Entity Linking via Joint Encoding of Types, Descriptions, and - PowerPoint PPT Presentation

Entity Linking via Joint Encoding of Types, Descriptions, and Context Author: Nitish Gupta, Sameer Singh, Dan Roth Source: EMNLP’17 Speaker: Ya-Wen Yu Date: 2018/7/31 � 1

Outline • Introduction • Method • Experiment • Conclusion � 2

Introduction • Challenge in Entity Linking ( mention ) (entity) � 3

Introduction • Goal - How to correctly resolve the mentions • local context • textual descriptions • fine-grained type � 4

Training Data Wikipedia anchors as mentions, links as true entity Description: first 100 tokens of the entity’s Wikipedia page Type: 112 fine-grained entity-types from Freebase � 5

Framework � 6

Outline • Introduction • Method • Experiment • Conclusion � 7

Method Mention Context Encoder, C Local-Context Encoder Document Context Encoder � 8

Method Mention Context Encoder, C Context Objective � 9

Method Entity Description Encoder, D � 10

Method Fine-grained Type Encoding, E Entity-Type Objective � 11

Method Type-Aware Context Representation, T Mention-context representation aware of entity type information should be helpful Learning Unified Entity Representation � 12

Method Entity Linking Top-30 Candidate entities ( Cm ) from Cross-Wikis � 13

Outline • Introduction • Method • Experiment • Conclusion � 14

Experiment Datasets CoNLL-YAGO (Hof- fart et al., 2011), ACE 2004 (NIST, 2004; Rati- nov et al., 2011), ACE 2005 (NIST, 2005; Ben- tivogli et al., 2010), Wikipedia (Ratinov et al., 2011). � 15

Experiment Entity Linking Performance � 16

Experiment Entity Linking Performance � 17

Experiment Cold-Start Entities • Dataset- 3791 mentions of 1000 rare entities from Wikipedia • New entities- Unknown during training. No linked mentions � 18

Experiment � 19

Experiment Typing Prediction predict fine-grained types for mentions from context (using vm and vt). � 20

Experiment Example Predictions � 21

Outline • Introduction • Method • Experiment • Conclusion � 22

Conclusion • Jointly learn entity embeddings using multiple information sources and leads to more accurate entity linking. • Unstructured: KB Description, Mention Contexts (Wikipedia) • Structured: Fine-grained Entity Types (Freebase) • Robust to missing or incomplete information : Not all 
 information sources needed for all entities • Cold-start entity linking : Accurate entity linking also to entities that were never observed during training 
 � 23