First Paper Friday: Stella Chang

University of Michigan School of Information PhD student Stella Chang has published her first paper as a UMSI student. The paper explores whether language models, trained on words and phrases young children hear, can learn to understand tricky grammar rules. 

“Mind the Gap: How BabyLMs Learn Filler-Gap Dependencies” is published in the proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP 2025). The paper finds that while language models can pick up some patterns in grammar, they often struggle to grasp more complicated structures and apply rules in new situations.

The publication of a PhD student’s first paper is a big milestone in their career, initiating them into the scholarly community as producers of knowledge. UMSI supports their work as part of our mission to share knowledge. 

“This paper is an extension of a linguistic course I was in last semester,” Chang says. “Our findings suggest that such models remain limited in their ability to generalize these linguistic phenomena, in contrast to human children, who acquire them even with sparse language evidence.”

Chang is a second-year PhD student advised by UMSI professor VG Vinod Vydiswaran. She expects to graduate by 2029. Before starting her PhD, Chang earned her Master of Science in Information at UMSI after spending time working as a data analyst and consultant for many years. Her current research focuses on health-related natural language processing (NLP), particularly investigating community behaviors and developing approaches that promote social good in health contexts.

At UMSI, Chang loves the open-minded environment and conversations that stem from that. 

“I love how we're so diverse, and how we get to exchange with and learn from all very different people!” she says. 

Read “Mind the Gap: How BabyLMs Learn Filler-Gap Dependencies” in the Proceedings of 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP 2025). See the abstract below: 

Humans acquire syntactic constructions like filler-gap dependencies from limited and often noisy input. Can neural language models do the same? We investigate this question by evaluating GPT-2 models trained on child-directed input from the BabyLM Challenge. Our experiments focus on whether these “baby” language models acquire filler-gap dependencies, generalize across constructions, and respect structural constraints such as island effects. We apply a suite of syntactic constructions to four models trained on child language, including two base models (trained on 10M and 100M tokens) and two well-performing models from the BabyLM Challenge (ConcreteGPT and BabbleGPT). We evaluate model behavior using wh-licensing scores, flip tests, and grammaticality contrasts across four constructions. Results show that BabyLM-scale models partially acquire filler-gap dependencies but often fail to generalize or fully capture island constraints.

RELATED

Learn more about Stella Chang by visiting her UMSI profile.  

Read about UMSI’s PhD in Information program!