?
Postdoctoral Fellow
Research group | Language Technology Group (LTG) / Integreat
Main supervisor |?Andrey Kutuzov
Co-supervisor |?Egor Kostylev
Affiliation |?Department of Informatics, UiO
Contact |?meihant@ifi.uio.no
Short bio
I completed my PhD in 2022, in computer science at the Tsinghua University in China, with a thesis titled “Low-Resource Event Extraction”, leverage multimodal learning, knowledge distillation, and self-labeling methods to handle the data scarcity issue in event extraction.
Following that, I worked as a postdoctoral fellow on a collaborative project between the China National Cleaning Center and Tsinghua University in Beijing, focusing on anti-money laundering. Currently, I am a postdoctoral research fellow at UiO under the DSTrain MSCA fellowship, working with the LTG group in the Department of Informatics on LLM knowledge separation.
Research interests and hobbies
Broadly speaking, I am interested in information extraction and knowledge extraction, particularly in low-source and multi-model setting. My research primarily focuses on developing knowledge distillation, multi-model methods to extraction information from unstructured text. This often involves using few-shot and multi-learning technology to address the inherent long-tail issue.
In my DSTrain project, I am shifting to a smaller-scale context, focusing on medical ultrasound elastography. The aim is to adapt the full waveform inversion (FWI), commonly used in seismic imaging, to estimate the physical properties of tissues—critical for tumor detection—while incorporating neural networks to accelerate simulations and enhance data analysis.
In my DSTrain project, I am shifting to title "Separating Linguistics from Knowledge in Large Language Models". This project designs modular LLMs that explicitly separate linguistic competence from factual knowledge. A lightweight base model is trained to capture linguistic structure, while an external, dynamically updated knowledge graph stores and retrieves facts. By disentangling language and knowledge from the outset, the architecture mitigates hallucinations, enables efficient updates, and improves interpretability.
Outside of research, I enjoy watching fantasy drama, swimming, raising my pets, cooking, as well as traveling to discover new cultures.
DSTrain project
Separation of linguistic and factual knowledge in large language models
Current large language models (LLMs) jointly encode linguistic competence (e.g., phonology, morphology, syntax, discourse) and factual knowledge (e.g., entities, relations, world events) in largely shared parameters. This tight entanglement makes factual updates expensive and risky: even small knowledge modifications often require fine-tuning substantial parts of the model, which is compute-heavy and can unintentionally degrade language fluency or previously learned facts. Our project aims to separate language and knowledge in existing LLMs to make them more maintainable, updatable, and trustworthy.
We will develop methods that isolate a relatively stable “linguistic core” from a more modular “knowledge component,” enabling targeted edits, faster refresh cycles, and controlled knowledge injection without full retraining. We will develop evaluation protocols to quantify (i) how well linguistic capabilities are preserved, (ii) how precisely factual content can be updated, and (iii) how robust the separation remains under distribution shifts and conflicting evidence.
By enabling efficient, localized knowledge maintenance while keeping language ability intact, this work supports scalable deployment of LLMs in fast-changing, high-stakes settings where correctness and updateability are critical.
?
?
Publications
DSTrain publications
ORCID ID: 0000-0003-2679-5641
Previous publications
?