Dohyeon Kim
Author directoryOther people with similar names: Dohyeon Kim
2026
SATS : Scenario-Anchored Topology-Inspired Structural Scoring
Dohyeon Kim | Jihoon Yang
Findings of the Association for Computational Linguistics: EMNLP 2026
Dohyeon Kim | Jihoon Yang
Findings of the Association for Computational Linguistics: EMNLP 2026
Sentence-level structural reasoning, reflected in relational connectivity patterns induced by attention distributions, plays an important role in grammatical acceptability judgment and figurative expression understanding. However, most existing scoring approaches rely primarily on token-level probability signals or sentence-level similarity, which do not explicitly capture structural relationships between tokens. We propose Scenario-Anchored Topology-Inspired Structural Scoring (SATS), a structure-aware comparison method that evaluates coherence between a scenario and multiple candidate interpretations using connectivity patterns derived from attention matrices. Unlike prior topology-inspired approaches that focus on pairwise comparisons, SATS directly supports multi-option reasoning while maintaining low computational cost. Experiments on BLiMP, FLUTE, and Pragmatics show that SATS improves encoder-based model performance in many settings and can approach the accuracy of much larger language models. These results suggest that explicit structural comparison can provide a useful complementary signal for the evaluated language understanding tasks.
Diagnosing Prompt Sensitivity and Interaction Instability in LLM-Based Humor Judgment
Dohyeon Kim | Jihoon Yang
Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
Dohyeon Kim | Jihoon Yang
Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
Humor judgment is subjective and often depends on ambiguity, expectation violation, and contextual reinterpretation, so the reliability of an LLM judge cannot be assessed from final detection performance alone. Existing LLM-based humor research and LLM-as-a-Judge studies largely focus on final performance, generation quality, or agreement with human evaluation, leaving prompt sensitivity and rebuttal robustness in subjective judgment tasks relatively underexplored. This paper analyzes how prompt design, structured interpretation, and user rebuttals shape LLM-based humor judgment. In zero-shot sentence-only judgments, models with similar F1 scores show substantially different FPR–FNR values, shifting between over-accepting and under-predicting humor depending on prompt formulation. Few-shot examples partially reduce these directional fluctuations, but their effects remain model-dependent. We further propose the Incongruity-Resolution Judgment framework, which compares a humor-supporting interpretation with a counter interpretation before producing a final judgment. Rather than uniformly improving performance, this framework serves as a diagnostic structure for examining how competing interpretations reshape model judgment behavior. Finally, explicit user rebuttals show that high initial performance does not guarantee robustness, highlighting the need to evaluate judgment persistence, directional error patterns, and interaction-level judgment trajectories.
2025
Enhancing Coreference Resolution with LLM-driven Data Augmentation and Adversarial Filtering
Dohyeon Kim | Gayeon Jung | Jeongseon Cho | Jihoon Yang
Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics
Dohyeon Kim | Gayeon Jung | Jeongseon Cho | Jihoon Yang
Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics
Coreference resolution is a fundamental task in natural language processing that involves linking different references to the same entity within a text. However, existing models often struggle to reliably identify referential relationships in contexts with extensive length or complex modifiers. This study proposes a data augmentation technique adding adjective phrases and employing a prompt-based adversarial filtering pipeline to address these challenges. Specifically, we generated and inserted contextually appropriate adjective phrases through the interaction between GPT-4o-mini based Few-shot Prompting and a Discriminative Language Model. The grammatical and semantic consistency of these phrases was validated via human evaluation and inter-annotator agreement (IAA) procedures. The generated synthetic dataset was integrated with existing data, leading to enhanced model performance. On the LitBank dataset, the CoNLL-F1 score increased by up to 1.7%, while the synthetic dataset improved linguistic diversity and the complexity of referential structures. The proposed pipeline represents a significant step towards developing coreference resolution models capable of better capturing linguistic variety and demonstrating robustness under challenging conditions.
Search