RAED: Retrieval-Augmented Entity Description Generation for Emerging Entity Linking and Disambiguation

Karim Ghonim, Pere-Lluís Huguet Cabot, Riccardo Orlando, Roberto NavigliEMNLP 2025

TL;DR

This paper introduces RAED, a novel framework that uses retrieval to generate entity descriptions, enabling better disambiguation and linking, especially for "emerging" entities not present in traditional knowledge bases. By grounding descriptions in external, up-to-date information, RAED significantly reduces hallucinations and improves performance on challenging tasks.

What Problem Does This Paper Address?

Entity Linking (EL) and Entity Disambiguation (ED) are crucial NLP tasks that aim to connect mentions in text to specific entities in a knowledge base. Traditional systems rely on predefined, static knowledge bases like Wikipedia. However, the real world is dynamic, with new entities constantly emerging (e.g., COVID-19). This creates a significant problem: when entities are new or poorly described in the knowledge base, existing models fail because they lack the necessary information. This limitation hinders applications in fact verification, knowledge base completion, and tasks requiring up-to-date information.

Methods

RAED tackles this by employing a Retrieval-Augmented Entity Description Generation approach. It consists of two main components:

  1. Passage Index: RAED first builds a non-parametric memory by indexing Wikipedia. Instead of just using page titles and definitions, it extracts relevant passages that mention entities, creating a richer source of contextual information. This index is filtered to manage size and relevance.
  2. Retriever: A dense passage retriever (similar to DPR) is trained to find the most relevant passages for a given entity mention. The input to the retriever is the entity mention, specially formatted to highlight the entity itself.
  3. Entity Description Generation: A generative model (like T5 or Llama) then takes the retrieved passages and the original query as input to generate a factual and contextually relevant description (title and definition) for the entity. This generation process is conditioned on the retrieved passages, grounding the output and reducing hallucinations.

The key novelty lies in using retrieved passages (not just candidate entities or static descriptions) to generate descriptions, enabling the system to handle entities not present in predefined inventories.

Datasets

  • English Wikipedia: Used to construct the passage index for retrieval.
  • AIDA-CONLL-YAGO: A standard benchmark for ED and EL, used for training and in-domain evaluation.
  • Out-of-Domain (OOD) Datasets: Including MSNBC, AQUAINT, ACE2004, WNED-CWEB, and WNED-WIKI, used to evaluate generalization to different domains.
  • TempEL: A temporal benchmark that provides yearly snapshots of Wikipedia entities, crucial for evaluating performance on emerging entities.

Key Insights and Ideas

  • Passages over Static Descriptions: The core idea is that retrieving relevant passages provides richer, more contextual information than relying solely on static entity definitions or titles, especially for emerging entities.
  • Generative Disambiguation: RAED doesn't just retrieve; it generates descriptions. This generative approach allows it to implicitly disambiguate entities by producing descriptions that are factually grounded and contextually appropriate, even without explicit disambiguation supervision.
  • Efficiency Gains: By leveraging retrieval, RAED can achieve strong performance with smaller models compared to larger models that rely solely on parametric knowledge. This suggests a path towards more efficient NLP systems.
  • Handling the Unknown: The framework is explicitly designed to handle entities not seen during training, a critical limitation in current EL/ED systems.

Key Results

  • Entity Description Generation: RAED consistently improved NLG metrics (BLEU, ROUGE-1, BERTScore, SIM) across various models. Crucially, it achieved a significant average increase of 27.5 points in Factual-NLI scores, indicating much higher factual consistency.
  • Entity Disambiguation: RAED variants outperformed their non-retrieval counterparts on both in-domain and out-of-domain ED benchmarks. Notably, RAED models matched or slightly surpassed specialized ED models like FusionED, despite not being explicitly trained for disambiguation.
  • Emerging Entity Linking: RAED significantly outperformed models without retrieval on the TempEL benchmark, showing an average improvement of 8.1 points in accuracy@64.
  • Qualitative Analysis: LLM-as-a-Judge experiments showed that RAED's descriptions were preferred over non-retrieval descriptions 82% of the time and were competitive with gold descriptions 40% of the time. Examples highlighted RAED's ability to avoid factual errors and disambiguation mistakes.

Limitations and Future Work

  • Reliance on Wikipedia: RAED currently relies solely on Wikipedia, which may contain outdated or biased information. Incorporating additional sources like news archives could improve robustness.
  • Retrieval Quality Dependency: The system's performance is sensitive to retrieval quality; if the retriever fails, descriptions can be inaccurate.
  • No Proactive Entity Discovery: RAED generates descriptions for given mentions but doesn't proactively discover new entities.
  • Future Directions: The authors suggest exploring larger datasets for better adaptability, integrating entity discovery with retrieval and generation into a unified framework, and improving robustness for knowledge base completion and graph construction.

Summary updated

BibTeX
@inproceedings{ghonim2025raed,
  author = {Karim Ghonim and Pere-Lluís Huguet Cabot and Riccardo Orlando and Roberto Navigli},
  title = {{RAED: Retrieval-Augmented Entity Description Generation for Emerging Entity Linking and Disambiguation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure