ScheMatiQ: From Research Question to Structured Data

Abstract: 

Research projects across domains often revolve around structuring a corpus of documents through the lens of a guiding research question: a legal scholar may compare immigration rulings by judges appointed by different U.S. presidents, an NLP researcher may ask when Chain-of-Thought prompting helps, and a computational biologist may investigate protein nuclear export signals across lab protocols. Answering such questions requires identifying the relevant observation units, deciding which attributes matter, extracting values across heterogeneous documents, and grounding each value in evidence.
In this talk, I will present ScheMatiQ, a framework which turns a natural-language research question and a document collection into a structured, human editable, evidence-backed database. I will discuss why this is difficult for current LLM-based tools, and show how we tackle this challenge through observation-unit discovery, schema induction, structured extraction, and an interactive expert-in-the-loop workflow.

Bio:

Gabriel Stanovsky is an Associate Professor in the School of Computer Science and Engineering at the Hebrew University of Jerusalem where he heads the program for BSc. in AI, and a Research Scientist at Ai2. He develops models for multidisciplinary research, with applications in archaeology, law, medicine, and beyond. His research is supported by grants from the ISF, Google, and IBM, among others. His work has received awards at ACL, NAACL, and CoNLL and has been featured in Science and The New York Times. He is also a recipient of the Krill Prize for Excellence in Scientific Research.