Knowledge Graph知识图谱—10. Data Quality and Linking

10. Data Quality and Linking

10.1 How well are the linked open data in practice?

Linked Open Vocabularies(LOV) project
– analyze usage of vocabularies

10.2 Quality

Linked Data Conformance vs. Quality
Conformance: – i.e., following standards and best practices, technical dimension, can be evaluated automatically

Quality: – i.e., how complete/correct/… is the data, content dimension, hard to evaluate automatically

Example: Crowd Evaluation of DBpedia

The Quality of Linked Open Data is far from perfect: conformance & content
Improving the quality is an active field of research
– Survey 2017: >40 approaches
– since then: a lot of work in KG embeddings

10.3 Links

Previously on Knowledge Graphs

• Integrate data from different sources
• Make connections between entities in those sources
• Facilitate cross data source queries
• Overcome data silos

Why do we need Links?

How do we Create the Links?

数据太多，很多将自己的跟其他数据集互连

10.3.1 Tool Support

A plethora of names
Mostly used for schema level:

• Ontology matching/alignment/mapping
• Schema matching/mapping

Mostly used for the instance level:

• Instance matching/alignment
• Interlinking
• Link discovery

10.3.2 Automating Interlinking

Basic Interlinking Techniques

Sources for Interlinking Signals

Simple String Based Metrics

• String equality
  e.g. foo:University_of_Mannheim, bar:University_of_Mannheim
• Common prefixes
  e.g. foo:United_States, bar:United_States_of_America
• Common postfixes
  e.g. foo:Barack_Obama, bar:Obama
• Typical usage of prefixes/postfixes: |common|/max(length)
  foo:United_States, bar:United_States_of_America → 12/22
  foo:Barack_Obama, bar:Obama → 5/12

Edit Distance

N-gram based Similarity

Typical Preprocessing Techniques

Language-specific Preprocessing

Using External Knowledge

From Matching Literals to Matching Entities

Preprocessing and Matching Pipelines

10.4 Schema Matching

10.5 Instance based Matching

Enforcing 1:1 Mappings

10.5 Matcher Combination

Evaluating Matchers

Challenges in Matching