Mentoring
Mentoring is the part of teaching I care about most, and the part I most want to keep. It also runs in both directions: my mentees' questions have repeatedly sent projects somewhere I would not have taken them.
I hand over a real open question early rather than an exercise with a known answer, because the experience of not knowing whether something will work is the thing being taught, and it cannot be simulated.
I start from what a student is trying to get out of the project rather than from what I need done, then define a problem that serves both. With someone new to research I start concrete and bounded, because early traction is what carries a person through the stretch of a first project where nothing works. With someone who already has research experience I start open and let the framing be negotiated.
A week spent on an approach that fails is sometimes the most valuable week of a project and sometimes purely discouraging, and which one it is depends on the student and the stage. I check in on reasoning rather than deliverables — asking why an approach was chosen before saying anything about whether it will work, since a reasoned choice and a borrowed one look identical in a write-up and completely different under one question.
My work sits across NLP, HCI, and scientific domains, so much of what I teach students is how to move between fields whose conventions are never stated aloud: what counts as evidence, what a contribution looks like, which claims need a citation rather than an experiment. Those rules are invisible to people already inside a field and are most of the difficulty for everyone else.
Owned the figure-decomposition stage that produces the validated inventory of regions the generated narration is permitted to point at — the component that keeps grounding from hallucinating a location.
Developed the invariance checks that verify an edit changes the intended element while provably preserving the rest, which became central to how the system reasons about cascading edits.
I was her teaching assistant across four course offerings at IIT Kharagpur. She taught the same core ideas to first-year programmers, to graduate researchers, and to a national online audience — which is where I learned that the material is not the course, the audience is.
My PhD advisor, whose QANTA project runs on incremental question answering: a question is read aloud and players interrupt the moment they think they know the answer. What gets measured is not whether you can produce an answer at the end but how early, and on what evidence, you will commit. His group also puts undergraduates alongside graduate students from computer science, language science, and information science on the same systems.
At Microsoft Research India she showed me that care and rigor are not competing demands on a mentor. She was generous with attention and immovable about depth, and she made clear that taking a student seriously means holding their work to a real standard rather than protecting them from one. It is the model I try to follow.
If you are a student interested in NLP, human-centered evaluation, or AI for scientific communication, I am always glad to hear from you.