Justin Chih-Yao Chen

I am a first-year Ph.D. student at The University of North Carolina at Chapel Hill, advised by Prof. Mohit Bansal. Prior to that, I earned my master’s degree from the Data Science Program at National Taiwan University, advised by Dr. Lun-Wei Ku and Prof. Hsin-Hsi Chen. My research interests lie generally in Natural Language Processing, LLM Reasoning, Representation Learning and Knowledge Graph.

When I was an undergraduate, I double majored in Computer Science and Statistics at National Cheng Kung University. I had the pleasure of working close with Prof. Cheng-Te Li, on topics related to Relation Extraction and Recommendation System.

ReConcile: Multi-Round Discussion Among Diverse LLMs

Improve LLM reasoning via multi-round discussions.

Enhanced NLI-based Fact-Checking System

Prevent the NLI model from learning short cuts by using HEX projection.

Fine-grained Emotion Classification (FEC)

Incorporate hyperbolic space and Parrott's emotion model from psychology.

Education

Ph.D. in Computer Science

UNC at Chapel Hill 2023-Present

M.S. in Data Science

National Taiwan University 2020-2022

B.S. in CS & B.B.A in Statistics

National Cheng Kung University 2016-2020

Work Experience

Institute of Information Science, Academia Sinica

Research Assistant NLPSA Lab • 2020-2022

Institute of Data Science, National Cheng Kung University

Research & Teaching Assistant NetAI Lab • 2017-2020

Codementor

Data Scientist Intern Summer 2019

News

Oct 7, 2023	We have a paper accepted to EMNLP 2023!
Sep 22, 2023	We have a new preprint on arxiv!
May 4, 2023	We have two papers accepted to ACL 2023.
Apr 1, 2023	I will be joining MURGe-Lab and begin my Ph.D. journey!
Feb 16, 2023	We have a paper accepted to IEEE ICASSP 2023.
Aug 18, 2022	Taught advenced natural language processing course at Taiwan AI Academy

Selected Publications

Preprint

ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs

Chen, Justin Chih-Yao, Saha, Swarnadeep, and Bansal, Mohit

2023

PDF
EMNLP’23

Location-Aware Visual Question Generation with Lightweight Models

Suwono, Nicholas Collin, Chen, Justin Chih-Yao, Hung, Tun Min, Huang, Ting-Hao Kenneth, Liao, I-Bin, Li, Yung-Hui, Ku, Lun-Wei, and Sun, Shao-Hua

2023

PDF
ACL’23

HonestBait: Forward References for Attractive but Faithful Headline Generation

In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) 2023

PDF BibTeX Video
ACL’23

Label-Aware Hyperbolic Embeddings for Fine-grained Emotion Classification

In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) 2023

PDF BibTeX Video
NAACL’21

ZS-BERT: Towards Zero-Shot Relation Extraction with Attribute Representation Learning

Chen, Chih-Yao, and Li, Cheng-Te

In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies 2021

Abs PDF BibTeX Video

While relation extraction is an essential task in knowledge acquisition and representation, and new-generated relations are common in the real world, less effort is made to predict unseen relations that cannot be observed at the training stage. In this paper, we formulate the zero-shot relation extraction problem by incorporating the text description of seen and unseen relations. We propose a novel multi-task learning model, Zero-Shot BERT (ZS-BERT), to directly predict unseen relations without hand-crafted attribute labeling and multiple pairwise classifications. Given training instances consisting of input sentences and the descriptions of their seen relations, ZS-BERT learns two functions that project sentences and relations into an embedding space by jointly minimizing the distances between them and classifying seen relations. By generating the embeddings of unseen relations and new-coming sentences based on such two functions, we use nearest neighbor search to obtain the prediction of unseen relations. Experiments conducted on two well-known datasets exhibit that ZS-BERT can outperform existing methods by at least 13.54% improvement on F1 score.