NLP  Natural Language Processing  I develop machine learning models that aim to understand and generate natural languages. We are currently witnessing the first generation of NLP tools that have been deployed at scale and are used by millions of people. However, the major component of this success is access to large amounts of training data which machines use to learn mappings between input and output. In many applications and languages, such annotations are not readily available, and are expensive and slow to collect. I am interested in designing algorithms that do not suffer from this annotation dependence. Specifically, we are developing deep learning models that can transfer annotations across domains and languages, that can learn from a few annotated examples by utilizing supplementary data sources, and that can take advantage of human-provided rationales to constrain model structure.     Oncology  Learning to Cure  Data collected about millions of cancer patients — their pathology slides, imaging, and other tests — contain answers to many open questions in oncology. Jointly with the MGH collaborators, we are developing algorithms that can learn from this data to improve models of disease progression, prevent over-treatment, and narrow down to the cure. On the NLP side, we are creating databases which record pertinent cancer features extracted from raw documents. On the computer vision side, we are working on deep learning models that compute personalized assessment from mammogram data focusing on early cancer detection.     Chemistry  ML Drug Discovery  Today, drug discovery involves practitioners with years of advanced training and is carried out in a trial-and-error, labor-intensive fashion. Our goal is to change a traditional pipeline. In a joint work with chemical engineers and chemists at MIT, we are working on deep learning methods for modeling chemical processes.