Papers by Mingi Kim
REZE: Representation Regularization for Domain-adaptive Text Embedding Pre-finetuning (2026.acl-long)
Copied to clipboard
| Challenge: | Recent text embedding models often introduce task-induced bias alongside domain knowledge, leading to performance degradation. |
| Approach: | They propose a representation regularization framework that explicitly controls representation shift during embedding pre-finetuning. |
| Outcome: | The proposed framework outperforms standard pre-finetuning and isotropy-oriented post-hoc regularization in most settings. |
KoCoSa: Korean Context-aware Sarcasm Detection Dataset (2024.lrec-main)
Copied to clipboard
| Challenge: | Sarcasm is a form of verbal irony where someone says the opposite of what they mean . misunderstanding this sarcasm may lead to fatal errors in dialogue systems . |
| Approach: | They propose a dataset for the Korean dialogue sarcasm detection task that uses 12.8K daily Korean dialogues and the labels on the last response. |
| Outcome: | The proposed system outperforms strong baselines like large language models in the Korean sarcasm detection task. |
Learning to See through Sound: From VggCaps to Multi2Cap for Richer Automated Audio Captioning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing AAC datasets suffer from short and simplistic captions, limiting expressiveness and semantic depth. |
| Approach: | They propose a multi-modal dataset that pairs audio with corresponding video and leverages large language models to generate rich, descriptive captions. |
| Outcome: | The proposed framework outperforms existing benchmarks in caption length, lexical diversity, and human-rated quality. |
CodeComplex: Dataset for Worst-Case Time Complexity Prediction (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Reasoning ability of large language models (LLMs) is crucial in complex decision-making tasks. |
| Approach: | They propose to use code time complexity prediction to assess LLMs' reasoning ability. |
| Outcome: | The proposed dataset comprises 4,900 Java codes and an equivalent number of Python codes. |