Postdoctoral Fellow
Research group | Language Technology Group (LTG)
Main supervisor | Yves Scherrer?
Co-supervisor |?-
Affiliation |?Department of Informatics, UiO
Contact |?shenbinq@ifi.uio.no
Short bio
I am a postdoctoral fellow at the Department of Informatics, University of Oslo, as part of the Marie Sk?odowska-Curie Actions (MSCA) programme DSTrain. My current research focuses on the evaluation and explainability of large language models (LLMs). I earned my PhD in Natural Language Processing (NLP) at the University of Surrey, UK, where I mainly worked on the evaluation of machine translation. During my doctoral studies, I also contributed to projects in diverse applied machine learning fields, including sentiment analysis, information retrieval, text‐to‐image generation, and video understanding.
Research interests and hobbies
I am mainly interested in applying deep learning methods to solve and explain language-related problems including but not limited to the text modality. Translation was the primary focus of my previous research. The investigation of model performance for neural machine translation and later LLMs on different language pairs led to my current postdoc project to explore the multilingual ability of LLMs and how cross-lingual information is represented in LLMs. While continuing my investigation on multilingual evaluation, I am shifting to explain the performance disparity for different languages by opening the black box of LLMs.
Apart from research, I enjoy running, hiking and photographing to record my life. In my spare time, I also love listening podcasts.
DSTrain project
Investigating the Multilingual and Cross-lingual Abilities of Large Language Models in Resource-varying Settings
Large Language Models (LLMs) have transformed natural language processing (NLP), yet remain skewed toward high-resource languages (e.g., English), leaving countless underrepresented languages disadvantaged. This project tackles that issue by systematically examining LLM performance across multilingual and cross-lingual tasks comparing with monolingual tasks. Objective 1 evaluates generation (translation, summarization), regression (translation quality scoring, sentiment intensity), and classification (sentiment label) tasks in both high- and low-resource languages, offering a nuanced view of LLM strengths and weaknesses. Objective 2 explores cross-lingual transfer abilities, assessing whether knowledge acquired in English or other resource-rich settings can effectively benefit low-resource languages or closely related language pairs (e.g., Swedish-Norwegian).
By featuring comprehensive task coverage, emphasis on linguistic inclusivity, and innovative transfer learning paradigms, the proposal surpasses existing evaluation approaches on LLMs. Achievability is supported by established NLP datasets, robust fine-tuning methods (adapters, quantization), and standard evaluation metrics (BLEU, ROUGE, correlation coefficients, F1, etc.). Baseline comparisons and iterative testing ensure transparency, reproducibility, and measurable progress.
Ultimately, this project aims to advance the field by strengthening multilingual LLM performance, refining cross-lingual transfer techniques, and promoting more equitable AI solutions. By bridging the digital language divide, empowering diverse communities, and informing broader applications of natural language technologies, these findings will support global linguistic inclusivity. The anticipated outcomes will not only offer practical benefits for multilingual NLP tools but also serve as a springboard for future innovations, ensuring that all language communities can participate fully in the AI revolution.
Publications
DSTrain publications
ORCID ID: 0000-0002-6313-3359.
Previous publications
?