Week of April 3
Language Universals Workshop
Najoung Kim (Boston University)
Title: Compositional Linguistic Generalization in Artificial Neural Networks: Taking Stock
Abstract: Compositionality is considered a central property of human language. One key benefit of compositionality is the generalization it enables---the production and comprehension of novel expressions analyzed as new compositions of familiar parts. Whether artificial neural networks can generalize in such a way has been part of a longstanding debate. In this talk, I will discuss the test for compositional linguistic generalization we developed in Kim and Linzen (2020), and attempt a nuanced answer to the question "Do contemporary artificial neural networks compositionally generalize?" by reviewing follow-up studies that have used this test. In short, models can achieve (behavioral) success if the syntactic structure of the generalization examples are shared with the training examples, but generalization to novel structures is difficult without additional structural scaffolds provided to the model. In addition to this general picture, I will discuss a common oversight in applying this test to pretrained models, namely the lack of control in lexical exposure.
Friday April 7 | 12:00-1:30 PM | Sever 214
Harvard LangCog
Jin Wang (HGSE)
Title: The neural basis of phonological awareness and its relation to reading skill in developing children from infancy to kindergarten to elementary school
Abstract: Reading is an essential skill for daily life and academic success. According to the connectionist model of reading, word recognition involves orthographic, phonological, and semantic processing, as well as the interactions among them. Language skill such as phonological processing, develops earlier than reading acquisition, and thus likely serves as a foundation for later reading development. We refer to this hypothesis as the scaffolding hypothesis. In addition, based on the connectionist model, the experience of learning to read changes the nature of spoken language processes, which we refer to as the sculpting hypothesis. The overarching goal of my dissertation was to use a language skill, phonological awareness, and its longitudinal relation to reading skill, to examine how the processes of scaffolding and sculpting unfold in elementary school children from 5- to 7- to 9-year-olds. The specific aims of my doctoral work were to determine whether the development of these processes depend on (1) the grain size of the representations including larger units like rhymes/rimes versus smaller units like phonemes/letters, (2) the mapping from phonology to orthography resulting in automatic orthographic activation during spoken language processing, and (3) the nature of phonological processing involving representations versus access to those representations. As is predicted by the connectionist model that developmental effects should be earlier for smaller compared to larger grain sizes, due to greater computational demands of the latter, we found that the scaffolding effects occurred at small grain sizes in younger children but at larger grain sizes in older children. In addition, we found that in younger children, better reading led to an increase of the mapping from phonology to orthography and that automatic orthographic activation during auditory processing scaffolded later reading acquisition, consistent with the connectionist model. Finally, we found that phonological representation played a role in the scaffolding and sculpting effects earlier than phonological access, suggesting that a critical component is missing in the connectionist model, which only includes phonological representations but no control systems to account for the role of phonological access. After understanding the specific roles of phonological awareness in reading development from preschool to elementary school years, an interesting question arises for my post-doctoral work: that is, how do the individual differences of phonological awareness skills emerge? Given that phonological awareness is closely related to children's exposure to speech, currently, we aim to examine whether the neural responses to speech processing in infants can serve as an early predictor for later phonological awareness, language, and/or reading related skills at preschool. Addressing this question will have implications for early identification, intervention, and education for children who have reading difficulties.
Tuesday April 4 | 5:30-7:00 PM | William James Hall #1550
Fieldwork Lab
Harvard Radcliffe Fellow Roberto Zariquiey will give a talk at the fieldwork lab meeting this week.
Title: Doing in fieldwork in Peruvian Amazon: reflections on language documentation, linguistic typology and language revitalization
Monday April 3 | 4:15-5:15 | Room 007 (90 Mt Auburn St)
Meaning and Modality Lab
Anna Lim (BU) will give a talk at the Meaning and Modality Lab meeting this week.
Thursday April 6 | 1:30-3:00 PM | Room 420 (2 Arrow St)
Visitors at Harvard Linguistics
We are happy to announce the arrival of Visiting Scholar Aqil Shamshy. He was an Associate Professor at Northwest Minzu University in Lanzhou, China. Aqil will be with us through the end of December 2023. Welcome, Aqil!