�ݺ�ߣ

ESBM: An Entity Summarization Benchmark
Qingxia Liu1, Gong Cheng1, Kalpa Gunaratna2, and Yuzhong Qu1
1 National Key Laboratory for Novel Software Technology, Nanjing University, China
2 Samsung Research America, Mountain View CA, USA

■ Introduction
■ Creating ESBM
■ Analyzing ESBM
■ Evaluating with ESBM
■ Conclusion
2020.06 2
Outline

2020.06 3
Entity Summarization
<Tim Berners Lee, alias, “TimBL”>
<Tim Berners Lee, name, “Tim Berners-Lee”>
<Tim Berners Lee, givenName, “Tim”>
<Tim Berners Lee, birthYear, “1955”>
<Tim Berners Lee, birthDate, “1955-06-08”>
<Tim Berners Lee, birthPlace, England>
<Tim Berners Lee, birthPlace, London>
<Tim Berners Lee, type, People Educated At Emanuel School>
<Tim Berners Lee, type, Scientist>
<Tim Berners-Lee, child, Ben Berners-Lee>
<Tim Berners-Lee, child, Alice Berners-Lee>
<Conway Berners-Lee, child, Tim Berners-Lee>
<Weaving the Web, author, Tim Berners-Lee>
<Tabulator, author, Tim Berners-Lee>
<Paul Otlet, influenced, Tim Berners-Lee>
<John Postel, influenced, Tim Berners-Lee>
<World Wide Web, developer, Tim Berners-Lee>
<World Wide Web Foundation, foundedBy, Tim Berners-Lee>
<World Wide Web Foundation, keyPerson, Tim Berners-Lee><Tim Berners Lee, type, Living People>
<Tim Berners Lee, type, Person>
<Tim Berners Lee, type, Agent>
<Tim Berners-Lee, award, Royal Society>
<Tim Berners-Lee, award, Royal Academy of Engineering >
<Tim Berners-Lee, award, Order of Merit>
<Tim Berners-Lee, award, Royal Order of the British Empire>
<Tim Berners-Lee, spouse, Rosemary Leith>
<Tim Berners Lee, birthDate, “1955-06-08”>
<Tim Berners Lee, birthPlace, England>
<Tim Berners Lee, type, Scientist>
<Tim Berners-Lee, award, Royal Society>
<World Wide Web, developer, Tim Berners-Lee>
Description of Tim Berners-Lee:
Summary:

■ RDF Data: T
● triple t∈T: <subj, pred, obj>
■ Entity Description: Desc(e)
● Desc(e) ={t∈T: subj(t)=e or obj(t)=e}
● triple t∈Desc(e): <e, property, value>
● values: class, entity, literal
■ Entity Summarization (ES): S(e, k)
● S⊆Desc(e) , |S|≤k
2020.06 4
Entity Summarization
Tim Berners-Lee
England
Scientist
Royal Society
Weaving the Web
Person
Paul Otlet
“Tim” “Tim Berners-Lee”
John Postel
“1955-06-08”
“1955”
valuesproperties
birthPlace type
type
author
influenced
influenced
name
givenName birthYear
birthDate
award

■ Limitations
● Task specificness
● Single dataset
● Small size
● Triple incomprehensiveness
2020.06 5
Existing Benchmarks
1 http://yovisto.com/labs/iswc2012
2 http://wiki.knoesis.org/index.php/FACES

■ Motivation
Research Challenges for Entity Summarization:
● Lack of good benchmarks
● Lack of evaluation efforts
■ Contributions
● Created an Entity Sumarization Benchmark (ESBM v1.2)
– overcoming the limitations of existing benchmarks
– meeting the desiderata for a successful benchmark
● Evaluated entity summarizers with ESBM
– made the most extensive evaluation effort to date
– evaluated 9 existing general-purpose entity summarizers
– evaluated 1 supervised learning-based entity summarizer for reference
2020.06 6
Our Work

■ To satisfy seven desiderata for a successful benchmark[18]
● accessibility, affordability, clarity, relevance, solvability, portability, scalability
■ To overcome limitations of available benchmarks
● General-purpose summaries
● Including class-, entity-, literal-valued triples
● Multiple datasets
● Currently largest available benchmark
2020.06 8
Design Goals
[18] Sim, S.E., Easterbrook, S.M., Holt, R.C.: Using benchmarking to advance research: A challenge to software engineering. In: ICSE 2003. pp. 74{83 (2003).

■ Datasets
● DBpedia
– imported dump files: instance types, instance types transitive, YAGO types, mappingbased
literals, mappingbased objects, labels, images, homepages, persondata, geo coordinates
mappingbased, and article categories
● LinkedMDB
– removed triples: owl:sameAs
■ Entities
sampled from seven large classes:
● DBpedia: Agent, Event, Location, Species, Work
● LinkedMDB: Film, Person
■ Triples per entity
● By class: 25.88-52.44 triples
● Overall: 37.62 triples
2020.06 9
Entity Descriptions

2020.06 10
Ground-Truth Summaries
■ Task
● 30 users
– each assigned 35 entities
● 175 entities
– each assigned to 6 users
● Each user created two
summaries for each entity
– for k=5 and k=10
■ Total
● 6 top-5 summaries
and 6 top-10 summaries
for each entity
● 175*6*2=2100 ground-truth summaries

■ Usage
● ESBM v1.2: specified training-validation-test splits for 5-fold cross validation
● Early versions: EYRE 2018 workshop, EYRE 2019 workshop
■ Desiderata
● Accessibility: permanent identifier on w3id.org
● Affordability: open-source, example code for evaluation
● Clarity: documented clearly and concisely
● Relevance: entities sampled from real datasets
● Solvability: not trivial and not too difficult
● Portability: any general-purpose entity summarizer that can process RDF data
● Scalability: reasonably large and diverse to evaluate mature entity summarizers
2020.06 11
The ESBM Benchmark

■ 175 entities, 6584 triples, 2100 ground-truth summaries
2020.06 13
Basic Statistics
Proportion of triples been selected into ground-truth summaries
Overlap: 4.91 triples
Top-5
summary
Top-10
summary
Overlap between top-5 and top-10 summaries

■ Literal-valued triples constitute a large proportion in ground-truth summaries.
● 30% in top-5 ground-truth summaries and 25% in top-10 summaries
■ Participants are not inclined to select multiple values of a property.
● The average number of distinct properties in top-5 ground-truth summaries is 4.70 (very close to 5)
2020.06 14
Triple Composition
Three bars in each group: Entity descriptions, Top-5 ground-truth summaries, Top-10 ground-truth summaries

■ Entity Description
● Jaccard similarity between property sets from each pair of classes is very low.
2020.06 15
Entity Heterogeneity

■ Ground-truth Summaries
● Popular properties:
– properties that appear in >50% ground truth summaries for each class
● Only 1~2/13.24 properties are popular in top-5 ground-truth summaries
– The importance of properties is generally contextualized by concrete entities.
2020.06 16
Entity Heterogeneity

■ Average overlap between 6 ground-truth summaries
● Moderate degree of agreement
● Comparable with those reported for other benchmarks
2020.06 17
Inter-Rater Agreement
[2] Cheng, G., Tran, T., Qu, Y.: RELIN: relatedness and informativeness-based centrality for entity summarization. In: ISWC 2011, Part I. pp. 114-129 (2011).
[7] Gunaratna, K., Thirunarayan, K., Sheth, A.P.: FACES: diversity-aware entity summarization using incremental hierarchical conceptual clustering. In: AAAI 2015. pp. 116-122 (2015).
[8] Gunaratna, K., Thirunarayan, K., Sheth, A.P., Cheng, G.: Gleaning types for literals in RDF triples with application to entity summarization. In: ESWC 2016. pp. 85-100 (2016).

Evaluating with ESBM
2020.06 18

■ Existing Entity Summarizers
● RELIN, DIVERSUM, LinkSUM, FACES, FACES-E, CD
● MPSUM, BAFREC, KAFCA
■ ORACLE Entity Summarizer
● k triples that are selected by the most participants into ground-truth summaries
■ Supervised Learning-Based Entity Summarizer
● 6 models:
– SMOreg, LinearRegression, MultilayerPerceptron, AdditiveRegression, REPTree,
RandomForest
● 7 features:
– gfT(global frequency of property), lf(local frequency of property), vfT(frequency of value),
si(self-information of triple)
– isC(value is class), isE(value is entity), isL(value is literal)
2020.06 19
Participating Entity Summarizers

■ Evaluation Criteria
● Sm: machine-generated entity summary
● Sh : human-made ground-truth summary
● P≠R if |Sm|<|Sh |=k
2020.06 20
Settings

2020.06 22
Results on Different Entity Types
k=5
k=10

■ F1 results
● RandomForest, REPTree
achieve the highest F1.
● Four methods
outperform all the
existing entity
summarizers.
● Two methods only fail to
outperform existing
entity summarizers in one
setting.
2020.06 23
Results of Supervised Learning
Demonstrated the powerfulness of supervised learning for entity summarization.

2020.06 24
Results of Supervised Learning
■ Features
for each t=<e, p,v> in Desc(e):
● gfT: # triples in the dataset where p appears
● lf: # triples in Desc(e) where p appears
● vfT: # triples in dataset where v appears
● si: self-information of triple t
● isC: whether v is a class
● isE: whether v is an entity
● isL: whether v is a literal
■ Results
● significantly effective: gfT, lf
– for LinkedMDB: vfT, si
● not significant: isC, isE, isL

■ Existing entity summarizers
● Leading systems: BAFREC, MPSUM
■ Supervised Learning method
● Outperforms existing entity summarizers
■ Comparing with ORACLE
● Still a large gap for improvement
2020.06 25
Summary of Evaluation Results
Entity summarization on ESBM is a non-trivial task.

■ Evaluation Criteria
● semantic overlap between triples
■ Representativeness of Ground Truth
● general-purpose VS. task-specific
■ Form of Ground Truth
● set-based VS. scoring-based
2020.06 27
Limitations

■ Contributions
● Created an Entity Summarization Benchmark: ESBM
– the most extensive evaluation effort to date
■ ESBM
● The currently largest available benchmark for entity summarization
● Entity summarization on ESBM is a non-trivial task
● Permanent link: https://w3id.org/esbm/
● GitHub repository: nju-websoft/ESBM
2020.06 28
Take-home Message
■ An Upcoming Paper
● Junyou Li, Gong Cheng, Qingxia Liu, Wen Zhang, Evgeny Kharlamov, Kalpa Gunaratna, Huajun Chen.
Neural Entity Summarization with Joint Encoding and Weak Supervision.
IJCAI-PRICAI 2020
– Deep learning based
– Significantly outperformed all the existing systems on ESBM

Thank you !
Questions ?
2020.06 29

■ Contributions
● Created an Entity Summarization Benchmark: ESBM
– the most extensive evaluation effort to date
■ ESBM
● The currently largest available benchmark for entity summarization
● Entity summarization on ESBM is a non-trivial task
● Permanent link: https://w3id.org/esbm/
● GitHub repository: nju-websoft/ESBM
2020.06 30
Take-home Message
■ An Upcoming Paper
● Junyou Li, Gong Cheng, Qingxia Liu, Wen Zhang, Evgeny Kharlamov, Kalpa Gunaratna, Huajun Chen.
Neural Entity Summarization with Joint Encoding and Weak Supervision.
IJCAI-PRICAI 2020
– Deep learning based
– Significantly outperformed all the existing systems on ESBM

�ݺ�ߣ

ESBM: An Entity Summarization Benchmark (ESWC 2020)

Recommended

More Related Content

Similar to ESBM: An Entity Summarization Benchmark (ESWC 2020) (20)

Recently uploaded (20)

ESBM: An Entity Summarization Benchmark (ESWC 2020)