We recently published one of the earliest works on formally evaluating the impressive text rewriting capability of GPT-3.5 (davinci), in particular, for paraphrase generation [EMNLP’22a] and text simplification [ACL’23a] . Our new LENS metric [ACL’23a] is the first learnt automatic evaluation metric for text simplification, which, when used as objective in minimum Bayes risk decoding (MBR), also set the newest state-of-the-art of open-sourced generation models, on par of GPT-3.5 and GPT-4. We also work on i
We analyze monolingual and multilingual LLMs for fairness [arXiv’23c] , distillation [ACL’23c] , cost effectiveness [EMNLP’21] , readability assessments [arXiv’23d] , and any other strengths/weaknesses that may lead to further development of better, fairer, smaller models. We also develop effective methods, such as label projection [ACL-F’23] , for cross-lingual transfer learning. That is, with only English annotated data, we directly train multilingual language models that can perform tasks (e.g., entity r
Sarah Wiegreffe – Personal website