About Me
I am currently a postdoctoral researcher at the Department of Computer Science, University of California, Los Angeles (UCLA), hosted by Prof. Nanyun (Violet) Peng and Prof. Kai-Wei Chang. I received my Ph.D. degree from the University of Science and Technology of China in June 2022, under the supervision of Prof. Zhen-Hua Ling. I have previously interned at Microsoft (2020-2021) and visited Queen’s University (2019-2020). I fortunately received the Best Paper Honorable Mention Award of ACL 2023 and Best Paper Award of DialDoc@ACL 2022.
My main research interests lie within deep learning for natural language processing, and the first-author papers during my Ph.D. study particularly focused on dialogue systems and retrieval-based language models:
- Retrieval-based dialogue systems: Interactiva Matching Network (IMN, CIKM 2019), Utterance-to-Utterance Interactiva Matching Network (U2U-IMN, TASLP 2020), and Speaker-Aware BERT (SA-BERT, CIKM 2020).
- Knowledge/Persona-grounded conversation: Dually Interactive Matching Network (DIM, EMNLP 2019), Filtering before Iteratively REferring (FIRE, Findings of EMNLP 2020), investigating the role of partners (SIGIR 2021), and detecting speaker personas (SPD, EMNLP 2021).
- Multi-party conversation: Pre-trained multi-party language model (MPC-BERT, ACL 2021), Heterogeneous graphical conversation model (HeterMPC, ACL 2022), a survey on multi-party conversation (IJCAI 2022), Graph-Induced Fine-Tuning (GIFT, ACL 2023), and Maximizing Addressee Deduction Expectation (MADNet, EMNLP 2023), which are systematically organized in our IJCNLP-AACL 2023 Tutorial.
These days, I am also interested in the techniques that can potentially democratize the training and deployment of large language models (LLMs). A list of topics that I am actively working on:
- Retrieval-Augmented LLMs: To augment the capabilities of given LLMs and scale down the size of LLMs via retrieval from knowledge indexes. Recent work: (1) Corrective Retrieval Augmented Generation (CRAG), (2) SyncCheck for Trustworthy RAG, (3) Sparse Context Selection for Accelerating RAG Inference (Sparse-RAG), and (4) TExt Generation via Task-specific and Open-world Knowledge (TEGTOK, Findings of ACL 2022).
- Model Editing: To reduce the cost of upgrading LLMs and enable data-efficient alterations to the behavior of LLMs, specifically within a designated realm of interest, while ensure no adverse impact on other inputs. Recent work: (1) Model Editing Harms General Abilities of LLMs, (2) Neighboring Perturbations of Knowledge Editing on LLMs (PEAK, ICML 2024), (3) Bidirectional Language Model Editing, and (4) Perturbation-Restrained Sequential Model Editing.
If you are interested in talking about research and life with me, please feel free to reach out.