TL;DR
This paper tackles the problem of "hallucinations" (generating false information) in entity abstract summarization. They propose SlotSum, a novel framework that decomposes summaries into factual "facts" and generic "templates," allowing for better error detection and correction using external knowledge.
What Problem Does This Paper Address?
Entity abstract summarization aims to create concise descriptions of entities (like people or places) based on multiple online documents. While powerful language models (PLMs) are good at this, they often "hallucinate" – generating plausible-sounding but factually incorrect information. This is a significant issue for applications like search engines or Wikipedia generation, where accuracy is paramount. Existing methods struggle to reliably prevent these factual errors, even with access to vast amounts of text.
Methods
The core innovation of SlotSum is its facts-template decomposition. Instead of generating a summary directly, SlotSum first creates a template with "slots" representing factual information. Then, it predicts the content for each slot. This decomposition allows the system to:
- Generate Templates: A sequence-to-sequence model generates a template from the input documents. The factual content within the template is replaced by special slot tokens (e.g.,
[SLT] birth_date [/SLT]). This step is designed to be "knowledge-agnostic" to prevent data leakage during training. - Predict and Correct Slots: SlotSum then predicts the actual facts to fill these slots. Crucially, it uses external knowledge sources (like knowledge bases) to verify and correct these predictions. This is where hallucinations are actively addressed. If a predicted fact doesn't align with external knowledge, it's corrected or discarded.
- Construct Summary: Finally, the corrected facts are inserted into the template to form the final summary.
SlotSum offers three strategies for filling slots: "Discard" (only uses corrected facts), "Predict" (predicts all slots and then corrects), and "All-Predict" (predicts all slots without external correction, serving as a baseline).
Datasets
- WikiFactSum: A new dataset constructed by combining the WikiSum dataset (for source documents and summaries) with WikiBio (for structured factual information). This dataset is designed to provide both the context for summarization and the ground truth facts needed for evaluation.
- WikiSum: A widely-used dataset for entity abstract summarization, derived from Wikipedia abstracts.
- WikiBio: A dataset containing structured biographical information.
Key Insights and Ideas
- Decomposition is Key: Separating factual content from the generic structure of a summary is a powerful way to combat hallucinations. Templates provide a scaffold that makes factual errors more apparent.
- External Knowledge is Crucial: While PLMs can generate fluent text, they lack inherent factual grounding. Integrating external knowledge sources is essential for ensuring accuracy.
- Templates Guide, But Don't Guarantee: While templates help identify where facts should be, the "All-Predict" strategy showed that simply guiding a model with fact keys can still lead to hallucinations if the slot prediction itself is flawed. This highlights the importance of the correction step.
- Data Leakage is a Concern: The authors note that large language models like ChatGPT might exhibit data leakage from their pre-training data, making direct comparison unfair.
Key Results
- Significant Improvement in Factual Correctness: SlotSum, especially with the "Predict" and "Discard" strategies, achieved significantly higher scores on fact-oriented metrics (FactCC and UniEval) compared to strong baselines like BART, T5, and PEGASUS.
- ROUGE and BERTScore Improvements: SlotSum also outperformed baselines on standard summarization metrics like ROUGE and BERTScore, indicating that improved factual correctness doesn't come at the cost of fluency or relevance.
- Human Evaluation Confirms: Human evaluations showed SlotSum's superiority in factual correctness, with a notable increase in QA scores when external knowledge was used for correction. Linguistic quality was comparable to other models, though completeness was slightly lower in some cases due to slot limitations.
Limitations and Future Work
- Dataset Domain Limitation: WikiFactSum is currently limited to biographical data. Generalizing SlotSum to other domains will require more diverse training data.
- LLM Performance: The authors did not test large language models like ChatGPT due to concerns about data leakage. Future work could explore how to fairly evaluate these models or adapt SlotSum for them.
- Template Completeness: While SlotSum significantly reduces hallucinations, it doesn't eliminate them entirely. Some facts might not be easily representable as slots, leading to potential omissions or misinterpretations (as seen in the case study with nationality).
- Future Directions: The authors suggest exploring better ways to fuse external knowledge with large language models and developing more sophisticated methods for incorporating knowledge.
Summary updated