Trusting Your Evidence: Hallucinate Less with Context-aware Decoding
Authors:
Preprint
Abstract
Language models (LMs) often struggle to pay enough attention to the input context, and generate texts that are unfaithful or contain hallucinations. To mitigate this issue, we present context-aware decoding (CAD), which follows a contrastive output distribution that amplifies the difference between the output probabilities when a model is used with and without context. Our experiments show that CAD, without additional training, significantly improves the faithfulness of different LM families, including OPT, GPT, LLaMA and FLAN-T5 for summarization tasks (e.g., 14.3% gain for LLaMA in factuality metrics). Furthermore, CAD is particularly effective in overriding a model's prior knowledge when it contradicts the provided context, leading to substantial improvements in tasks where resolving the knowledge conflict is essential.Keywords
All Related Versions
This publication has 18 references indexed in Scilit:
- Benchmarking Large Language Models for News SummarizationTransactions of the Association for Computational Linguistics, 2024
- CLIFF: Contrastive Learning for Improving Faithfulness and Factuality in Abstractive Summarization2021
- Entity-Based Knowledge Conflicts in Question Answering2021
- Improving Factual Consistency of Abstractive Summarization via Question Answering2021
- Annotating and Modeling Fine-grained Factuality in Summarization2021
- Understanding Factuality in Abstractive Summarization with FRANK: A Benchmark for Factuality Metrics2021
- Multi-Fact Correction in Abstractive Text Summarization2020
- On Faithfulness and Factuality in Abstractive Summarization2020
- Improved Natural Language Generation via Loss Truncation2020
- Natural Questions: A Benchmark for Question Answering ResearchTransactions of the Association for Computational Linguistics, 2019