• Designed prompt suites and rubric-based evaluations for groundedness/hallucination, with failure-mode categories tied to discourse phenomena, such as attribution, consistency, and hedging.
• Developed annotation guidelines to assess socio-cultural bias and pragmatic harm, emphasizing appropriateness in Korean register and stance.
• Validated a Korean multimodal dialogue dataset by aligning utterance-level pragmatic functions, such as agreement, refusal, and hesitation, with nonverbal cue labels.

Key insight: many reliability issues are not isolated “wrong facts,” but emerge from how meaning is constructed across turns, including implicature, underspecification, and discourse coherence.

• Created semantics-informed risk probes for single-turn inputs, including cases where meaning depends on polysemy or indirect phrasing.
• Developed multi-turn scenarios to capture escalation patterns, such as justification, concretization, and actionable requests, and measured how risk accumulates over dialogue context.
• Analyzed evasion behaviors after refusals, such as repair, paraphrase, and euphemisms, and incorporated these patterns into testing protocols.

Key insight: risk is often discourse-driven. What looks harmless in isolation can become unsafe when reference resolution, implicature, and turn-by-turn intent shifts are taken into account.

• Built a linguistically grounded ambiguity taxonomy, including ellipsis, anaphora, spacing/noise, and informal variants, to standardize how ambiguous inputs are handled.
• Applied text normalization and preprocessing guided by the taxonomy to reduce interpretation noise and improve NLU robustness.
• Led annotation efforts to ensure sociolinguistic coverage, including honorific levels, informal slang, and code-mixing, so the system behaves consistently across styles.

Key insight: in Korean conversational products, performance depends not only on intent classification accuracy but also on pragmatic consistency under register shifts.

• Supported corpus collection and preprocessing pipelines, emphasizing consistent unit definitions, such as sentence and utterance, and annotation-ready formatting.
• Assisted in annotation design and QA routines to improve label clarity and reduce ambiguity in training data.
• Helped construct NLU datasets for chatbots with attention to semantic coverage and linguistic variation.

• Compared Direct, Translate-then-Summarize, Summarize-then-Translate, and Semantic pipeline configurations.
• Evaluated generated Chinese summaries using ROUGE, BERTScore, and OmniScore.
• Analyzed semantic representation errors to identify information bottlenecks, schema instability, and model-specific failure patterns.
• Found that multi-agent decomposition does not automatically improve performance, but can make model failures more inspectable.

Key insight: more complex agentic pipelines are not necessarily better, but structured intermediate representations can reveal where models lose or distort information.

• Designed pattern-based rules grounded in lexical semantics and compositionality, such as “costs rise” versus “stock prices rise.”
• Built a small, interpretable pipeline to extract attribute-predicate relations from financial text and analyzed common error patterns.
• Reflected on limitations and extensions, such as industry metadata and context windows, for more robust interpretation.
• Recognized with the Outstanding Undergraduate Thesis Award.