A Hierarchical Utterance Generation Method for Mitigating Topic-Shift Hallucinations in Dialogue Systems

Q: What does this research mean for the field?

A hierarchical utterance generation approach effectively mitigates topic-shift hallucinations in dialogue systems, significantly improving response quality and contextual consistency across multiple large language model families compared to existing methods. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Yunjin Park; Juoh Sun; Harksoo Kim

doi:10.5626/ktcp.2026.32.8.352

August 25, 2026정보과학회논문지 : 컴퓨팅의 실제 및 레터

A Hierarchical Utterance Generation Method for Mitigating Topic-Shift Hallucinations in Dialogue Systems

View at Publisher

Ask AI

Bookmark

Share

Key Points

To develop and evaluate a hierarchical utterance generation method aimed at mitigating topic-shift hallucinations and maintaining conversational consistency in dialogue systems.
Implemented a hierarchical utterance generation framework and tested it across seven language models across Gemma, Llama, and Qwen families.
Evaluated response quality and contextual coherence using both GPT-4o-based automatic evaluation and human assessment with inter-annotator agreement metrics.
Automatic evaluation via GPT-4o showed higher Mean Reciprocal Rank (MRR) across all seven tested models, with improvements of 2.7% to 5.1% in Gemma, 22.2% to 38.1% in Llama, and 11.1% to 17.9% in Qwen2.5 families relative to prior baselines.
Human evaluations confirmed the highest average MRR for the proposed method across all models, achieving strong inter-annotator agreement (Kendall’s W = 0.825).

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

Comparative evaluation demonstrates improved contextual consistency across seven large language models, indicating effective reduction of topic-shift hallucinations in dialogue systems.

Abstract

No abstract yet. The journal hasn't published one we can show.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Park et al. (2026) studied this question.

synapsesocial.com/papers/6a8fe96710c91c1e92620f5c https://doi.org/10.5626/ktcp.2026.32.8.352

View at Publisher

Ask AI

Bookmark

Share

A Hierarchical Utterance Generation Method for Mitigating Topic-Shift Hallucinations in Dialogue Systems

Key Points

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

A Hierarchical Utterance Generation Method for Mitigating Topic-Shift Hallucinations in Dialogue Systems

Key Points

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading