Abstract
Most studies on abstractive summarization report ROUGE scores between system and reference summaries. However, we have a concern about the truthfulness of generated summaries: whether all facts of a generated summary are mentioned in the source text. This paper explores improving the truthfulness in headline generation on two popular datasets. Analyzing headlines generated by the state-of-the-art encoder-decoder model, we show that the model sometimes generates untruthful headlines. We conjecture that one of the reasons lies in untruthful supervision data used for training the model. In order to quantify the truthfulness of article-headline pairs, we consider the textual entailment of whether an article entails its headline. After confirming quite a few untruthful instances in the datasets, this study hypothesizes that removing untruthful instances from the supervision data may remedy the problem of the untruthful behaviors of the model. Building a binary classifier that predicts an entailment relation between an article and its headline, we filter out untruthful instances from the supervision data. Experimental results demonstrate that the headline generation model trained on filtered supervision data shows no clear difference in ROUGE scores but remarkable improvements in automatic and manual evaluations of the generated headlines.
Introduction
Abstractive text summarization has seen significant advancements through encoder-decoder architectures. However, these models frequently suffer from "hallucinations"—generating headlines that contain facts, entities, or temporal information not present in the source text. This lack of truthfulness undermines the utility of summarization systems in real-world applications.
The authors hypothesize that this behavior is a direct result of noisy training data. In benchmarks like English Gigaword, models are often trained to predict a headline using only the first sentence of an article. Because headlines frequently synthesize information from the entire article, the target headline often contains facts missing from the provided lead sentence. When models are forced to learn from these unentailed mappings, they learn to guess or hallucinate missing information.
Methodology
To address this, the researchers propose a data-cleaning pipeline based on textual entailment:
- Entailment Classification: Binary classifiers (RoBERTa for English, BERT for Japanese) were trained to determine if a source text logically entails its target headline.
- Data Filtering: Training pairs were aggressively filtered out if the source text did not entail the headline. This process removed approximately 29% of the English Gigaword dataset and 51% of the Japanese dataset.
- Self-Training (Pseudo-labeling): To prevent performance degradation from reduced data volume, the authors used a baseline model to generate truthful "pseudo-headlines" for the discarded source texts, which were then added back to the training pool.
- Model Training: A standard Transformer model was trained on this refined, augmented dataset.
The Truthfulness Metric
A key contribution of this work is the introduction of a "support score," defined as a recall-oriented ROUGE-1 score measuring word overlap between the generated headline and the source document rather than the reference headline. This metric serves as a proxy for factual consistency.
Analysis revealed that standard ROUGE scores are largely blind to truthfulness. As shown in the figure above, there is almost zero correlation (Pearson's r = 0.189) between ROUGE-1 and the support score. A high ROUGE score indicates similarity to a reference headline but does not guarantee that the generated content is actually supported by the source text.
Results
Experimental results on the Gigaword dataset demonstrate the effectiveness of the filtering approach:
- Entailment Rate: The baseline model's headlines were entailed by the source text 85.78% of the time. The model trained on filtered and pseudo-labeled data improved this to 93.56%.
- Human Evaluation: Human judges rated 85.32% of the filtered model's headlines as truthful, compared to 77.06% for the baseline.
- ROUGE Stability: Crucially, these gains in truthfulness did not come at the cost of standard performance metrics. ROUGE-1 scores remained stable (35.80 for baseline vs. 35.85 for the proposed model).
Conclusion
This study highlights that model hallucinations are often a learned behavior stemming from noisy supervision data. By treating summarization as a textual entailment task during data curation, researchers can significantly improve the factual consistency of abstractive models. Future work may focus on more granular filtering, such as removing specific untruthful spans rather than entire training pairs, and developing more robust automated metrics for factual consistency.
Summary updated