Slide 54
Slide 54 text
54
Text-based Bi-encoder
1. Training the encoders for encoding mention and candidates.
2. Project mention and candidates to the same continuous space.
3. Retrieve top-N candidates. (L2-distance, Dot-product, etc.)
mention Candidate
entity
Following [Wu et al., ‘20] and [Logeswaran et al, ‘19]
• for scoring, Dot-product is used.
• for candidates’ feature, entity’s title and description are
used.