Wenyan Li
Biography
Hi! I am Wenyan, a PhD student at the CoAStaL NLP Group, University of Copenhagen. I’m supervised by Anders Søgaard.
Before this, I worked as a Senior NLP Researcher at Sensetime and Comcast AI Research Lab (mentored by Ferhan Ture).
I completed my MS at University of Maryland, College Park, where I worked with Prof. Jordan Boyd-Graber on Natural Language Processing.
In my free time, I enjoy painting, cooking, yoga, and table tennis :)
NEWS:
Interests
- Natural Language Processing
- Multimodal Learning
- Information Retrieval
- Speech Technology
Education
-
M.S. in EECS (Thesis Track), 2016-2018
University of Maryland, College Park
-
B.E. in Electrical Engineering, 2012-2016
Northwestern Polytechnical University
Selected Publications
(2024).
Words Worth a Thousand Pictures: Measuring and Understanding Perceptual Variability in Text-to-Image Generation.
In EMNLP 2024 (Outstanding Paper Award).
(2024).
FoodieQA: A Multimodal Dataset for Fine-Grained Understanding of Chinese Food Culture.
In EMNLP 2024.
(2024).
Understanding Retrieval Robustness for Retrieval-Augmented Image Captioning.
In ACL 2024 (oral, 8%).
(2024).
The Role of Data Curation in Image Captioning.
EACL 2024 (oral).
(2023).
MAP: Low-data Regime Multimodal Learning with Adapter-based Pre-training and Prompting.
In Learning with Small Data (LSD), 2023.
(2022).
Systems and Methods for Training Voice Query Models.
U.S. filed patent.
(2020).
An Attentive Recurrent Model for Incremental Prediction of Sentence-final Verbs.
In EMNLP findings 2020.
Experience
Senior NLP Researcher (Artificial General Intelligence team)
- Knowledge-enhanced QA and dialogue system * Multimodal and prompt learning
Senior Machine Learning Research Engineer (NLP)
- Designed an unsupervised auto-annotation system for voice queries with user behavioral modeling to automatically identify errors in speech recognition and NLP systems and suggest corrections
- Built an active learning pipeline with auto-labeled user transcriptions to improve ASR system for comcast X1, increasing system recognition accuracy by 9% (summarized the work into a conference paper as the first-author and filed a patent as the main inventor)
- Developed a context-based approach that discovered misclassified user queries in question answering systems by performing semantic search with Sentence-BERT
- Leveraged subword-level query representation and adversarial training in customer care dialogue system for misspelled user queries, which improved classification accuracy by 18% and increased user experience stability
- Mentored interns and new-hires on projects relevant to multi-task learning and query representation
Recent Posts
Certificates
Applied Text Mining in Python
Using Python for Research
Object Oriented Programming in Java