Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Sequence-Level Certainty Reduces Hallucination In Knowledge-Grounded Dialogue Generation

Yixin Wan, Fanyou Wu, Weijie Xu, Srinivasan H. Sengamedu

TL;DR

This work targets hallucination in knowledge-grounded dialogue generation by introducing sequence-level certainty as a unifying framework. It splits certainty into probabilistic and semantic components and proposes Certainty-based Response Ranking (CRR), with two variants: Probabilistic CRR (P-CRR) and Semantic CRR (S-CRR), to select faithful responses at decoding time. Empirical results show that higher sequence-level certainty correlates with reduced hallucination, and that CRR methods significantly improve faithfulness across multiple models, decoding methods, and datasets, with S-CRR often providing the strongest gains. The approach offers a practical, entailment-based decoding-time mitigation that can generalize to broader knowledge-grounded generation tasks and inspires future work on sequence-wide uncertainty metrics.

Abstract

In this work, we propose sequence-level certainty as a common theme over hallucination in Knowledge Grounded Dialogue Generation (KGDG). We explore the correlation between the level of hallucination in model responses and two types of sequence-level certainty: probabilistic certainty and semantic certainty. Empirical results reveal that higher levels of both types of certainty in model responses are correlated with lower levels of hallucination. We further propose Certainty-based Response Ranking (CRR), a decoding-time hallucination mitigation method that samples several response candidates, ranks them based on sequence-level certainty, and outputs the response with the highest certainty level. Aligning with our definitions of sequence-level certainty, we design 2 types of CRR approaches: Probabilistic CRR (P-CRR) and Semantic CRR (S-CRR). P-CRR ranks individually sampled model responses using the arithmetic mean log-probability of the entire sequence. S-CRR approaches certainty estimation from meaning-space, and ranks model response candidates based on their semantic certainty level as measured by an entailment-based Agreement Score (AS). Through extensive experiments across 3 KGDG datasets, 3 decoding methods, and 4 KGDG models, we validate the effectiveness of CRR for reducing hallucination in KGDG task.

Sequence-Level Certainty Reduces Hallucination In Knowledge-Grounded Dialogue Generation

TL;DR

This work targets hallucination in knowledge-grounded dialogue generation by introducing sequence-level certainty as a unifying framework. It splits certainty into probabilistic and semantic components and proposes Certainty-based Response Ranking (CRR), with two variants: Probabilistic CRR (P-CRR) and Semantic CRR (S-CRR), to select faithful responses at decoding time. Empirical results show that higher sequence-level certainty correlates with reduced hallucination, and that CRR methods significantly improve faithfulness across multiple models, decoding methods, and datasets, with S-CRR often providing the strongest gains. The approach offers a practical, entailment-based decoding-time mitigation that can generalize to broader knowledge-grounded generation tasks and inspires future work on sequence-wide uncertainty metrics.

Abstract

In this work, we propose sequence-level certainty as a common theme over hallucination in Knowledge Grounded Dialogue Generation (KGDG). We explore the correlation between the level of hallucination in model responses and two types of sequence-level certainty: probabilistic certainty and semantic certainty. Empirical results reveal that higher levels of both types of certainty in model responses are correlated with lower levels of hallucination. We further propose Certainty-based Response Ranking (CRR), a decoding-time hallucination mitigation method that samples several response candidates, ranks them based on sequence-level certainty, and outputs the response with the highest certainty level. Aligning with our definitions of sequence-level certainty, we design 2 types of CRR approaches: Probabilistic CRR (P-CRR) and Semantic CRR (S-CRR). P-CRR ranks individually sampled model responses using the arithmetic mean log-probability of the entire sequence. S-CRR approaches certainty estimation from meaning-space, and ranks model response candidates based on their semantic certainty level as measured by an entailment-based Agreement Score (AS). Through extensive experiments across 3 KGDG datasets, 3 decoding methods, and 4 KGDG models, we validate the effectiveness of CRR for reducing hallucination in KGDG task.
Paper Structure (25 sections, 1 figure, 7 tables)

This paper contains 25 sections, 1 figure, 7 tables.

Table of Contents

  1. Introduction
  2. Sequence-Level Certainty
  3. Probabilistic Certainty
  4. Semantic Certainty
  5. Certainty-based Response Ranking
  6. Method
  7. Experiments
  8. Experiment Setup
  9. Results
  10. Sequence-level Certainty and Hallucination
  11. Effectiveness of CRR for Hallucination Mitigation
  12. Background on Uncertainty and Hallucination
  13. Uncertainty Estimation in Generative Models
  14. On Uncertainty and Hallucination
  15. Conclusion
  16. ...and 10 more sections

Figures (1)

  • Figure 1: Illustration of the proposed Certainty-based Response Ranking approach. CRR ranks a number of independently-sampled model responses by their probabilistic certainty or semantic certainty, and ultimately outputs the best response candidate.