【Team】Explainable AI Team
【Date】2026/September/3(Thursday) 10:00-11:00(JST)
【Speaker】Qingcheng Zeng, Northwestern University
Title: Reliable Embedding Systems for Real-World Search and Discovery
Abstract: Text embeddings have become a basic interface between language and computation,
powering search and large-scale text analysis. Yet real-world information needs rarely arrive
in the clean, uniform form assumed by standard benchmarks: people move across languages,
express nuanced preferences, and interpret meaning through domain and social context. In
this talk, I argue that embedding systems can support reliable search and discovery, but doing
so requires more than larger models and generic semantic similarity. Drawing on my work, I
show how mixed-language queries expose hidden representational failures, how targeted
training can make retrieval respond to user-defined criteria, and how context-sensitive
representations can turn embeddings into interpretable instruments for studying social
meaning. Together, these studies point toward adaptive embedding systems that account for
language, intent, and domain knowledge as information needs evolve. The broader goal is to
build systems that help people find relevant evidence and derive defensible insights from
complex collections of text.
