Correction with Backtracking Reduces Hallucination in Summarization

TL;DR

This paper tackles hallucinations in neural abstractive summarization by introducing Correction with Backtracking (CoBa), an inference-time method that detects non-grounded content via $p_{\theta}(s_t|\mathcal{C},\mathcal{S}_{<t})$ and token-to-context distance thresholds, then backtracks to re-generate prior tokens up to a limit $L$. CoBa does not require retraining and is compatible with most decoding schemes, offering two detection strategies—uncertainty-based and similarity-based—and a depth-first backtracking mitigation. Experiments on Flan-T5 XL and LLaMA across Newsroom, CNN/DM, and XSUM show CoBa and its variant CoBa-d improve faithfulness metrics such as AlignScore, FactCC, and BS-Fact, while preserving generation efficiency relative to heavier lookahead-based methods. The approach is complementary to existing training-time and generation-time techniques and has practical impact for deploying more faithful summarization systems without additional model training.

Abstract

Abstractive summarization aims at generating natural language summaries of a source document that are succinct while preserving the important elements. Despite recent advances, neural text summarization models are known to be susceptible to hallucinating (or more correctly confabulating), that is to produce summaries with details that are not grounded in the source document. In this paper, we introduce a simple yet efficient technique, CoBa, to reduce hallucination in abstractive summarization. The approach is based on two steps: hallucination detection and mitigation. We show that the former can be achieved through measuring simple statistics about conditional word probabilities and distance to context words. Further, we demonstrate that straight-forward backtracking is surprisingly effective at mitigation. We thoroughly evaluate the proposed method with prior art on three benchmark datasets for text summarization. The results show that CoBa is effective and efficient in reducing hallucination, and offers great adaptability and flexibility. Code can be found at https://github.com/zhenzhel/CoBa.

Figures (6)

• Figure 1: Schematic illustration of CoBa (using only token probability as the detection metric with threshold $0.2$). After the partial summary "I live", the token "in" has a higher probability than "with". However, "I live in" will pressure the model into hallucinating a place. We detect this because all the next tokens have a probability lower than our threshold $0.2$. Backtracking enables the model to find an alternative continuation that avoids hallucination down the line.
• Figure 2: Average token probability (top) and token-to-context distance (bottom) around the hallucination span. Token offset 0 stands for the token where hallucination starts, negative offsets stand for the tokens before hallucination and positive ones are for the hallucinated tokens. On average, the token which starts the hallucination has the lowest probability and is the furthest away from the context tokens compared to surrounding ones.
• Figure 3: AlignScore vs. Generation Time. Note that the the x-axis is in log scale. The curve for Lookahead represents doing lookahead every $k$ tokens for $k$ from 200 to 1. CoBa attains the highest AlignScore with more than 10x speedup.
• Figure 4: Ablation on the token confidence threshold for CoBa. High is better for all metrics. Most metrics saturate around threshold 0.2-0.25.
• Figure 5: Qualitative examples of greedy decoding vs. CoBa and CoBa-d. The hallucinated content is marked in red and the corrected details are marked in green. CoBa and CoBa-d correctly remove the hallucinated content by triggering backtracking at corresponding positions and generate summaries with more and faithful details.