We describe in this paper our participation in the product search task of LL4IR CLEF 2015 Lab. This task aims to evaluate, with living labs protective point of view, the retrieval effectiveness over e-commerce search engines. During the online shopping process, users would search for interesting products and quickly access those that fit with their needs among a long tail of similar or closely related products. Our contribution addresses head queries that are frequently submitted on e-commerce Web sites. Head queries usually target featured products with several variations, accessories, and complementary products. We propose a probabilistic model for product search based on the intuition that descriptive fields and the category might fit with the query. Finaly, we present results obtained during the second round of the product search task.

1. IRIT at CLEF 2015:
A product search model for head queries
Lamjed Ben Jabeur, Laure Soulier, and Lynda Tamine
IRI, University of Paul Sabatier, Toulouse, France
1

2. Head Queries
What are head queries ?
– Most frequent queries
– Target featured products
2
Hello Ketty
• Miniature World (5)
• Dolls and accessories (3)
• Skill Game (3)
• Makeup Kit (2)
• Entertainment (1)
• Book (1)
• Creative stock (1)
• Decorative objects (1)
• …

3. Head Queries
3
Samsung
• What category is relevant?

4. Product search model
Probabilistic model for product search
– the likelihood that product descriptive fields match the query
– The likelihood that the product category is relevant in regard of
the query
Product relevance
4

5. Topical relevance
BM25F scoring schema
Weighting fields
Term-frequency saturation
5

6. Category relevance
Similarity between query and category
Ranking categories
• Scores of category’s documents
• Log-scaled category size
• 95th percentile aggregation
6

7. Submit run
1. Gather the list queries, stored in cache
2. Get the list of candidate documents to be ranked,
stored in cache
3. Get document content, stored in cache
4. Apply Hungarian Snowball stemmer
5. We compute document scores
6. Format run and submit run
7

8. Run Setup
Field weighting
BM25F configuration
– K1=2.0
– bf=0
8

9. Results
9
Teamname Outcome #Wins #Losses #Ties #Impressions
Baseline 0.4691 91 103 467 661
UiS-Mira 0.3413 71 137 517 725
UiS-Jern 0.3277 58 119 488 665
UiS-UiS 0.2827 54 137 508 699
GESIS 0.2685 40 109 374 523
Teamname Outcome #Wins #Losses #Ties #Impressions
Baseline 0.5284 93 83 598 774
UiS-Jern 0.4795 82 89 596 767
GESIS 0.4520 80 97 639 816
UiS-Mira 0.4389 79 101 577 757
UiS-UiS 0.4118 84 120 527 731
IRIT 0.3990 79 119 593 791
CLEF LL4IR
LL4IR Round #2

10. Results
10
Teamname Outcome #Wins #Losses #Ties #Impressions
IRIT 0.4890 89 93 533 715
UiS-Mira 0.4507 64 78 527 669
Baseline 0.4430 66 83 498 647
GESIS 0.4134 74 105 513 692
UiS-Jern 0.3702 67 114 511 692
Teamname Outcome #Wins #Losses #Ties #Impressions
IRIT 0.4654 101 116 767 984
GESIS 0.4292 103 137 804 1044
Baseline 0.3783 87 143 781 1011
LL4IR Round #3: social engagement
LL4IR Round #4: social engagement

11. Reranking with social engagement
11
Hello Ketty 1
2
" Buborékfújó Hello Kitty"
3
User engagement

12. Conclusion
• Variants of product search model including
field descriptors and/or users' feedback
• Effectiveness improved while integrating
users' feedback
• Future work:
– More in-depth analysis for identifying relevant
evident sources
– Experiments on other search tasks
12

13. Thank you for your attention
@amjedbj
13
amjedbj/vulter
@LaureSoulier
@LyndaTamine