PaperlayerView exact revision on arXiv

Source-linked AI summary

Improving Faithfulness in Abstractive Summarization with Contrast Candidate Generation and Selection

Sihao Chen, Fan Zhang, Kazoo Sone, Dan Roth

arXiv:2104.09061v1cs.CL

TL;DR

Faithfulness remains a challenge for neural abstractive summarizers, which can generate hallucinated content. This paper generates source-compatible summary alternatives and selects among them with a discriminative model. On XSum, it corrects extrinsic hallucinations, while mistakes can introduce a small fraction of intrinsic hallucinations.

  • Problem

    Neural summarization systems still produce intrinsic and extrinsic hallucinations, despite progress on automatic metrics, leaving faithfulness less improved.

    Source: §1 · p. 1

  • Method

    The method generates summary variants by replacing named entities or quantities with source-compatible alternatives, then uses a discriminative model to select a candidate.

    Source: §1 · p. 2, §2.1 · p. 2

  • Results

    On XSum, the method corrects extrinsic hallucinations but incurs a small fraction of intrinsic hallucinations on mistakes; corrected summaries significantly improve FEQA scores over originals.

    Source: §3.2 · p. 3, §6 · p. 5

  • Takeaways & Limitations

    The authors offer their analysis of hallucinations across neural summarization systems as insights for future work.

    Source: §1 · p. 2, §6 · p. 5

    Because the method focuses on entities and quantities, the authors say it is likely insufficient to solve the entire summary-faithfulness problem.

    Source: §4.3 · p. 5

Abstract

from arXiv · show

Despite significant progress in neural abstractive summarization, recent studies have shown that the current models are prone to generating summaries that are unfaithful to the original context. To address the issue, we study contrast candidate generation and selection as a model-agnostic post-processing technique to correct the extrinsic hallucinations (i.e. information not present in the source text) in unfaithful summaries. We learn a discriminative correction model by generating alternative candidate summaries where named entities and quantities in the generated summary are replaced with ones with compatible semantic types from the source document. This model is then used to select the best candidate as the final output summary. Our experiments and analysis across a number of neural summarization systems show that our proposed method is effective in identifying and correcting extrinsic hallucinations. We analyze the typical hallucination phenomenon by different types of neural summarization systems, in hope to provide insights for future work on the direction.

1 Introduction

Neural abstractive summarizers have improved fluency and salience, but faithfulness remains less addressed as current systems still hallucinate. The paper studies model-agnostic post-processing to correct extrinsic hallucinations.

1 Introduction

  • Neural and pretrained transformer methods have improved the fluency and salience of generated summaries.The introduction distinguishes these gains from the less-developed progress on faithfulness.Source: §1 · p. 1
  • Faithfulness has received less progress than automatic-metric performance, while current state-of-the-art models still produce intrinsic and extrinsic hallucinations.Intrinsic hallucinations contain information present in the source; extrinsic hallucinations contain information absent from it.Source: §1 · p. 1
  • The paper studies post-processing as a model-agnostic way to correct hallucinations in settings where much of the ground-truth summarization data is itself hallucinated.Its contrast-candidate method replaces named entities or quantities with source-compatible alternatives and selects a candidate summary.Source: §1 · p. 2
  • The authors report preliminary effectiveness on XSum in correcting unfaithful summaries with extrinsic hallucinations.They also validate the method across neural summarization systems trained on XSum and analyze their typical hallucination types.Source: §1 · p. 2

2 Contrast Candidate Generation & Selection

The method targets hallucinations in entities and quantities, motivated by their presence in XSum summaries. It generates source-compatible variants, then uses a discriminative model to rank and select a candidate.

2 Contrast Candidate Generation & Selection

  • Hallucinations in XSum ground-truth summaries motivate correcting entity and quantity errors with a model-agnostic post-processing method.The study frames correction as conceptually less complex than summarization, requiring less training data while inheriting the fluency of data-intensive systems.Source: §2 · p. 2, Table 2 · p. 2

2.1 Contrast Candidate Generation

  • For generation, the method replaces summary entities with different source-document entities carrying the same NER label, creating fluent but not necessarily faithful variants.A Stanza NER system extracts entities of different semantic types from the source and summary.Source: §2.1 · p. 2

2.2 Contrast Candidate Selection

  • Candidate selection uses a discriminative classifier trained on faithful ground-truth summaries and synthetic negatives made by replacing entities with same-type alternatives.A margin-ranking term encourages higher scores for positive than negative candidates.Source: §2.2 · pp. 2–3
  • At test time, the trained model scores the generated variants and original summary, and the highest-scoring candidate becomes the output.Source: §2.2 · p. 3

3 Experiments

The experiments test whether contrast-candidate selection preserves standard summary metrics while improving faithfulness. On corrected summaries, automatic and human evaluation indicate improved faithfulness, with intrinsic hallucinations arising from some mistaken replacements.

3 Experiments

  • The method changed 13.3% of BART-generated summaries, left 48.3% unchanged, and had no applicable replacement for 38.4%.The 38.4% either lacked a hallucinated entity or a source entity of compatible type.Source: §3 · p. 3

3.1 ROUGE and BERTSCORE Evaluation

  • ROUGE and BERTSCORE were slightly lower with correction, which the authors attribute to matching extrinsic hallucinations in reference summaries.Source: §3.1 · p. 3, Table 3 · p. 3

3.2 Faithfulness Evaluation

  • FEQA scores improved significantly for the 1,510 summaries corrected by the system.The reported significance test used a two-tailed paired t-test (p < 0.001).Source: §3.2 · p. 3, Table 3 · p. 3
  • Human evaluation found improved faithfulness among corrected summaries, but mistaken replacements also introduced intrinsic hallucinations.The table reports faithful, extrinsically hallucinated, and intrinsically hallucinated judgments for 95 corrected summaries.Source: Table 4 · p. 3, §3.2 · p. 4

4 Analysis and Discussion

Across four summarization systems, the selection model has consistently high precision, while overall recall varies with the systems’ hallucination types. The analysis also identifies mistaken replacements and the limits of entity-focused correction.

4.1 Identifying Hallucination Across Systems

  • The selection model achieves consistently high precision across four systems, while overall recall varies with their entity and quantity hallucination patterns.Its relative recall is high with respect to the share of hallucinations involving entities and quantities.Source: §4.1 · p. 4
  • Recall drops on pretrained models such as BERTS2S, possibly because entity and quantity hallucinations make up a smaller share of their outputs.Source: §4.1 · p. 4

4.2 Intrinsic vs. Extrinsic Hallucinations Trade-off

  • Wrong entity replacements can introduce intrinsic hallucinations, and the method may lack a correct replacement when none appears in the source.The authors report this issue most for pretrained BERTS2S and BART, potentially due to pretraining artifacts or priors.Source: §4.2 · p. 4

4.3 Entity Faithfulness ⫋Summary Faithfulness

  • Entity faithfulness alone does not guarantee summary faithfulness; entity-and-quantity correction likely misses claims needing commonsense reasoning or knowledge retrieval.The discussion notes that a statement may be verifiable in the real world independently of the source text.Source: §4.3 · pp. 4–5

5 Related Work

Prior work measures summary faithfulness through content analysis and question answering, while also exploring pre- and post-processing methods to improve it. This study focuses on a setting where much of the training data contains hallucinations.

5 Related Work

  • ROUGE and BERTScore correlate poorly with human judgments of faithfulness, motivating categorical content analysis and question-answering consistency tests.The cited approaches include Narayan et al. (2018b), Deutsch et al. (2020), and Durmus et al. (2020) for question answering.Source: §5 · p. 5
  • Prior correction strategies include entailment-based reranking and truncating potentially unfaithful training summaries; this study addresses a setting with substantial training-data hallucination.The passage also contrasts the study with Cao et al. (2020), which uses similar strategies on a dataset with a very small fraction of hallucinations.Source: §5 · p. 5

6 Conclusion

The paper studies post-hoc contrast candidate generation and selection to correct extrinsic hallucinations involving entities and quantities. On XSum, it reports correcting extrinsic hallucinations while introducing a small fraction of intrinsic hallucinations on mistakes.

6 Conclusion

  • The method targets extrinsic hallucinations in entities and quantities through contrast candidate generation and selection, even when the dataset has intrinsic and extrinsic hallucinations.The conclusion frames this as a post-hoc fix.Source: §6 · p. 5
  • On XSum, the method corrects extrinsic hallucinations but incurs a small fraction of intrinsic hallucinations on mistakes.The authors also provide detailed analysis and discussion of the method's capabilities and limitations.Source: §6 · p. 5

A Candidate Selection Model

The candidate-selection model uses pretrained BART-base with a classification layer to label summaries as faithful or hallucinated.

A Candidate Selection Model

  • The selector uses pretrained BART-base, a linear layer over the max-pooled embedding, and labels candidates FAITHFUL or HALLUCINATED.Experiments use r = 1e-5, margin γ = 0, and three training epochs.Source: A Candidate Selection Model · p. 7

B Complete ROUGE and BERTSCORE Results

Table 7 reports ROUGE and BERTSCORE F1 results for baseline BART-large summaries and post-processed summaries on the XSum test set.

B Complete ROUGE and BERTSCORE Results

  • Table 7 compares baseline BART-large and post-processed summaries on XSum test-set ROUGE-1, ROUGE-2, ROUGE-L, and BERTSCORE F1.The passage provides the table's scope but no result values.Source: Table 7 · p. 7

C Estimating Confidence Interval for Human Evaluation

The human-evaluation confidence interval is estimated by bootstrapping annotations for 95 instances. The adjudicated annotation serves as ground truth, and the adjusted mean and 95% interval are derived from 1,000 resamples.

C Estimating Confidence Interval for Human Evaluation

Loading 2104.09061v1…