Concentric Semantic Snapshot

Preview:

Citation preview

THE CONCENTRIC NATURE OF NEWS SEMANTIC SNAPSHOTS

JOSÉ LUIS REDONDO GARCIAGIUSEPPE RIZZORAPHAËL TRONCY

@peputo / redondo@eurecom.fr

@giusepperizzo / giuseppe.rizzo@eurecom.fr

@rtroncy / raphael.troncy@eurecom.fr

2

Overview

May 1, 2023 8th International Conference on Knowledge Capture

1. Introducing the Problem: Contextualizing News Items o The News Semantic Snapshot (NSS)

2. Previous Work:o Frequency-based Functionso Multidimensional Relevancy Approach

3. A Concentric Model for Generating NSS

3

Overview

May 1, 2023 8th International Conference on Knowledge Capture

1. Introducing the Problem: Contextualizing News Items o The News Semantic Snapshot (NSS)

2. Previous Work:o Frequency-based Functionso Multidimensional Relevancy Approach

3. A Concentric Model for Generating NSS

4

The Problem: Contextualizing News

May 1, 2023 8th International Conference on Knowledge Capture

Wolfgang Schäuble

Finance Minister Ruling Party in Ger.

Christian Democratic Union

1 2 3

5May 1, 2023 8th International Conference on Knowledge Capture

Sarah Harrison

WikiLeaks Editor Airport in Moscow

Sheremetyevo

The Problem: Contextualizing News1 2 3

6

Contextualizing News: Applications

May 1, 2023 8th International Conference on Knowledge Capture

1 2 3

7

1 2 3

News Semantic Snapshot (NSS) [1]

May 1, 2023 8th International Conference on Knowledge Capture

News Semantic Snapshot (NSS)

[1] Redondo et al., Generating the Semantic Snapshot of Newscasts using Entity Expansion, ICWE 2015, Rotterdam.

May 1, 2023 8

Recreating the NSS

News Semantic Snapshot8th International Conference on Knowledge Capture

ea eb ec ed ef eg eh ei ej ek el em

ea ec eh ej ek em

(2) SELECTION: filtering, clustering, ranking…

(1) EXPANSION: query generation, search, document retrieval…

ea eb ec ed

1 2 3

May 1, 2023 9

Involving: (experts in the news domain + users)Dimensions:

Play with the data and help us to extend it at: https://github.com/jluisred/NewsConceptExpansion/wiki/Golden-Standard-Creation

News Semantic Snapshot: Gold Standard

(1) Video Subtitles(2) Image in the video(3) Text in the video image(4) Suggestions of an expert(5) Related articles

8th International Conference on Knowledge Capture

1 2 3

May 1, 2023 10

Recreating the NSS

News Semantic Snapshot8th International Conference on Knowledge Capture

ea eb ec ed ef eg eh ei ej ek el em

ea ec eh ej ek em

(2) SELECTION: filtering, clustering, ranking…

(1) EXPANSION: query generation, search, document retrieval…

ea eb ec ed

1 2 3

May 1, 2023

(1) Bringing in Missing Entities: News Entity Expansion

11

1.a)

8th International Conference on Knowledge Capture

Web sites to be crawled:- Google- L1 : A set of 10

internationals English speaking newspapers

- L2 : A set of 3 international newspapers used in GS

Temporal Window:- 1W: - 2W: Annotation filtering- Schema.org

1.b)Parameters [1]:

1 2 3

[1] Redondo et al., Generating the Semantic Snapshot of Newscasts using Entity Expansion, ICWE 2015, Rotterdam.

May 1, 2023 12

News Semantic Snapshot8th International Conference on Knowledge Capture

ea eb ec ed ef eg eh ei ej ek el em

ea ec eh ej ek em

(2) SELECTION: filtering, clustering, ranking…

(1) EXPANSION: query generation, search, document retrieval…

ea eb ec ed

Recall (E. Expansion) = 0.91

Recall (NER on Subtitles) = 0.42

Recreating the NSS1 2 3

May 1, 2023 138th International Conference on Knowledge Capture

(NSS)

(Entity Expansion)

0

N

FIdeal(ei)

(NSS)

FX(ei)

=?MNDCG

The Selection Problem: 1 2 3

14

Overview

May 1, 2023 8th International Conference on Knowledge Capture

1. Introducing the Problem: Contextualizing News Items o The News Semantic Snapshot (NSS)

2. Previous Work:o Frequency-based Functiono Multidimensional Relevancy Approach

3. A Concentric Model for Generating NSS

May 1, 2023 15

1º Entity Frequency SNOW Workshop 2014 [2]

8th International Conference on Knowledge Capture

A

1 2 3

[2] Redondo et al., Describing and Contextualizing Events in TV News Show}, SNOW Workshop, WWW 2014, Seoul, Korea.

May 1, 2023 16

Frequency Based: Results

8th International Conference on Knowledge Capture

(NSS)

(Expansion)

FREQ0

N

(NSS

)

F(Laura Poitras) = 2

F(Glenn Greenwald) = 1

1 2 3

May 1, 2023 15th International Conference on Web Engineering (ICWE) 17

(Fr) (FrGaussian)

Multidimensional ApproachICWE 2015 [1]

1 2 3

[1] Redondo et al., Generating the Semantic Snapshot of Newscasts using Entity Expansion, ICWE 2015, Rotterdam.

May 1, 2023

POPULARITY (FPOP) EXPERT RULES (FEXP)

18

- Based on Google Trends- w = 2 months- μ + 2*σ (2.5%)

Example:- [ Location, = 0.48 ]- [ Person, = 0.74 ]- [ Organization, = 0.95 ]- [ < 2 , = 0.0 ]

15th International Conference on Web Engineering (ICWE) 18

Multidimensional Approach1 2 3

May 1, 2023 19

- News Entity Expansion + Dimensions Generate the News Semantic Snapshot

- Best score: 0.667 in MNDCG at 10, better than BS1/2

• Collection: CSE (Google + 2W + Schema.org)• Ranking:

• Expert Rules• Popularity

8th International Conference on Knowledge Capture

Multidimensionality: Results1 2 3

May 1, 2023 208th International Conference on Knowledge Capture

(NSS))

(Expansion)

FREQ POP EXP

+ + =

(NSS

)

Multidimensionality: Results1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 21

Follow up: Fine-Tuning

1. Exploit Google Relevance (+1.80%)2. Promote Subtitle Entities (+2.50%)3. Exploit Named Entity Extractor’s confidence (+0.20%)4. Interpret popularity Dimension (+1.40%)5. Performing Clustering before Filtering (-0.60%)

- NO SIGNIFICANT IMPROVEMENT -

1 2 3

May 1, 2023 228th International Conference on Knowledge Capture

(NSS)

TuneFunction XFREQ POP EXP

No Improvement: Why?Re-ShuffleOriginal

(NSS

)

How many Dimensions?How to combine them?

1 2 3

23

Overview

May 1, 2023 8th International Conference on Knowledge Capture

1. Introducing the Problem: Contextualizing News Items o The News Semantic Snapshot (NSS)

2. Previous Work:o Frequency-based Functiono Multidimensional Relevancy Approach

3. A Concentric Model for Generating NSS

May 1, 2023 8th International Conference on Knowledge Capture 24

Thinking Outside the Box:

1. Is there room for improvement?2. Is MNDCG a good measure to

evaluate NSS? 3. How to significantly improve the

approach?

1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 25

Room for Improvement?

GAIN

1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 26

Room for Improvement?1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 27

How to Evaluate NSS? MNDCG:• Too focused on success at first positions (decay

Function)

• NSS intends to be flexible, ranking is application-dependent

COMPACTNESS:• Prioritizes coverage over ranking• Compromise between: Recall and NSS size• Recall*: positives are weighted according to score in GT

(NSS)

1 2 3

May 1, 2023 288th International Conference on Knowledge Capture

Compactness:Recall: 22/33 = 0.66

Sa = 27

Sb = 33

Sc = 54

Sa = 27

Sb = 33

Sc= 54

(NSS

)

A B CA

B

C

> >

1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 29

Re-thinking the Approach: Concentric Snapshot

Duality in News Entity Spectrum:• REPRESENTATIVE entities:

• Driving the plot of the story, sometimes evident for users. • RELEVANT entities

• Related to former via specific reasons

Exploit the entity semantic relations

Popular?

Suggested by Expert?

Informative?Unexpected?

Interesting?

Explicative?Highly

1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 30

Hypothesis: Concentric SnapshotCORE:• Representative entities

• Spottable via Frequency dimensions

• High degree of cohesiveness

CRUST:• Attached to the Core via

particular relations

• Agnostic to relevancy nature: informativeness, interestingness, etc.

1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 31

Core Generationa) Representative entities: Frequency Dimension

(NSS)

b) Cohesiveness (DBpedia)

1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 32

Crust Generation

The number of Web documents talking simultaneously about a particular entity e and the Core:

??

1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 33

Experimental Settings

1. Entity Frequency• Core1: Jaro-Winkler > 0.9 • Core2: Frequency based on Exact String matching

2. Cohesiveness: • Everything is Connected Engine [3]• Skb(e1, e2) > 0.125

CORE: (2 configurations)

[3] Everything is Connected Engine:

https://github.com/mmlab/eice

1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 34

1. Candidates for CRUST generation: • Ex1: 1° ICWE2015 by R*(50): L2+Google, F3 1W, Gauss+ POP• Ex2: 2° ICWE 2015 by R*(50): L2+Google, F3 1W, Freq + POP

2. Function for attaching entities to CORE:• SWEB(ei, Core) over Google CSE, default Configuration

CRUST:

Experimental Settings1 2 3

(2 configurations)

May 1, 2023 8th International Conference on Knowledge Capture 35

• Core+Crust: • CrustOnly:

Projecting CORE and CRUST:

(NSS)

(Expansion)

CORE CRUST Core+Crust CrustOnly

Experimental Settings1 2 3

(2 configurations)

May 1, 2023 8th International Conference on Knowledge Capture 36

Baselines:

BAS01: best run in ICWE 2015 at R*(50)BAS02: second best run in ICWE 2015 at R*(50)

FREQPOPEXP

Experimental Settings1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 37

Results: Compactness

Percentage decrease of 36.9% over BAS01

IdealGT: size of SSN according to Gold Standard

(2*2*2 + 2) Runs

1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 38

Results: Recall* over N1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 39

Conclusion• News applications can benefit from the News Semantic Snapshot (NSS)

• Proposed a concentric based model for generating the NSS:• Formalizes duality in entities (Representative VS Relevant)• Exploit the entity semantic relations between Core and Crust.• Accommodate into a single model different relevancy dimensions via the notion of

web presence ( SWeb )

• Concentric model better reproduces the NSS:• Better Compactness: 36.9% over BAS01• Similar recall, Smaller size

• Concentric model easier to implement:

• Core can be reproduced via Frequency Dimension• Crust brings up relevant entities without having to deal with fuzzy dimensions

1 2 3

May 1, 2023 8th International Conference on Knowledge Capture 40

Future• Extend the number of videos considered in GT:

From 5 to 23 (+18), check [4] for more information

• Spot not only relationships between Crust and the Core but also predicates that characterize them:

[4] https://github.com/jluisred/NewsConceptExpansion/wiki/Golden-Standard-Creation

Editor in WikiLeaks

1 2 3

JOSÉ LUIS REDONDO GARCIAGIUSEPPE RIZZORAPHAËL TRONCY

@peputo / redondo@eurecom.fr

@giusepperizzo / giuseppe.rizzo@eurecom.fr

@rtroncy / raphael.troncy@eurecom.fr

http://www.slideshare.net/joseluisredondo/concentric-semantic-snapshot

Visit poster at booth:

34

Recommended