Source-linked AI summary
Improving Factual Consistency of Abstractive Summarization via Question Answering
Feng Nan, Cicero Nogueira dos Santos, Henghui Zhu, Patrick Ng, Kathleen McKeown, Ramesh Nallapati, Dejiao Zhang, Zhiguo Wang, Andrew O. Arnold, Bing Xiang
TL;DR
Factually inconsistent summaries remain a serious problem for abstractive summarization, while existing metrics such as ROUGE do not adequately measure factual consistency. The paper introduces QUALS, an efficient question-answering-based metric, and CONSEQ, a contrastive training algorithm that maximizes factualness. Experiments show improved factual consistency according to automatic metrics and human evaluation, although FactCC is poorly suited to highly abstractive XSUM summaries and online training may use less representative samples.
Problem
Abstractive summarization models can generate factually inconsistent summaries, and ROUGE is inadequate for measuring factual consistency.
Method
The paper introduces QUALS as an efficient factual-consistency metric and CONSEQ as a contrastive learning algorithm that maximizes QUALS during training.
Results
The proposed method improves summarization factual consistency according to automatic metrics and side-by-side human evaluation.
Takeaways & Limitations
The approach incorporates factualness into seq2seq training and may extend non-differentiable evaluation metrics to broader seq2seq model training.
Takeaways & Limitations
FactCC is not well suited to highly abstractive XSUM summaries, and online training can produce less representative samples because construction is restricted to each batch.
Abstract
from arXiv · showhide
A commonly observed problem with the state-of-the art abstractive summarization models is that the generated summaries can be factually inconsistent with the input documents. The fact that automatic summarization may produce plausible-sounding yet inaccurate summaries is a major concern that limits its wide application. In this paper we present an approach to address factual consistency in summarization. We first propose an efficient automatic evaluation metric to measure factual consistency; next, we propose a novel learning algorithm that maximizes the proposed metric during model training. Through extensive experiments, we confirm that our method is effective in improving factual consistency and even overall quality of the summaries, as judged by both automatic metrics and human evaluation.
1 Introduction
Abstractive summarization has improved, but generated summaries can still hallucinate or contradict source documents. The paper proposes an efficient factual-consistency metric and a training method that maximizes factualness.
- About 30% of summaries generated by neural seq2seq models suffer from fact fabrication.
- ROUGE captures n-gram overlap but is inadequate for factual consistency, leaving effective automatic evaluation as a major training obstacle.
- The paper identifies an efficient factual-consistency metric and a training algorithm that maximizes factualness as the two required components.
- The proposed metric simplifies QAGS, achieves a 55x speedup, and correlates closely with QAGS.
2 An Efficient Metric for Factual Consistency
The paper introduces QUALS, an efficient QA-driven metric that evaluates factual consistency by comparing QAGen likelihoods for question–answer pairs under a summary and its source document. It simplifies QAGS while retaining its interpretable comparison between summary-derived and document-supported answers.
- 2.1 Background on QAGS: QAGS compares answers to the same generated questions in the summary and document, using word-level F1 overlap as its factual-consistency score.Its four-step pipeline uses answer extraction, question generation, question answering, and answer comparison.
- 2.1 Background on QAGS: QAGS is interpretable and correlates with human evaluation, but its three separate AE, QG, and QA models require substantial memory and computation during training.These costs motivate a metric that can be incorporated into summarization-model training.
- 2.2 QUALS (ours): QUALS uses one QAGen language model to generate question–answer pairs from a summary and evaluate those same pairs under the input document.Given a summary, QAGen jointly produces q and a; the document likelihood of the same pair is then evaluated separately.
- 2.2 QUALS (ours): QUALS generates 60 diverse question–answer pairs per summary, filters pairs whose answers are absent from the summary, and retains one pair when answers duplicate.The procedure uses diverse beam search with 60 beam groups and strength 0.5.
- 2.2 QUALS (ours): QUALS subtracts the summary log likelihood from the document log likelihood to baseline pair quality and normalize shifts caused by document style, vocabulary, and topic.This difference avoids treating intrinsically unlikely summary-generated pairs as inconsistent and reduces domain-related likelihood variation.
3 Improving Factual Consistency Through Contrastive Learning
The paper introduces CONSEQ, a contrastive learning algorithm that uses QUALS scores to train summarization models toward factually consistent outputs. It contrasts high- and low-scoring candidate summaries, while retaining ground-truth targets to improve stability relative to direct policy-gradient optimization.
- 3.1 CONSEQ: CONSEQ starts from an MLE-trained model, combines ground-truth and sampled sequences, and constructs S+ and S− according to their evaluation scores.The candidate pool is S = {y^(i), ŷ^(i)}, after which score-based positive and negative subsets are formed.
- 3.1 CONSEQ: Unlike REINFORCE, CONSEQ also uses ground-truth targets, which the authors report helps avoid instability that made REINFORCE summaries unintelligible after a few hundred updates.REINFORCE relies entirely on sampled sequences for its gradient updates.
- 3.2 CONSEQ + QUALS for Imposing Factual Consistency: The QUALS-based implementation selects high-scoring ground-truth summaries as positives and low-scoring sampled summaries as negatives before matching them by input document.Six top-k samples are generated per document, the lowest-scoring sample is selected, and p% extremes form the candidate pools.
- 3.1 CONSEQ: CONSEQ forms positive and negative summary sets from QUALS scores and trains the model to increase the probability of high-scoring sequences while decreasing the probability of low-scoring ones.Positive and negative candidates are paired for the same input document so the contrastive loss pushes generation toward the more consistent summary.
- 3.2 CONSEQ + QUALS for Imposing Factual Consistency: Online CONSEQ updates parameters more frequently using candidates from the latest model, but its within-batch construction can produce less representative positive and negative samples than offline training.The trade-off is between fresher candidates and broader pooling across input documents.
- 3.2 CONSEQ + QUALS for Imposing Factual Consistency: The method assumes that some ground-truth summaries are factually inconsistent, especially for XSUM, so only the top p% are used to form positive examples.The paper attributes this issue to imperfect data collection.
4 Experiments
Experiments on XSUM and CNN/DailyMail evaluate QUALS-CONSEQ against BART-large MLE using automatic metrics, ablations, human judgments, and qualitative examples. QUALS-CONSEQ improves factual consistency across QAGS and human evaluation while preserving ROUGE, though FactCC behaves poorly on highly abstractive XSUM summaries.
- Metric validation: QUALS correlates well and monotonically with QAGS on both XSUM and CNNDM, supporting its use for ranking summaries during contrastive training.The evaluation bins summaries by QUALS and compares average QAGS within each bin.
- Results: More than 4 QAGS points on XSUM and about 2 on CNNDM separate QUALS-CONSEQ from BART-large MLE, while ROUGE is not degraded.These test-set comparisons establish the main result across both datasets.
- Ablations: Replacing QUALS with ROUGE in contrastive training lowers QAGS below MLE and can also lower ROUGE, indicating that factual-consistency training requires an appropriate metric.The authors attribute the ROUGE variant’s weakness partly to focusing on difficult examples that may contain hallucinations in the ground-truth summaries.
- Ablations: The F1-based Q-F1-C variant slightly raises QAGS over Q-C on CNNDM but performs worse overall, consistent with log likelihood capturing semantically similar answers more softly than F1.The weighted Q-C-W variant improves factual consistency less than Q-C, while online Q-C-O tends to raise ROUGE but lower factual consistency.
- Human evaluation: Human evaluation finds QUALS-CONSEQ more factually consistent than BART-large MLE on both datasets, with an even larger improvement in informativeness.Annotator agreement is fair across factual consistency, informativeness, and grammatical correctness, with many disagreements involving ties.
- Qualitative analysis: QUALS-CONSEQ corrects entity hallucinations, relation errors, and coreference mistakes, but can miss errors involving modifier phrases.The qualitative analysis includes hallucinated flight or contract details, incorrect numbers, and an incorrect assignment of who locked a cockpit door.
5 Related work
The paper situates its approach among factual-consistency metrics, training methods, and post-processing techniques for abstractive summarization. It distinguishes QUALS as a general metric not limited to entities.
- Prior work measures factual consistency using QAGS, FactCC, or OpenIE and dependency parsing to match relations between documents and summaries.
- CONSEQ differs from unlikelihood training because its positive and negative loss terms are treated equally rather than weighting the negative term as regularization.
- Entity-level metrics and entity-focused improvements address a narrower scope than QUALS, which is designed to evaluate factual consistency beyond entities.
- Another approach post-processes summaries by swapping generated entities with same-type entities from the source, then ranking the resulting candidates for factual consistency.
6 Conclusion
The paper concludes that QUALS measures factual consistency efficiently and that CONSEQ maximizes QUALS during seq2seq training. Experiments show improved factual consistency for state-of-the-art summarization models under automatic and human evaluation.
- QUALS is introduced as an evaluation protocol, while CONSEQ is a contrastive learning algorithm for seq2seq models that maximizes QUALS during training.
- The proposed method significantly improves factual consistency in state-of-the-art summarization models according to automatic metrics and side-by-side human evaluation.
- The authors suggest CONSEQ may also help incorporate non-differentiable evaluation metrics into seq2seq model training more broadly.
A.1 Our implementation of QAGS
The QAGS implementation extracts answer candidates, generates questions for them, and uses a question-answering model to support factual-consistency evaluation.
- The implementation extracts named entities and noun phrases, removes uninformative stop words, and retains 10 answer candidates.
- For each answer candidate, a BART-large model trained on NewsQA generates questions using beam size 10, retaining the top 3 questions per answer.
- With 10 answer candidates and 3 questions per candidate, the implementation produces 30 questions per summary and answers them with ALBERT-xxlarge trained on SQuAD2.0.
A.2 Speed Estimate for QUALS and QAGS
QUALS is substantially faster than QAGS, making QA-based factual-consistency evaluation more practical for model training.
- 1.7 hours versus 93.6 hours on one NVIDIA V100 GPU, giving QUALS more than a 55x speedup over QAGS on the CNNDM test set.Using QAGS during training would require an estimated 14,033 GPU hours, or 584 days, on one GPU.
A.3 Additional Experimental Details of CONSEQ
CONSEQ constructs positive and negative summary sets by intersecting examples, yielding much smaller training subsets than the original datasets. QUALS and QAGS show the same monotonic relationship across percentile-bin analyses.
- Training-set construction: Intersecting positive and negative examples produced |S+| = |S−| = 6492 for XSUM and 48809 for CNNDM.The original training sets contained 203540 XSUM and 287112 CNNDM examples.
- Training-set construction: Using the intersection of examples produced better overall summary quality than using the constructed sets directly in CONSEQ.
- Metric relationship: QUALS increases monotonically with QAGS across 100 percentile bins on both XSUM and CNNDM.The corresponding Figure 4 reports QAGS as a function of QUALS, with average QAGS increasing across QUALS bins.
A.4 QUALS Approximates QAGS
QUALS closely tracks QAGS on both XSUM and CNNDM, including across finer-grained bins. A worked example shows that QUALS distinguishes a hallucinated flight number from a factually supported airline name.
- Qualitative illustration: The model hallucinated flight number “4U 9525” because the input document described an AirAsia crash without mentioning that number.The number belonged to a Germanwings flight, not the AirAsia crash described in the document.
- Qualitative illustration: QUALS scored the inconsistent and consistent question-answer pairs −2.615 and −0.054, respectively.The hallucinated answer was much more likely under the summary than the document, whereas “AirAsia” was similarly likely under both.
- Additional materials: Additional examples cover generated summaries, QAGen question-answer pairs, and the human-evaluation interface and instructions.