Core Concepts
BEER2 proposes a bidirectional training framework to enhance the interaction between retriever and reader for improved end-to-end Entity Linking performance.
Abstract
1. Introduction:
End-to-End Entity Linking (EL) extracts mentions and links them to entities in a knowledge base.
2. Retriever-Reader Structure:
Retriever retrieves candidate entities, while the reader predicts span positions of entities in text.
3. Methodology:
BEER2 consists of a retriever and reader trained jointly with bidirectional data flows.
4. Data Flow:
"Retriever → Reader" and "Reader → Retriever" data flows guide training interactions.
5. Retriever Module:
Dual-encoder model selects candidate entities based on retrieval scores using sentence and entity representations.
6. Span Information:
Reader's predicted span information aids in auxiliary retrieval for more accurate entity selection.
Stats
Entity Linking (EL) is a fundamental task for Information Extraction and Knowledge Graphs.
The general form of EL aims to find mentions in the given document and then link them to corresponding entities in a specific knowledge base.
Benefiting from dense retrieval, EntQA achieves state-of-the-art performance in end-to-end EL.
Quotes
"The BEER2 contains two data flows: Retriever → Reader and Reader → Retriever."
"Extensive experiments demonstrate the effectiveness of our proposed BEER2."