melt - semantics 2020...data and web science group, university of mannheim...

37
MELT Matching EvaLuation Toolkit SEMANTiCS 2019 – Karlsruhe – 2019/09/11 Sven Hertling; Jan Portisch; Heiko Paulheim 1

Upload: others

Post on 08-Jun-2020

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

MELTMatching EvaLuation Toolkit

SEMANTiCS 2019 – Karlsruhe – 2019/09/11

Sven Hertling; Jan Portisch; Heiko Paulheim

1

Page 2: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Joint Work

2

Heiko PaulheimData and Web Science Group, University of Mannheim

[email protected]

Sven HertlingData and Web Science Group, University of Mannheim

[email protected]

Jan PortischData and Web Science Group, University of Mannheim / SAP SE

[email protected]

Page 3: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Agenda

• Motivation

• What is MELT?

• Usage Example

• Q&A

3

Page 4: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

MOTIVATION

4

Page 5: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Ontology Alignment Evaluation Initiative (OAEI)

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 5

Ontology Alignment Evaluation Initiative• running campaigns since 2005• structured in tracks (similar to task

sets)• researchers submit their

implementation• centrally evaluated by track

organizers• results published

Page 6: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Tooling

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 6

Semantic Evaluation at Large Scale• (among others) packaging and

evaluation• OAEI support since 2010

Holistic Benchmarking of Big Linked Data • (among others) packaging and

evaluation• OAEI support since 2017• OAEI 2018: 6/19 matchers support

HOBBIT

Page 7: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Pain Points

• Limited evaluation capabilities in SEALS, HOBBIT, and the Alignment API

• No easy-to-use parameter tuning

• Packaging process might be complicated for new entrants to the community

• Tooling Java-focused (no Python)

• Implementation of the Alignment API not maven-based

• Tool breaks

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 7

Page 8: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

What is MELT?

8

Page 9: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

What is MELT?

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE

9

• Facilitate matcher packaging• Facilitate matcher

submission

• Easy matcher development• Non-Java matcher

development• Maven support

• Advanced evaluation capabilities • Evaluation before packaging• Allow for interactive visualization

• Allow for parameteroptimization

• Streamlined development process• Integration with existing tooling• OAEI support• Extensibility

Page 10: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

What is MELT?

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 10

Matcher Development

Matcher Fine-Tuning

Matcher Evaluation

Matcher Submission

Page 11: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

What is MELT?

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 11

Matcher Development

Matcher Fine-Tuning

Matcher Evaluation

Matcher Submission

Page 12: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Yet Another Alignment API (YAAA)

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 12

Page 13: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Full SEALS/HOBBIT Support

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 13

Page 14: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

External Matching

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 14

• Simple wrapping

• Packageable for HOBBIT and SEALS

• Matcher can still be evaluated in MELT

• Documentation and demo project available on GitHub

Page 15: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

What is MELT?

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 15

Matcher Development

Matcher Fine-Tuning

Matcher Evaluation

Matcher Submission

Page 16: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Parameter Tuning

• Run matcher configurations in parallel (i.e., multi-threaded)

• Hand over ExecutionResultSet to Evaluator and pick best value according to what you want to optimize

• Out-of-the-box classes that assist you

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 16

Page 17: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

What is MELT?

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 17

Matcher Development

Matcher Fine-Tuning

Matcher Evaluation

Matcher Submission

Page 18: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Evaluation

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 18

Track track = TrackRepository.Multifarm.getSpecificMultifarmTrack("ar", "cn");

ExecutionResultSet ers = new ExecutionResultSet();

ers.addAll(Executor.run(track.getTestCases(), new Matcher(), "Matcher"));

EvaluatorCSV evaluatorCSV = new EvaluatorCSV(ers);

evaluatorCSV.write();

Page 19: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Evaluation

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 19

Track track = TrackRepository.Multifarm.getSpecificMultifarmTrack("ar", "cn");

ExecutionResultSet ers = new ExecutionResultSet();

ers.addAll(Executor.run(track.getTestCases(), new Matcher(), "Matcher"));

EvaluatorCSV evaluatorCSV = new EvaluatorCSV(ers);

evaluatorCSV.write();

Full OAEI support: All tracks available (one-time automated download)

Page 20: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Evaluation

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 20

Track track = TrackRepository.Multifarm.getSpecificMultifarmTrack("ar", "cn");

ExecutionResultSet ers = new ExecutionResultSet();

ers.addAll(Executor.run(track.getTestCases(), new Matcher(), "Matcher"));

EvaluatorCSV evaluatorCSV = new EvaluatorCSV(ers);

evaluatorCSV.write();

Full OAEI support: All tracks available (one-time automated download)

Multiple evaluators available, extensible.

Page 21: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Exemplary Evaluation

“Show me the false positive class-class mappings for Multifarm on track en-de for matcher WiktionaryMatcher.”

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 21

Page 22: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Exemplary Evaluation

“Show me the false positive class-class mappings for Multifarm on track en-de for matcher WiktionaryMatcher.”

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 22

Page 23: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Exemplary Evaluation

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 23

“Show me the false positive class-class mappings for Multifarm on track en-de for matcher WiktionaryMatcher.” → Just filter the correspondences!

Page 24: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

More Evaluation

Filter for Scores or Correspondences

• Micro Average Precision and Macro Average Precision over track Conference

• All residual true positives for track Anatomy

• Macro Average Class-F1 for all tracks

• …

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 24

Page 25: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

What is MELT?

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 25

Matcher Development

Matcher Fine-Tuning

Matcher Evaluation

Matcher Submission

Page 26: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Matcher Submission

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 26

pom

xml

(fill out a template –no maven wizardry required)

> mvn install

> mvn deploy

• Creates SEALS zip file → can be submitted right away

• Creates HOBBIT docker container

• Uploads and deploys container directly in HOBBITautomatically

Page 27: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Usage Example

27

Page 28: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Usage Example

Motivation

“Which matchers are interesting candidates for combination?”

→ Matchers with a high F1 score and high “diversity”.

28MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE

Page 29: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Quantitative Analysis I

Analysis OAEI 2018 results for Conference and Anatomy: Jaccard overlap of Alignments rendered as heatmap in LaTex.

Executor.loadFromAnatomyResultsFolder(“myPath”);

// few lines of other code (available on GitHub as

// example)

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 29

Page 30: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Results for Anatomy

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 30

Page 31: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Results for Conference

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 31

Page 32: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Quantitative Analysis II

Mean Absolute Deviation (MAD) of Similarities plotted against F1.

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 32

Page 33: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Results for Anatomy

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 33

Page 34: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Results for Anatomy

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 34

Page 35: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Results for Conference

MOTIVATION | WHAT IS MELT? | USAGE EXAMPLE 35

Page 36: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

There is MUCH more to MELT

Ontology Caching Services

36

Matcher Pipelining

Baseline Matchers

One-Time Auto-Downloadof OAEI Tracks

Automatic Reading of OAEI Result Alignments

OAEI-Track Organizer Tools

TRY IT!

Execution of SEALS packages from within MELT

Alignment Extensions

Multi-Threaded Matcher Execution

ExecutionResultIndexing

Alignment Refiners

Page 37: MELT - SEMANTiCS 2020...Data and Web Science Group, University of Mannheim heiko@informatik.uni-mannheim.de Sven Hertling ... Holistic Benchmarking of Big Linked Data • (among others)

Thank you!

Sven HertlingData and Web Science Group, University of [email protected]

Jan PortischData and Web Science Group, University of [email protected]

Heiko PaulheimData and Web Science Group, University of [email protected]

37