Our lab works on topics in ML, NLP, and AI to create models, systems, and agents with intelligent behavior. Our roots are in computational semantics and large-scale data – automated understanding at the corpora level – but we now work at all levels of the stack, from ML fundamentals to cognitive science.
Large models are the backbone of systems with intelligent behavior. We seek to improve and understand these models, with work on understanding datasets for LLMs (Data Portraits; Dated Data), extending and compressing long sequences (Nuggets; Toucan), and parameter efficient adaptation (ReAdapt).
A large part of our lab is focused on systems that process information and make decisions - either autonomously or in service of human information seekers. This ranges from models that can extract complex structured information from documents or images (IterX), to neuro-symbolic reasoning agents (Nellie; Treewise) that derive logical inferences from knowledge sources. We also work on topics in information retrieval - how can we determine whether information is useful to an automated agent or a human seeking to make a decision (FollowIR)?
Our lab often extends text-oriented reasoning and information seeking systems to other modalities: images and videos. We’ve built systems around event detection in videos (MultiVENT) and visual reasoning (TV-Trees).
It is often useful for these systems to work in multiple languages. We enable this through the creation of multilingual datasets (MegaWika; Multilingual Multiparty Coref) and cross-lingual transfer learning methods (SpanFinder).
Automated systems and agents should be safe, controllable, and truthful. We work on aligning and controlling language models to reduce hallucinations (According-To; Verifiable by Design). Another line of work investigates claims made by speakers: what does the speaker believe? Can we verify factual claims (DecompScore)? How can we handle uncertainty and calibration (UncertainNLI; Cross-lingual Calibration)?
The legal domain is a challenging testbed. We’ve worked on systems that can apply logical reasoning with tax laws (SARA) as well as fundamental NLP tasks like retrieval (CLERC), generation, and information extraction from legal corpora (Statutory IE).
We’ve recently begun studying Collaborative Intelligence in the classroom alongside colleagues at the JHU Center for Talented Youth and School of Education (AI as Co-Teacher).
Projects linked here are representative highlights - see Publications for the full list