This paper describes the rst participation of the SINAI team in the CLEF 2009 wikipediaMM task. This year, we only want to establish a rst contact with the task and the collections. Thus, we have generated a new collection expanding with WordNet terms in order to perform the information included in this collection. In addition, we have expanded de queries with WordNet too. We have used the LEMUR toolkit as the Information Retrieval system in our experiments.
This paper presents the rst participation of the SINAI research group at the CLEF wikipediaMM
task, although this is our fth participation at the ImageCLEF campaign. We have experience in
tasks that involve the retrieving of relevant images using the text associated to each image. In
this eld, we have obtained promising results combining some information retrieval systems and
using several ontologies for the expansion of queries and the textual information of the collection
[
Our main goal in this work is to study the use of the WordNet expansion technique over a
collection and the queries. The integration of external resources to achieve a better information is a
technique that has been used to improve the accuracy of the systems. For example, the integration
of knowledge through the use of ontologies has been very successful in many systems. Speci cally,
WordNet1 [
The following section describes the collection and queries expansion using WordNet. In Section 3, we explain the experiments and obtained results. Finally, conclusions are presented in Section 4.
WordNet is a large lexical database of English, developed in Princeton University. It groups nouns, verbs, adjectives and adverbs into sets of synonyms called synsets and records various semantic relations between these synsets. In our experiments, we have expanded collection and topic set with WordNet ontology using only the words includes in all the synset of each word. Next steps have been followed to expand the text:
3
The purpose of these experiments was to compare the performance of collection expansion when the query is expanded in the same way. We have used two textual collections and two sets of topics. The collections are:
NT: It contains the text of narrative and title labels from original collection.
NTWn: Constains narrative and title labels expanded with WordNet.
The sets of topics use only the title label of original topics to investigate if the expansion of a short text topic improves the system performance. The name of the topic sets are:
TWn: The title label of original topics expanded with WordNet
The name of the experiment includes the name of the group, the name of the collection and the name of the topic set.
sinai NT T: Collection include narrative and title and topics include only title. sinai NTWn T: Collection include narrative and title expanded with WordNet and topics include only title. sinai NT TWn: Collection include narrative and title and topics include title expanded with WordNet. sinai NTWn TWn: Collection include narrative and title expanded with WordNet and topics include title expanded with WordNet.
The dataset of the collection has been indexed using Lemur2 IR system, by applying KLdivergence weighing function and using Pseudo-Relevance Feedback (PRF). Table 1 shows the main average precision (MAP) of our runs.
As we can see in Table 1 the query expansion does not improve the results although using only a expansion for the collection slightly achieve better MAP.
The expasion algorithm has several problems. The query expansion includes all synsets and add noise to the query. For example, the query number 102 is `build site' and the word site has
sinai NT T sinai NTWn T sinai NT TWn sinai NTWn TWn
NTWn
T
T TWn TWn synsets with the multi-words `internet site' and `web site'. These multi-words are included in the expansion of the query but its have not relation with the query. Other problem with the expansion methodology is the eliminition of repeated words. The multi-words `internet site' and `web site' contain the word site. Therefore, the expansion includes the word site several time, because the algorithm does not compare each word in the multi-word. The algorithm handles each multi-word like a simple word. 4
In this paper we describe our rst participation at ImageCLEFwiki. We have experimented with di erent kind of expansion using the external resource WordNet. However, the obtained results show that it is necessary to continue investigating the expansion methodology.
Thus, our next goal will be to improve the expansion by applying some more techniques. For example, it will be interesting to prove a word disambiguation procedure before to incorporate the synset. In addition, further ltering of words included in multiword expressions could achieve better results.
This work has been supported by the Regional Government of Andaluc a (Spain) under excellence project GeOasis (P08-41999), the Spanish Government under project Text-Mess TIMOM (TIN2006-15265-C06-03) and the local project RFC/PP2008/UJA-08- 16-14.