Papers by Meeyoung Cha
Parallel Communities Across the Surface Web and the Dark Web (2025.findings-emnlp)
Copied to clipboard
Wenchao Dong, Megha Sundriyal, Seongchan Park, Jaehong Kim, Meeyoung Cha, Tanmoy Chakraborty, Wonjae Lee
| Challenge: | Sense of Community is a social motivation that is reflected in the social behavior of humans. |
| Approach: | They compile a large collection of parallel community datasets comprising over 7 million posts and comments from Reddit and 200,000 posts and comment from Dread, a dark web discussion forum, covering similar topics. |
| Outcome: | The results show that users on Reddit exhibit a stronger sense of community membership despite the dark web’s restricted accessibility. |
SGT: Securing Open-Source LLMs Against Malicious Fine-tuning via Safety Guidance Trigger (2026.acl-long)
Copied to clipboard
| Challenge: | open-weight large language models increase vulnerability to malicious fine-tuning . despite these advantages, open-source LLMs increase the potential for misuse . |
| Approach: | They propose a safety guide for open-weight large language models that guides fine-tuning toward the safety manifold to preserve alignment. |
| Outcome: | The proposed safety guidance trigger significantly improves robustness against malicious fine-tuning. |
Machine Behavior in Relational Moral Dilemmas: Moral Rightness, Predicted Human Behavior, and Model Decisions (2026.findings-acl)
Copied to clipboard
| Challenge: | Human moral judgment is context-dependent and changes based on interpersonal relationships. |
| Approach: | They characterize LLM behavior using the Whistleblower’s Dilemma . they find moral rightness remains consistently fairness-oriented . |
| Outcome: | The model decisions mirror moral rightness judgments, rather than their behavioral predictions. |
Unified Neural Topic Model via Contrastive Learning and Term Weighting (2023.eacl-main)
Copied to clipboard
| Challenge: | Recent techniques employ pretrained language models to improve topic quality. |
| Approach: | They propose a topic-based model that uses contrastive learning and term weighting to learn from a pretrained language model and discover influential terms from semantically coherent clusters. |
| Outcome: | The proposed model outperforms baselines across multiple topic coherence measures and can be used as an add-on to existing topic models and improves their performance. |
How Training Data Shapes the Use of Parametric and In-Context Knowledge in Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models leverage parametric and in-context knowledge in training . however, when these sources conflict, models arbitrate based on their internal confidence . |
| Approach: | They conduct controlled experiments using synthetic corpora to identify data properties that shape knowledge utilization. |
| Outcome: | The results show that the robust use of both knowledge sources is an emergent property . the results provide guidance for designing training data that supports the reliability of parametric and in-context knowledge in language models. |
The Fallacy of Echo Chambers: Analyzing the Political Slants of User-Generated News Comments in Korean Media (D19-55)
Copied to clipboard
| Challenge: | a new study analyzes the political slants of user comments on partisan media in Korea . the classifiers detect political leaning on conservative and liberal news outlets . |
| Approach: | They built a BERT-based classifier to detect political leaning of short comments . they found a high presence of conservative bias on conservative and liberal news outlets . |
| Outcome: | The proposed classifier produced an F1 score of 0.83 for 21.6K comments . it shows that more liberals comment on stories resonating with their political perspectives . |
SQuARe: A Large-Scale Dataset of Sensitive Questions and Acceptable Responses Created through Human-Machine Collaboration (2023.acl-long)
Copied to clipboard
Hwaran Lee, Seokhee Hong, Joonsuk Park, Takyoung Kim, Meeyoung Cha, Yejin Choi, Byoungpil Kim, Gunhee Kim, Eun-Ju Lee, Yong Lim, Alice Oh, Sangchul Park, Jung-Woo Ha
| Challenge: | Existing studies focus on coping with social harms that large language models pose . however, discussions on sensitive issues can become toxic even if the users are well-intentioned. |
| Approach: | They propose to use Korean dataset to test whether LLMs can generate offensive content and propagate prejudices. |
| Outcome: | The proposed dataset shows that acceptable response generation improves for HyperCLOVA and GPT-3. |
How Do Moral Emotions Shape Political Participation? A Cross-Cultural Analysis of Online Petitions Using Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Using cross-cultural online petition data, we examine how moral emotions influence user participation and political participation. |
| Approach: | They construct and share a moral emotion dataset comprising 50,000 petition sentences in Korean and English each, along with emotion labels annotated by a fine-tuned LLM. |
| Outcome: | The results show that moral emotions like other-suffering increase both forms of participation and help petitions go viral, while self-conscious have the opposite effect. |
Detecting Offensive Language in an Open Chatbot Platform (2024.lrec-main)
Copied to clipboard
Hyeonho Song, Jisu Hong, Chani Jung, Hyojin Chin, Mingi Shin, Yubin Choi, Junghoi Choi, Meeyoung Cha
| Challenge: | Existing efforts to automatically filter offensive language are vulnerable to users’ deliberate text manipulation tactics, such as misspelling words. |
| Approach: | They propose a contrastive learning model that embeds chat content with a random masking strategy to detect offensive language in open-domain chat conversations. |
| Outcome: | The proposed model outperforms existing models in detecting offensive language in open-domain chat conversations while also showing robustness against users’ deliberate text manipulation tactics when using offensive language. |
Detecting Contextomized Quotes in News Headlines by Contrastive Learning (2023.findings-eacl)
Copied to clipboard
| Challenge: | Existing research has found a link between the use of direct quotations and fake news. |
| Approach: | They propose a contrastive learning framework that allows embedding news quotes based on domain-driven positive and negative samples to identify such an editorial strategy. |
| Outcome: | The proposed framework maximizes the semantic similarity between the headline quote and the matched quote in the body text while minimizing similarity for other unmatched quotes in the same or other articles. |