Papers by Gunhee Lee
Who Wrote this Code? Watermarking for Code Generation (2024.acl-long)
Copied to clipboard
| Challenge: | Existing methods to detect machine-generated text by embedding watermarks fail to function appropriately in code generation tasks due to the task’s nature of having low entropy. |
| Approach: | They propose a logit-modifying watermark method which enhances detection ability and mitigates code quality degeneration by removing low-entropy segments at generating and detecting watermarks. |
| Outcome: | The proposed method outperforms baseline methods in detecting machine-generated code text while preserving code quality. |
When Should Dense Retrievers Be Updated in Evolving Corpora? Detecting Out-of-Distribution Corpora Using GradNormIR (2025.findings-acl)
Copied to clipboard
Dayoon Ko, Jinyoung Kim, Sohyeon Kim, Jinhyuk Kim, Jaehoon Lee, Seonghak Song, Minyoung Lee, Gunhee Kim
| Challenge: | Dense retrievers encode text into embeddings to retrieve relevant documents . however, real-world corpora evolve, resulting in degraded retrieval performance . identifying when a dense retriever requires an update is critical for robust retrieval systems . |
| Approach: | They propose a task of predicting whether a corpus is out-of-distribution (OOD) relative to a dense retriever before indexing. |
| Outcome: | The proposed method detects whether a corpus is out-of-distribution (OOD) relative to a dense retriever before indexing. |
Is a Peeled Apple Still Red? Evaluating LLMs’ Ability for Conceptual Combination with Property Type (2025.naacl-long)
Copied to clipboard
| Challenge: | Conceptual combination is a cognitive process that merges basic concepts, enabling the creation of complex expressions. |
| Approach: | They propose to use a Conceptual Combination with Property Type dataset to evaluate LLMs for conceptual combination thoroughly. |
| Outcome: | The proposed method improves performance in all generative tasks. |
KoSBI: A Dataset for Mitigating Social Bias Risks Towards Safer Large Language Model Applications (2023.acl-industry)
Copied to clipboard
| Challenge: | Existing research and resources are not readily applicable in South Korea due to the differences in language and culture, both of which significantly affect the biases and targeted demographic groups. |
| Approach: | They propose a social bias dataset of 34k pairs of contexts and sentences in Korean covering 72 demographic groups in 15 categories. |
| Outcome: | The proposed dataset reduces social biases by 16.47%p on average for HyperClova (30B and 82B), and GPT-3. |
TimeChara: Evaluating Point-in-Time Character Hallucination of Role-Playing Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) can be used to simulate human behaviors, but point-in-time role-playing is a key component of fandom role-players. |
| Approach: | They propose a benchmark to evaluate point-in-time character hallucination in role-playing LLMs. |
| Outcome: | The proposed method reduces point-in-time character hallucinations effectively by decomposing reasoning steps and using narrative experts. |
Recursion of Thought: A Divide-and-Conquer Approach to Multi-Context Reasoning with Language Models (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to generate intermediate steps (CoT) are limited by the maximum context size due to various reasons. |
| Approach: | They propose a new inference framework that introduces several special tokens that the models can output to trigger context-related operations. |
| Outcome: | Extensive experiments with multiple architectures including GPT-3 show that the proposed framework significantly improves LMs’ inference capability. |
AudioCaps: Generating Captions for Audios in The Wild (N19-1)
Copied to clipboard
| Challenge: | a dataset of 46K audio clips with human-written text pairs is used to generate captions for audio . the task of translating a multimedia input source into natural language has been extensively studied over the past few years . |
| Approach: | They propose a top-down multi-scale encoder and aligned semantic attention for audio captioning. |
| Outcome: | The proposed captions are faithful to audio inputs and better than existing models. |
Think, Verbalize, then Speak: Bridging Complex Thoughts and Comprehensible Speech (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to decouple LLMs from spoken communication produce suboptimal results due to mismatches between optimal textual and verbal delivery. |
| Approach: | They propose a framework that decouples reasoning from spoken delivery to preserve the full reasoning capacity of LLMs. |
| Outcome: | The proposed framework preserves full reasoning capacity of large language models . it improves speech naturalness and conciseness with minimal impact on reasoning . |
See It All: Contextualized Late Aggregation for 3D Dense Captioning (2024.findings-acl)
Copied to clipboard
| Challenge: | Recent approaches to 3D dense captioning struggle with contradicting objectives . SIA generates captions with different region of interest and aggregates them afterwards . |
| Approach: | They propose a transformer pipeline that engages in 3D dense captioning with a new paradigm . SIA decodes two sets of queries—context query and instance query—and then aggregates them afterwards . |
| Outcome: | The proposed pipeline generates captions with different region of interest and aggregates them afterwards to enhance local-global sensitivity. |
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. |
Converting the Point of View of Messages Spoken to Virtual Assistants (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Using a voice message, virtual assistants extract the message and send it to the user’s contact, rather than properly converting it to “I love you.” |
| Approach: | They propose to take a voice message from one user, convert it to “I love you” and deliver it to its target user. |
| Outcome: | The proposed system can take a voice message from one user, convert the point of view of the message, and then deliver the result to its target user. |
Behavior-SD: Behaviorally Aware Spoken Dialogue Generation with Large Language Models (2025.naacl-long)
Copied to clipboard
| Challenge: | Spoken dialogues lack explicit modeling of behavior traits that are often overlooked in language models . et al.: our work opens new possibilities for developing behaviorally-aware dialogue systems . |
| Approach: | They propose a large-scale dataset with over 100K spoken dialogues (2,164 hours) they propose BeDLM, the first dialogue model capable of generating natural conversations . |
| Outcome: | The proposed model outperforms baseline models in generating natural dialogues . the proposed model can generate natural conversations conditioned on behavioral and narrative contexts - a key feature of spoken language models . |
Can Language Models Laugh at YouTube Short-form Videos? (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing datasets that focus on verbal cues and focus on short-form funny videos focus on focusing on verbs and visual cue. |
| Approach: | They curate a user-generated dataset of 10K multimodal funny videos from YouTube and annotate each video with timestamps and explanations for funny moments. |
| Outcome: | The proposed dataset improves the ability of large language models to understand humor. |