SLIDE 1
Speaker: Shih-Han Lo Advisor: Professor Jia-Ling Koh Author: Aasish Pappu, Roi Blanco, Yashar Mehdad, Amanda Stent, Kapil Thadani Date: 2017/09/19 Source: WSDM ’17
1
Outline
2
Extraction and Linking Speaker: Shih-Han Lo Advisor: Professor - - PowerPoint PPT Presentation
Extraction and Linking Speaker: Shih-Han Lo Advisor: Professor - - PowerPoint PPT Presentation
Lightweight Multilingual Entity Extraction and Linking Speaker: Shih-Han Lo Advisor: Professor Jia-Ling Koh Author: Aasish Pappu, Roi Blanco, Yashar Mehdad, Amanda Stent, Kapil Thadani Date: 2017/09/19 Source: WSDM 17 1 Outline
SLIDE 2
SLIDE 3
Introduction
3
Key tasks for text analytic systems:
Named Entity Recognition (NER) Named Entity Linking (NEL)
Some systems perform NER and NEL jointly.
SLIDE 4
Introduction
4
Most approaches involve (some of) the following
steps:
Mention detection Mention normalization Candidate entity retrieval for each mention Entity disambiguation for mentions with multiple
candidate entities
Mention clustering for mentions that do not link
to any entity
Motivation
SLIDE 5
Outline
5
Introduction Method Experiment Conclusion
SLIDE 6
Mention Detection
6
Typically consists of running an NER system over
input text.
We use simple CRFs and only a few lexical,
syntactic and semantic features.
SLIDE 7
System Description
7
SLIDE 8
Candidate Entity Retrieval
8
Entity Embeddings
We aim to simultaneously learn D-dimensional
representations of Ent and W in a common vector space.
Training our embedding model: continuous skip-
grams with 300 dimensions and a window size of 10.
SLIDE 9
Candidate Entity Retrieval
9
Entity Embeddings
SLIDE 10
Candidate Entity Retrieval
10
Fast Entity Linking
Fast Entity Linker (FEL) is an unsupervised
approach.
FEL imposes contextual dependencies by
calculating the cosine distance between two entities.
Candidate From the substrings of the input string
Minimal perfect hash function Elias-Fano integer coding
SLIDE 11
Entity Disambiguation
11
Task of figuring out to which candidate entity a
mention refers.
The task is complex because mentions may refer
to different entities, depend on local context.
SLIDE 12
Entity Disambiguation
12
Forward-Backward Algorithm (FwBw)
SLIDE 13
Entity Disambiguation
13
Exemplar (Clustering)
SLIDE 14
Entity Disambiguation
14
Label Propagation (LabelProp)
Modified adsorption (MAD) For , we inject seed labels L on a
few nodes.
For nodes V’, we assign a label distribution: Along with , MAD takes three hyper-
parameters as input.
We pick the highest ranked label for each node in
V as the final candidate.
SLIDE 15
Outline
15
Introduction Method Experiment Conclusion
SLIDE 16
Experiment
16
Datasets:
Cross-lingual TAC KBP 2013 Mono-lingual AIDA-CONLL 2003
SLIDE 17
Experiment
17
Setup
N-best: N = 10 FwBw: λ = 0.5 Exemplar: max_iterations = 300, λ = 0.5 LabelProp: μ1 = 1, μ2 = 1e − 2, μ3 = 1e − 2
SLIDE 18
Experiment
18
TAC KBP Evaluation Results
SLIDE 19
Experiment
19
Analysis
SLIDE 20
Experiment
20
Analysis
SLIDE 21
Experiment
21
AIDA Evaluation
SLIDE 22
Experiment
22
Runtime Performance
SLIDE 23
Outline
23
Introduction Method Experiment Conclusion
SLIDE 24
Conclusion
24