[Colloquium] Chao-Chun Hsu MS Presentation/Mar 1, 2023

[Megan Woodward]

This is an announcement of Chao-Chun Hsu's MS Presentation
===============================================
Candidate: Chao-Chun Hsu

Date: Wednesday, March 01, 2023

Time:  3 pm CST

Location: JCL 298

Remote Location: https://uchicago.zoom.us/j/94132202998?pwd=TFRpUXJrUW1KVW9udTJ2Zy81OVJ2UT09<https://www.google.com/url?q=https://uchicago.zoom.us/j/94132202998?pwd%3DTFRpUXJrUW1KVW9udTJ2Zy81OVJ2UT09&sa=D&source=calendar&ust=1678026497926401&usg=AOvVaw1pxA7LKZ_mNapnGc7nK396>  (ID: 94132202998, passcode: H2965Wmi)

M.S. Paper Title: On Predictability of Physician Notes

Abstract: The success of large language models not only leads to substantial progress in NLP systems, but also provides exciting opportunities to characterize the predictability of human texts.
In this work, we conduct the first large-scale systematic study of predictability in a wide range of text corpora, with an emphasis on physician notes.
We show that physician notes are much more predictable than other texts.
For example, the perplexity is only about 3 on physician notes, compared to 19 on Wikipedia and 12 on PubMed abstracts.
We further examine the predictability of different physicians and patients, characterize what information is hard to predict, and investigate the flow of predictability in texts.
Results show that notes for females and minorities groups have significantly lower perplexity compared to males and whites respectively.
Perplexity flow analysis further reveals which sections in the physician note drive the differences between groups.
Last but not least, we show that perplexity can serve as useful features for predicting clinical outcomes.

Advisors: Chenhao Tan

Committee Members: Chenhao Tan, Allyson Ettinger, and Ziad Obermeyer

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20230228/cff84ce2/attachment.html>