exploratory search on topics through different perspectives with dbpedia

Post on 02-Jul-2015

507 Views

Category:

Technology

0 Downloads

Preview:

Click to see full reader

DESCRIPTION

Presentation at SEMANTICS2014: http://www.semantics.cc/www.semantics.cc/index.html Paper: Exploratory search on topics through different perspectives with DBpedia https://hal.inria.fr/hal-01057031 A promising scenario for combining linked data and search is exploratory search. During exploratory search, the search objective is ill-defined and favorable to discovery. A common limit of the existing linked data based exploratory search systems is that they constrain the exploration through single results selection and ranking schemes. The users can not influence the results to reveal specific aspects of knowledge that interest them. The models and algorithms we propose unveil such knowledge nuances by allowing the exploration of topics through several perspectives. The users adjust important computation parameters through three operations that help retrieving desired exploration perspectives: specification of interest criteria about the topic explored, controlled randomness injection to reveal unexpected knowledge and choice of the processed knowledge source(s). This paper describes the corresponding models, algorithms and the Discovery Hub implementation. It focuses on the three mentioned operations and presents their evaluations.

TRANSCRIPT

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Exploratory search on topics through differentperspectives with DBpedia

Nicolas Marie, Fabien Gandon, Alain Giboin, Émilie Palagi

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

2

CONTEXTPROPOSITIONEVALUATIONCONCLUSION

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

3

CONTEXTPROPOSITIONEVALUATIONCONCLUSION

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Search is only a partially solved problem [White, 2009]Ambiguous queries, natural language queries, exploratory search tasks…

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

10 blue links paradigm, simple, fast

Exploratory searchbottleneck

Discovery Hub

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Overviews

Faceted interfaces

Results clustering

Low-cost of browsing (going back-and-forth functionalities)

Query-suggestions and refinement

Serendipitous discoveries provocation

In-session of account related memory features

Exploratory search systems are optimized to support exploratory search tasks, common functionalities:

Discovery Hub

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Discovery Hub

Linked data are promising for supporting exploratory search: • new algorithms • new interaction models optimized for exploration.

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Discovery Hub

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

MaturityDiscovery Hub

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

1 topic of interest => 1 entity => 1 results set to explore

1 perspective

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Discovery HubTopics are complex, multifaceted,

One entity => multiple perspectives &knowledge nuances

Entourage Art. movement

Curiosities…In French culture

In American culture

I want to discover Claude Monet (painter)...

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

MORE

Aemoo

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

13

CONTEXTPROPOSITIONEVALUATIONCONCLUSION

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Discovery Hub

The models and algorithms we proposeunveil topic knowledge nuances by allowing the exploration of topics through several perspectives.

In the graph context of linked data these perspectives correspond to different non exclusive sets of objects and relations that are informative on a topic regarding specific aspects.

Flexible querying and data processing

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Refer to the papers for the complete formalization

Building perspectives thanks to spreading activation

……

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Discovery Hub

3 perspective-operations to expose knowledge nuances :

• Criteria of interest specification• Controlled randomness injection• Data source selection

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Discovery Hub Criteria of interest specification

, dcterms:category, ?x

, dcterms:category, ?x

Classic similarity measure

, dcterms:category, ?a | ?b | ?c |...

, dcterms:category, ?a | ?b | ?c |...

Criteria spec. similarity

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Discovery Hub Criteria of interest specification

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Classic – top 5 artists

« French / not impressonist » criteria specification – top 5 artists

« Not French / Impressonist » criteria specification – top 5 artists

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Chosen level of randomness

Randomness injection

* r + (1-r)*

* r + (1-r)*

* r + (1-r)*

* r + (1-r)*

* r + (1-r)*

* r + (1

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Discovery Hub

Local Kgram instance

Data source selection

fr.dbpedia.org/sparql

it.dbpedia.org/sparql

de.dbpedia.org/sparql

es.dbpedia.org/sparql

dbpedia.org/sparql

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Discovery HubData source selection

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

28

CONTEXTPROPOSITIONEVALUATIONCONCLUSION

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Evaluated algorithm versions

•Basis algorithm of Discovery Hub

•Personalized algorithm through criteria specification

•Randomized algorithm, with 0.5 threshold

•Highly randomized algorithm (Highly R.), with 1.0 threshold

Discovery Hub

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

• Hypothesis 1:

Users who specify their criteria of interest about a topic find the results of the search more relevant.

• Hypothesis 2:

Users who specify their criteria of interest about a topic do not find the results of the search less novel.

• Hypothesis 3:

The stronger is the level of randomness the more surprising the results are for the users.

• Hypothesis 4:

Even if the level of surprise is high, the majority of the top results are still relevant to the users.

Discovery Hub

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

𝐻3 𝐻𝑖𝑔ℎ𝑙𝑦 𝑅. ; 𝑆𝑢𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑠𝑢𝑙𝑡 > 𝑅𝑎𝑛𝑑𝑜𝑚𝑖𝑧𝑒𝑑 ; 𝑆𝑢𝑟𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑠𝑢𝑙𝑡 > [𝐵𝑎𝑠𝑖𝑠 ; 𝑆𝑢𝑟𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑠𝑢𝑙𝑡]

[ " ; Suprising Relation] > " ; 𝑆𝑢𝑟𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑙𝑎𝑡𝑖𝑜𝑛 > [ " ; 𝑆𝑢𝑟𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑙𝑎𝑡𝑖𝑜𝑛]

𝐻2 𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝑖𝑧𝑒𝑑 ; 𝑆𝑢𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑠𝑢𝑙𝑡 > [𝐵𝑎𝑠𝑖𝑠 ; 𝑆𝑢𝑟𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑠𝑢𝑙𝑡]

[ " ; Surprising Relation] > [ " ; 𝑆𝑢𝑟𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑙𝑎𝑡𝑖𝑜𝑛]

𝐻1 𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝑖𝑧𝑒𝑑 ; 𝐼𝑛𝑡𝑒𝑟𝑒𝑠𝑡 > [𝐵𝑎𝑠𝑖𝑠 ; 𝐼𝑛𝑡𝑒𝑟𝑒𝑠𝑡]

" ; 𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒 < [ " ; 𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒]

𝐻4 𝐻𝑖𝑔ℎ𝑙𝑦 𝑅𝑎𝑛𝑑𝑜𝑚𝑖𝑧𝑒𝑑 ; 𝐼𝑛𝑡𝑒𝑟𝑒𝑠𝑡 > 𝐴𝑣𝑒𝑟𝑎𝑔𝑒 (2,5)

𝑅𝑎𝑛𝑑𝑜𝑚𝑖𝑧𝑒𝑑 ; 𝐼𝑛𝑡𝑒𝑟𝑒𝑠𝑡 > 𝐴𝑣𝑒𝑟𝑎𝑔𝑒 (2,5)

(Highly R. : Highly Randomized)

Discovery Hub

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

A « good »

result in

ESS is…

Our definitions Chosen metrics : Questions

(Likert Scale)

… A

surprising

result

A result is surprising if :

• You discovered an

unknown resource or

relation

• You discovered

something unexpected

Surprising Result

This result is suprising

?

Surprising

Relation

This relation between

the topic searched and

the result is surprising

?

… An

intersting

result

A result is interesting if :

• You think it is similar to

the topic explored

• You think you will

remind or reuse it

Interesting Result

This result is interesting

?

Distance between

the Result and the

topic searched

This result is too distant

from the topic searched

?

Discovery Hub

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

• 16 participants

• Phase 1 : Selection of 2 topics in a list of 20 queries randomly choose in the query log of Discovery Hub

- Information Visualization

- Serge Gainsbourg (french singer)

• Phase 2 : Specification of the categories of interest

• Phase 3 : User Test (~1h)

- Before the test

- Interview (name, age, do they know Discovery Hub ?,…)

- Presentation of Discovery Hub and the objective of the test

- Presentation of the questions and simulation

Discovery Hub

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

𝐻1 𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝑖𝑧𝑒𝑑 ; 𝐼𝑛𝑡𝑒𝑟𝑒𝑠𝑡 > [𝐵𝑎𝑠𝑖𝑠 ; 𝐼𝑛𝑡𝑒𝑟𝑒𝑠𝑡]

" ; 𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒 < [ " ; 𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒]

H1 : Users who specify their criteria of interest about a topic find the results of the search more relevant.

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

𝐻2 𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝑖𝑧𝑒𝑑 ; 𝑆𝑢𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑠𝑢𝑙𝑡 > [𝐵𝑎𝑠𝑖𝑠 ; 𝑆𝑢𝑟𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑠𝑢𝑙𝑡]

[ " ; Surprising Relation] > [ " ; 𝑆𝑢𝑟𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑙𝑎𝑡𝑖𝑜𝑛]

H2: Users who specify their criteria of interest about a topic do not find the results of the search less novel

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

𝐻3 𝐻𝑖𝑔ℎ𝑙𝑦 𝑅. ; 𝑆𝑢𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑠𝑢𝑙𝑡 > 𝑅𝑎𝑛𝑑𝑜𝑚𝑖𝑧𝑒𝑑 ; 𝑆𝑢𝑟𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑠𝑢𝑙𝑡 > [𝐵𝑎𝑠𝑖𝑠 ; 𝑆𝑢𝑟𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑠𝑢𝑙𝑡]

[ " ; Suprising Relation] > " ; 𝑆𝑢𝑟𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑙𝑎𝑡𝑖𝑜𝑛 > [ " ; 𝑆𝑢𝑟𝑝𝑟𝑖𝑠𝑖𝑛𝑔 𝑅𝑒𝑙𝑎𝑡𝑖𝑜𝑛]

(Highly R. : Highly Randomized)

H3: The stronger is the level of randomness the more surprising the results are for the users

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

𝐻4 𝐻𝑖𝑔ℎ𝑙𝑦 𝑅𝑎𝑛𝑑𝑜𝑚𝑖𝑧𝑒𝑑 ; 𝐼𝑛𝑡𝑒𝑟𝑒𝑠𝑡 > 𝐴𝑣𝑒𝑟𝑎𝑔𝑒 (2,5)

𝑅𝑎𝑛𝑑𝑜𝑚𝑖𝑧𝑒𝑑 ; 𝐼𝑛𝑡𝑒𝑟𝑒𝑠𝑡 > 𝐴𝑣𝑒𝑟𝑎𝑔𝑒 (2,5)

H4: Even if the level of surprise is high, the majority of the top results are still relevant to the users

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

38

CONTEXTPROPOSITIONEVALUATIONCONCLUSION

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

Discovery Hub

• We proposed a framework to enable multi-perspective exploratory search:

- Formalization

- Implementation

- Evaluation

• 3 operators : criteria spec., randomization, data selection

• Evaluations globally positive, slight adjustements needed

• Interesting propositions from the reviewers, thank you

COPYRIGHT © 2011 ALCATEL-LUCENT. ALL RIGHTS RESERVED.

http://semreco.inria.fr

werarediscoveryhub@gmail.com

Thank you ! Questions ?

Discovery Hub

top related