Core Concepts
Entity abstract summarization can be improved by disentangling facts from templates and introducing external knowledge to reduce hallucinations.
Abstract
Introduction
Entity abstract summarization aims to generate concise descriptions of entities based on relevant internet documents.
Previous methods suffer from hallucinations, leading to factual errors in summaries.
Data Extraction
"Hallucinations refer to the nonsensical or unfaithful contents in the generated texts."
"Hallucinations are difficult to eliminate under the traditional sequence-to-sequence paradigm."
Experiments
SlotSum framework effectively reduces hallucinations by disentangling facts and introducing external knowledge.
SlotSum outperforms baseline models in factual correctness and linguistic quality.
Ablation Study
SlotSum maintains competitiveness and improves factual correctness.
Guiding models with keys degrades performance on fact-oriented metrics.
Case Study
SlotSum reduces hallucinations but may still include factual errors in summaries.
Related Work
Entity abstract summarization and template-based text generation are key areas of research.
Limitations
Limited dataset domain and potential risks of using frozen Wikipedia data are acknowledged.
Ethics Statement
Data sources, human annotation, intended use, licenses, and terms are discussed.
Stats
"Hallucinations refer to the nonsensical or unfaithful contents in the generated texts."
"Hallucinations are difficult to eliminate under the traditional sequence-to-sequence paradigm."
Quotes
"Hallucinations refer to the nonsensical or unfaithful contents in the generated texts."
"Hallucinations are difficult to eliminate under the traditional sequence-to-sequence paradigm."