🎉Excited to announce that I will be starting as an Assistant Professor of Computer Science at Rutgers University in Fall 2024! In the meantime, I will spend the next year at Johns Hopkins
@jhuclsp
as a postdoc with
@mdredze
!
I’m on the faculty job market! My research focuses on Responsible AI and investigates undiscovered and underrepresented ethical issues in NLP across fairness, trustworthiness, and safety. If you have relevant positions, let me know! Please share/RT!
✨I am hiring PhD students
@RutgersCS
for Fall 2024!✨
If you are interested in working with me on Responsible AI at Rutgers, you can read more at:
Also, if you are attending EMNLP and would like to chat about this, please reach out!
I will be at
#ACL2023NLP
next week! I am recruiting Ph.D. students in the next academic cycle, so if you are interested in working with me and would like to connect at the conference, please reach out! Looking forward to meeting everyone🎉
🎉Excited to announce that I will be starting as an Assistant Professor of Computer Science at Rutgers University in Fall 2024! In the meantime, I will spend the next year at Johns Hopkins
@jhuclsp
as a postdoc with
@mdredze
!
Excited to share my work from my Amazon internship last summer!
We compare biases across English, Spanish, Chinese, Hebrew, and Italian. We analyze biases for race, religion, nationality, and gender attributes on the downstream sentiment analysis task.
Sharon Levy is a welcome new addition to the team! She is interested in exploring the intersection of public health and AI, with expertise in natural language processing and Responsible AI.
She has a PhD in Computer Science from the University of California at Santa Barbara.
Excited to share my work from my Amazon internship last summer!
We compare biases across English, Spanish, Chinese, Hebrew, and Italian. We analyze biases for race, religion, nationality, and gender attributes on the downstream sentiment analysis task.
Excited to share two papers accepted to
#EMNLP2022
!
“SafeText: A Benchmark for Exploring Physical Safety in Language Models” w/ Emily Allaway, Melanie Subbiah,
@hmslydia
,
@DrDesmondPatton
, Kathleen McKeown, and
@WilliamWangNLP
accepted to main conference!
[1/2]
🎉 Dr.
@sharonlevy21
has successfully defended her Ph.D. on "Responsible AI via Responsible Large Language Models" at UCSB, & bagging the Fiona and Michael Goodchild Graduate Mentoring Award! 🏆 She's
#7
PhD from my lab, and
#5
to pursue a faculty job. Congrats Dr. Levy! 🚀🎓👏
Happy to share that our paper “HybriDialogue: An Information-Seeking Dialogue Dataset
Grounded on Tabular and Textual Data” was accepted to
#ACL2022
Findings! w/ Kai Nakamura,
@pascaltuan
,
@WenhuChen
, and
@WilliamWangNLP
Excited to share that our demo paper “Open-Domain Question-Answering for COVID-19 and Other Emergent Domains” was accepted to
#EMNLP2021
! w/ Kevin Mo,
@XiongWenhan
, and
@WilliamWangNLP
!
Presenting this at
#EMNLP2022
on Friday at 2PM in the Collaboratorium! Looking forward to meeting everyone this week🙂 Let me know if you would like to chat!
Huge thank you to my advisor
@WilliamWangNLP
and my wonderful colleagues, friends, and family for supporting me through this journey
@ucsbNLP
@ucsbcs
!
I’m recruiting Ph.D. students in the upcoming application cycle. If you’re interested in Responsible NLP, please reach out!✨
“Mitigating Covertly Unsafe Text within Natural Language Systems” w/
@alexmeigz
,
@aniyipyip
, Melanie Subbiah, Emily Allaway, John N. Judge,
@DrDesmondPatton
, Bruce Bimber, Kathleen McKeown and
@WilliamWangNLP
accepted to findings!
Preprints coming soon!
[2/2]
In our
#EMNLP
2022 paper, we introduce SafeText, a dataset to benchmark the evaluation of implicitly unsafe text, i.e., text that does not contain explicitly unsafe keywords (e.g. stab) and instead requires additional commonsense knowledge to recognize its physical harm.
For those of you who missed the
#EMNLP2021
demo session, check out our demo paper on open-domain question-answering for COVID-19 and other emergent domains! w/ Kevin Mo,
@XiongWenhan
, and
@WilliamWangNLP
Congratulations to our Ph.D. student Sharon Levy
@sharonlevy21
who has been selected to receive an Amazon Fellowship in Responsible AI and
#NLProc
! Sharon's
#ACL2021nlp
paper on memorization of conspiracy theories in large LMs:
Using SafeText, we evaluate several LLM’s for their generation and reasoning skills. We find that models have the capability to generate unsafe text and are not able to reason well between safe and unsafe advice within different scenarios.
Also, check out
@alexmeigz
and
@aniyipyip
's
#EMNLP2022
Findings position paper on mitigating this covertly unsafe text for a comprehensive overview of unsafe language!
To remove ambiguities when translating to grammatically gendered languages, we adapt existing sentiment analysis bias templates in English and explicitly define male/female subjects.
Our findings show that biases can manifest differently across various languages for individual attributes while also revealing similarities across languages as well such as favoritism of groups dominant in each language’s culture (e.g. majority religions and nationalities).
We also analyze the impact of multilingual training and how this affects biases! We break this down for pretraining and finetuning data. Our results show bias amplification from multilingual finetuning while multilingual pretraining does not have a consistent effect.
@fredahshi
Thank you! We did not try increasing data size (this was something we were limited to in certain languages) and our hypothesis was that each language/culture has its own societal biases, but it would definitely be interesting to see if biases differ and/or reduce with more data!