The document discusses type-aware entity retrieval, focusing on the transition from traditional information retrieval to entity-oriented search for satisfying complex information needs. It outlines research questions related to representing, combining, and estimating type-based information using various type taxonomies and experimental setups. The findings suggest different approaches and models for type representation and the integration of type and textual information in entity retrieval tasks.

1. Type-aware Entity Retrieval
Darío Garigliotti
IAI
Universitetet i Stavanger (UiS)
NTNU - March 3, 2016

2. From Information Retrieval to Entity
Retrieval
• Traditional Information Retrieval recently extended
to an Entity-oriented Search
• Satisfaction of more complex information needs
• Current support on search engines

3. From Information Retrieval to Entity
Retrieval

4. From Information Retrieval to Entity
Retrieval

5. Countries where one can pay with the euro
Entity Retrieval

6. Countries where one can pay with the euro
• Related entities (via a relation or predicate)
• Types or categories or classes
Entity Retrieval

7. Countries where one can pay with the euro
Impressionist art museums in The Netherlands
• Related entities (via a relation or predicate)
• Types or categories or classes
Entity Retrieval

8. Countries where one can pay with the euro
Impressionist art museums in The Netherlands
• Related entities (via a relation or predicate)
• Types or categories or classes
Entity Retrieval

9. Countries where one can pay with the euro
Impressionist art museums in The Netherlands
• Related entities (via a relation or predicate)
• Types or categories or classes
Entity Retrieval

10. Entity Retrieval
Evaluated tasks
• Entity ranking (given a textual query and target categories)
• List completion (given Q and entity examples, and? types)
• Related entity finding (given entity E, relation R and type T)
e.g. E = "Schumacher", R = "His teammates when he was on
Ferrari", T = "Person"
from Q = "Schumacher teammates when he was on Ferrari"

11. Type-aware entity retrieval
Our research questions
1. How to represent type-based information?
2. How to combine type-based and textual
information?
3. How to estimate type-based information?

12. Type-aware entity retrieval
RQ2. How to combine type-based and textual information?
• Basics: term-based models
• A variety of related tasks across the literature
• Entity retrieval approaches
• Where to look for entities? How to find them? How to rank
them?
• Major model families
• Common main insight: types help!

13. Type-aware entity retrieval
RQ1. How to represent type-based information?
• Dimensions we identified
• type taxonomies
• hierarchical structure
• dataset version
• Minimal concerning in the related work

14. Type taxonomies
• We consider four well-known type taxonomies
Type system Wikipedia DBpedia Freebase YAGO
#types 753,524 591 1719 568,672
#top-level types NA 58 92 61
#most-specific-level types 753,524 472 1626 549,623
depth NA 7 2 19
entities w/ type 4.12M 3.24M 3.77M 2.89M
avg #types/entity 4.02 6.30 9.57 16.44

15. Type representation
• We consider different ways of modeling type
assignments:
Top level, most specific level, and path-to-top
r
e
r
e
r
e

16. Experimental setup
• Our experimental environment looks like this:
Term-based representation
Query model Entity model Query model Entity model
Type-based representation
P(e|q) / P(q|e)P(e)
p(t|✓T
e )p(t|✓T
q )
KL(✓T
q ||✓T
e )
p(t0
| ✓T 0
e )p(t0
| ✓T 0
q )
KL(✓T 0
q k ✓T 0
e )
P(q | e) = (1 )P(✓T 0
q | ✓T 0
e ) + P(✓T
q | ✓T
e )

17. Experimental setup
• Term-based component: Mixture of LM method
• We obtain combinations of these elements:
• Type taxonomies
• Models
• Type-based representations

18. Ingredients
• Model instantiations for
• M1 (Mixture):
• M2 (Multiplicative):
• M3 (Filtering):
P(e | q) / P(q | e)P(e)
P(q | e) = (1 )P(✓T 0
q | ✓T 0
e ) + P(✓T
q | ✓T
e )
P(q | e) = P(✓T 0
q | ✓T 0
e )P(✓T
q | ✓T
e )
P(✓T
q | ✓T
e ) 2 {0, 1}

19. Ingredients
• Query model for the type-based representation
is provided by a target types oracle
P(t|✓T
q )
Query: guitar origin blues DBpedia Types:
<dbo:Album>: 4
<dbo:MusicalArtist>: 43
...
Freebase Types:
<fb:music.group_member>: 34
<fb:people.deceased_person>: 17
...
Wikipedia Categories:
<dbpedia:Category:Blues_musicians_from_New_Orleans,_Louisiana>: 2
<dbpedia:Category:Blues_songs>: 2
...

20. Ingredients
• Our experimental environment looks like this:
Query model Entity model
Type-based representation
P(e|q) / P(q|e)P(e)
p(t|✓T
e )p(t|✓T
q )
KL(✓T
q ||✓T
e )
P(q | e) = (1 )P(✓T 0
q | ✓T 0
e ) + P(✓T
q | ✓T
e )

21. Ingredients
• Entity model for the type-based representation is a
distribution estimated through the entity types
Query: guitar origin blues
Relevant entities:
<dbpedia:The_Merle_Travis_Guitar>
<dbpedia:Blues_Breakers_with_Eric_Clapton>
<dbpedia:Poor_Boy_Blues>
...
...
Freebase Types:
... DBpedia Types:
<dbo:Album>
<dbo:MusicalWork>
...
...
Freebase Types:
... Wikipedia Categories:
<Category:1950_albums>
<Category:Merle_Travis_albums>
...

22. Results (1)
RQ1. How to represent
type-based information?
Type representation - Model M1
MAP
0
0.058
0.115
0.173
0.23
all assigned types most specific level path-to-top top level
YAGO Freebase Wikipedia DBpedia
Type representation - Model M2
MAP
0
0.045
0.09
0.135
0.18
all assigned types most specific level path-to-top top level
YAGO Freebase Wikipedia DBpedia
Type representation - Model M3
MAP
0
0.055
0.11
0.165
0.22
all assigned types most specific level path-to-top top level
YAGO Freebase Wikipedia DBpedia

23. Results (2)
RQ2. How to combine type-based and textual information?
Combining information - All assigned types
MAP
0
0.06
0.12
0.18
0.24
YAGO Freebase Wikipedia DBpedia
M1 M2 M3
Combining information - Most-specific-level types
MAP
0
0.06
0.12
0.18
0.24
YAGO Freebase Wikipedia DBpedia
M1 M2 M3

24. Future work
RQ3: How to estimate type-based information?
Term-based representation
Query model Entity model Query model Entity model
Type-based representation
P(e|q) / P(q|e)P(e)
p(t|✓T
e )p(t|✓T
q )
KL(✓T
q ||✓T
e )
p(t0
| ✓T 0
e )p(t0
| ✓T 0
q )
KL(✓T 0
q k ✓T 0
e )
P(q | e) = (1 )P(✓T 0
q | ✓T 0
e ) + P(✓T
q | ✓T
e )

25. Future work
• Main focus will be on query typing, but eventually
on entity typing as well
• How to take the best from different type taxonomies