Основні поняття
In this work, the authors propose sequence-level certainty as a key factor in reducing hallucination in Knowledge Grounded Dialogue Generation (KGDG) models. They introduce Certainty-based Response Ranking (CRR) methods to mitigate model hallucination during decoding time.
Анотація
The content explores the correlation between sequence-level certainty and hallucination in KGDG models. It introduces two types of CRR approaches, Probabilistic CRR (P-CRR) and Semantic CRR (S-CRR), to reduce model hallucination effectively. Experimental results across various datasets, decoding methods, and models validate the effectiveness of these methods.
Key points:
- Proposal of sequence-level certainty as crucial for reducing hallucination in KGDG.
- Introduction of Certainty-based Response Ranking (CRR) methods.
- Validation through experiments on different datasets, decoding methods, and models.
Статистика
Sequence-level probabilistic certainty can be calculated as: 1/N * Σ(log p(si|s<i)), where p(si|s<i) is the conditional probability of generating token si given past tokens.
Agreement Score (AS) is used to measure semantic certainty by calculating the summed probability of semantic entailment between a response candidate and all others.
Цитати
"Higher levels of both types of certainty are correlated with lower levels of hallucination in model outputs."
"CRR significantly reduces hallucinations across all experiment settings."