Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Entity Disambiguation via Fusion Entity Decoding

Junxiong Wang, Ali Mousavi, Omar Attia, Ronak Pradeep, Saloni Potdar, Alexander M. Rush, Umar Farooq Minhas, Yunyao Li

TL;DR

This work tackles entity disambiguation and end-to-end entity linking by leveraging detailed entity descriptions. It introduces FusionED, an encoder-decoder architecture where the encoder derives per-candidate representations by jointly encoding the input text with each candidate's title and description, and the decoder fuses these representations to output the correct entity, enabling greedy decoding at inference. The approach is complemented by a retrieval-augmented EL framework (bi-encoder retrieval plus Fusion EL Reader) and a study of retrieval-augmented LLMs, showing strong gains on ZELDA and GERBIL benchmarks while highlighting the ongoing gap between fine-tuned models and LLM-based prompts. The findings underscore the value of incorporating rich entity descriptions for disambiguation and discuss limitations, biases, and avenues for multilingual and prompt-based enhancements in EL.

Abstract

Entity disambiguation (ED), which links the mentions of ambiguous entities to their referent entities in a knowledge base, serves as a core component in entity linking (EL). Existing generative approaches demonstrate improved accuracy compared to classification approaches under the standardized ZELDA benchmark. Nevertheless, generative approaches suffer from the need for large-scale pre-training and inefficient generation. Most importantly, entity descriptions, which could contain crucial information to distinguish similar entities from each other, are often overlooked. We propose an encoder-decoder model to disambiguate entities with more detailed entity descriptions. Given text and candidate entities, the encoder learns interactions between the text and each candidate entity, producing representations for each entity candidate. The decoder then fuses the representations of entity candidates together and selects the correct entity. Our experiments, conducted on various entity disambiguation benchmarks, demonstrate the strong and robust performance of this model, particularly +1.5% in the ZELDA benchmark compared with GENRE. Furthermore, we integrate this approach into the retrieval/reader framework and observe +1.5% improvements in end-to-end entity linking in the GERBIL benchmark compared with EntQA.

Entity Disambiguation via Fusion Entity Decoding

TL;DR

This work tackles entity disambiguation and end-to-end entity linking by leveraging detailed entity descriptions. It introduces FusionED, an encoder-decoder architecture where the encoder derives per-candidate representations by jointly encoding the input text with each candidate's title and description, and the decoder fuses these representations to output the correct entity, enabling greedy decoding at inference. The approach is complemented by a retrieval-augmented EL framework (bi-encoder retrieval plus Fusion EL Reader) and a study of retrieval-augmented LLMs, showing strong gains on ZELDA and GERBIL benchmarks while highlighting the ongoing gap between fine-tuned models and LLM-based prompts. The findings underscore the value of incorporating rich entity descriptions for disambiguation and discuss limitations, biases, and avenues for multilingual and prompt-based enhancements in EL.

Abstract

Entity disambiguation (ED), which links the mentions of ambiguous entities to their referent entities in a knowledge base, serves as a core component in entity linking (EL). Existing generative approaches demonstrate improved accuracy compared to classification approaches under the standardized ZELDA benchmark. Nevertheless, generative approaches suffer from the need for large-scale pre-training and inefficient generation. Most importantly, entity descriptions, which could contain crucial information to distinguish similar entities from each other, are often overlooked. We propose an encoder-decoder model to disambiguate entities with more detailed entity descriptions. Given text and candidate entities, the encoder learns interactions between the text and each candidate entity, producing representations for each entity candidate. The decoder then fuses the representations of entity candidates together and selects the correct entity. Our experiments, conducted on various entity disambiguation benchmarks, demonstrate the strong and robust performance of this model, particularly +1.5% in the ZELDA benchmark compared with GENRE. Furthermore, we integrate this approach into the retrieval/reader framework and observe +1.5% improvements in end-to-end entity linking in the GERBIL benchmark compared with EntQA.
Paper Structure (29 sections, 2 figures, 6 tables)

This paper contains 29 sections, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Entity Disambiguation.
  4. Entity Linking.
  5. Model
  6. Entity Disambiguation
  7. Entity Linking
  8. Bi-encoder EL Retrieval
  9. Entity Embedding.
  10. Passage Embedding.
  11. Training.
  12. Fusion EL Reader
  13. Experiment
  14. Entity Disambiguation
  15. Setup.
  16. ...and 14 more sections

Figures (2)

  • Figure 1: Pipeline of the fusion entity decoding for entity disambiguation. Given a text 'DUBLIN 1996-12-07 Jack Charlton's relationship with the people of Ireland was cemented on Saturday when the Englishman was officially declared one of their own. (few sentences are abbreviated here) That is why this is so emotional a night for me , <s1> Charlton <e1> said'. Follow de2020autoregressive, we add special tokens <s1> and <e1> to denote the corresponding mention to disambiguate. Given candidate entities 'Charlton Athletic F.C.', 'Jack Charlton', 'Bobby Charlton', 'Suzanne Charlton' from KB, we concatenate text with each entity candidate, including its entity title and its description. The Encoder learns interactions between the text and each entity candidate and produces suitable representations for each entity candidate; decoder concatenates those representations and selects the correct entity.
  • Figure 2: Example of document level entity linking from AIDA test. Given a document, FusionED splits it into smaller passage chunks. Given the current passage 'That is why this is so emotional a night for me, Charlton said.', the bi-encoder entity retrieval picks up top 100 entity candidates, e.g., 'Charlton Athletic F.C.', 'Bobby Charlton', 'Jack Charlton'. FusionED then decodes linked entities and mentions using entity candidate lists.