differences surviving correction for multiple comparisons, due to a lack of statistical power. Future research directions and implications for educations will be discussed.