LinkedIn | Scholar | Vitae
zhexiong [at] cs.pitt.edu

Quote: The most beautiful thing we can experience is the mysterious. It is the source of all true art and science. -- Albert Einstein


Hi, I am a Ph.D. candidate of Computer Science in the Department of Computer Science at University of Pittsburgh. I work with Prof. Diane Litman in PETAL Lab. I received a master of science degree in Computer Science at Emory University and a bachelor degree in Computer Science and Technology at Sichuan University. Prior to my Ph.D. adventure, I was blessed to work with Prof. Jinho Choi in EmoryNLP Lab.

I am passionate about using natural language processing (NLP) and large language models (LLMs) to build writing applications (e.g., automated writing evaluation, revision simulation with LLM agents). I am also interested in multimodal learning (e.g., visual question answering, multimodal argument mining) and interpretable machine learning (e.g., neuro-symbolic reasoning, adaptive meta-learning). I focus on using NLP to address challenges in education, science, and technology. Please feel free to email me at zhexiong [at] cs.pitt.edu if you are interested in a discussion or collaboration!

News


Selected Projects

Efficient Layer-wise LLM Fine-tuning for Revision Intention Prediction
This project investigates a dynamic approach to fine-tuning large language models (LLMs) by selectively updating important layers while freezing redundant ones, which achieves fast fine-tuning convergence and requires low GPU allocations. It is the first work to introduce efficient layer-wise LLM fine-tuning in the revision community.
[video] [paper] [slides] [poster] [code]
A System to Assess Student Essay Revisions and Provide Formative Feedback
This project presents eRevise+RF, an enhanced automated writing evaluation (AWE) system, to assess argumentative essay revisions and provide formative feedback on both evidence usage and revision success in response to feedback. It shows promise of using NLP to scaffold young students in writing and revising argumentative essays.
[video] [paper] [poster] [code] [system] [webpage] [news]
Predicting the Quality of Revisions in Argumentative Writing
This project studies the relationship between argument contexts (ACs) and argument revisions (ARs) in argumentative writing. It leverages Chain-of-Thought prompts to facilitate ChatGPT to generate ACs for distinguishing between successful and unsuccessful ARs. It opens a promising avenue in revision research.
[video] [paper] [slides] [poster] [code]
Task-Adaptive Meta-Learning Framework for Advancing Spatial Generalizability
This project proposes a model-agnostic meta-learning framework that integrates geographically heterogeneous data into region-sensitive meta tasks using an easy-to-hard task hierarchy. It enables models to adapt to numerous heterogeneous tasks, significantly improving their generalizability.
[video] [paper] [slides] [poster] [code]
ImageArg: A Multi-modal Tweet Dataset for Image Persuasiveness Mining
This project develops a multimodal dataset featuring annotation on the persuasiveness of tweet images. The annotation is based on a persuasion taxonomy developed to explore image functionalities and the means of persuasion. The ImageArg wins Best Paper Award in the 9th ArgMining Workshop.
[video] [paper] [slides] [code] [shared task] [news]
Graph-Symbolic VQA with Rich Visual Estimators and No QA Labels
This project proposes a graph-symbolic framework that mimics human reasoning through a three-stage graph-based approach: image graph construction, question graph construction, and answering based on symbols in the two graphs. It requires no question-answer labels and can be applied to new domains without data annotation.
[video] [paper] [slides] [poster]
Reviving Literary Heritage Using RaspberryPi
This project presents a compact retriever system built on RaspberryPi that retrieves and prints archived poems. With a simple press of a button on the RaspberryPi board, users can effortlessly access a carefully curated poem. It showcases a seamless integration of art and science to reconnect audiences with literary heritage.
[news]

Selected Publication


Experience


Education


Teaching


Student Supervision


Academic Service


Leadership and Community Engagement


Selected Awards


Miscellaneous


Contact