💬 click anything to find out more!
🙋 Hello! I'm a MS candidate at Stanford University studying Computer Science and Music.
🔬 I research NLP: the art of making computers understand and work with human language.
👩💻 In 2020-2021, I led Stanford's Chirpy Cardinal research team. We're competing in the Amazon Alexa Prize Challenge for a $500K prize.
⚒️ I've worked at DE Shaw Research and Amazon AWS AI. In Fall 2021, I will be at Google Research.
🎻 I'm a concert pianist and organist! Click here to see my videos!
Publications click any to find out more!
Alexa Prize Proceedings 2021
Fertility and Sterility, Jul 2021
JAMA Network Open, Jul 2021
Fertility and Sterility, May 2021
Neural, Neural Everywhere: Controlled Generation Meets Scaffolded, Structured Dialogueto appear, Proceedings of the 2021 Alexa Prize
We present the second iteration of Chirpy Cardinal, an open-domain dialogue agent developed for the Alexa Prize SGC4 competition. We focus on improving conversational flexibility, initiative, and coherence, introducing an array of new neural methods and make major improvements in entity linking, topical transitions, and latency. These components come together to create a refreshed Chirpy Cardinal that is able to initiate conversations filled with interesting facts, engaging topics, and heartfelt responses.
Align-Refine: Non-Autoregressive Speech Recognition via Iterative RealignmentNAACL 2021
Sequence-to-sequence Transformer speech recognition models achieve extremely high performance, but are difficult to deploy in the wild due to long decoding times. Non-autoregressive models are significantly faster, but at the cost of degraded performance. We introduced a new non-autoregressive method— Align-Refine—which significantly narrows this gap. Our method achieves SoTA-level performance at a 6x speedup over previous work.
Deep Subjecthood: Higher-Order Grammatical Features in Multilingual BERTEACL 2021
Previous work has demonstrated that Transformer encoders learn certain linguistic features, like part-of-speech and syntax. In this work, we demonstrate that this is also the case for deeper linguistic features like morphosyntactic alignment, a core attribute of all languages. We find that the predictive distributions of subjecthood classifiers reflect the morphosyntactic alignment of their training languages, demonstrating the influence of high-level grammatical features not manifested in any one input sentence.
Finding Universal Grammatical Relations in Multilingual BERTEACL 2021
We show that Multilingual BERT is able to learn a general syntactic structure applicable to a variety of natural languages. Additionally, we find evidence that mBERT learns cross-lingual syntactic categories like “subject” and “adverb”—categories that largely agree with traditional linguistic concepts of syntax! Our results imply that simply by reading a large amount of text, mBERT is able to represent syntax—something fundamental to understanding language—in a way that seems to apply across many of the languages it comprehends.
A ML algorithm can optimize the day of trigger to improve in vitro fertilization outcomesFertility and Sterility
Injecting hormones to trigger ovulation is one of the most important parts of IVF, but the impact of timing decisions has not been previously studied. We applied causal inference methods to optimize the timing, achieving strong results with a simple model!
Changing stimulation protocol on repeat conventional ovarian stimulation cycles does not lead to improved laboratory outcomesFertility and Sterility
Physicians often change stimulation protocols after an unsuccessful cycle. We demonstrate empirically that this is not beneficial; on the contrary, we found a minor, but statistically significant improvement in certain laboratory outcomes in the group where the same approach was used a second time
Development and Validation of an Artificial Intelligence System to Optimize Clinician Review of Patient RecordsJAMA Network Open, Jul 2021
PDF clinical records are inconsistent and hard to understand, so physician review is time-consuming and fraught with error. We developed a system which summarizes and presents these records in an easy-to-understand web interface. A study with first-time physician users demonstrated a 20% time speedup!