Recent researches demonstrate that the use and integration of several knowledge sources improves the quality and e ciency of information systems. In this paper, we present the system developed for the ImageCLEF 2010 medical task. We show the e ect of using the medical ontology MeSH to expand terms found in textual queries. This year we have applied a strategy for deciding when include the information extracted from MeSH. The experiments carried out show that our machine learning approach to determine when to perform expansion did not resulted in any improvement over our base line, that where no expansion was performed at all.
This paper presents the system developed by the SINAI research group at the ImageCLEF medical retrieval task 2010.
The goal of the medical task is to retrieve relevant images based on an image query[
The e cient access to multimodal information is becoming more and more di cult every day. For this reason, it is necessary to develop search strategies for easier retrieval of useful information. One of these strategies includes the use of linguistic resources in order to improve the access and management of information by expanding queries in information retrieval systems, enriching the databases semantically or extracting unknown data from collections.
In previous years we have experimented with the expansion of the queries with medical
ontologies[
This year, our main goal is to study the expansion of the collection and queries with the same ontology MeSH, applying a strategy for deciding when include the information extracted.
This year, the same collection as 2009 and 2008 is used but with a larger number of images. The data set used contains all images from articles published in Radiology and Radiographics including the text of the captions and a link to the html of the full text articles, more than 77,000 images.
Previous years we developed a system that test di erent aspects, such as the application of
Information Gain in order to improve the results[
The following section describes the strategy for deciding when include the information extracted. In Section 3, we explain the experiments and obtained results. Finally, conclusions are presented in Section 4. 2
We have used the MeSH ontology to expand the queries included in the GoldMiner collection. First, we have extracted two types of terms from the MeSH descriptors:
MeSH Heading (MH), which is a term composed by one or more words. Each record contains only one MH term.
Entry, which is composed of one or more words too. Each record contains several Entry terms. These terms are a di erent way of writing the MH term, that is, they are synonyms.
This set of terms of a same descriptor constitutes a bag of terms. We have used the bags of terms to expand the queries. If all the words of a term are in the query, we generate a new expanded query by adding all the terms in this bag. To compare the words of a particular term and those of the query, we rst put all the words in lowercase without removing stopwords. It does not matter the order in which these component words occur in the query.
In order to reduce the number of terms available to expand the query, we have only used those
that are in the MeSH categories [A] Anatomy, [C] Diseases, and [E] Analytical, Diagnostic and
Therapeutic Techniques and Equipment[
Given the hardness of the \when to expand" problem, a Machine Learning perspective was followed. Using the gain or loss in MAP as a label, with a true value as label when the MAP increased with the expansion and a false value as label when the MAP decreased, the direction was to generated models in order to determine, automatically, when to expand or not.
The well known learning algorithms Support Vector Machine (SVM)[
By using the data in 2008 and 2009 ImageCLEFmed campaigns, we have carried out several experiments. We found that there were no clear correlation (slightly more than 0,20 in the Pearson's coe cient) between the success of the expansion for a query and the number of relevant documents found in associated qrels lists. Therefore, the expansion seemed not related to the expected number of relevant documents.
Despite the hardness of the \when to expand" problem, a Machine Learning perspective was followed. Using the gain or loss in MAP as a label, with a true value as label when the MAP 1http://www.nlm.nih.gov/mesh/ 2http://www.nlm.nih.gov/research/umls/ increased with the expansion and a false value as label when the MAP decreased, the direction was to generate a model in order to determine, automatically, when to expand or not.
Based on the Ranking Status Value (RSV) returned by the retrieval engine for both queries (expand and original): the RSV values of the top 100 documents in each of the two retrieved lists are used as features for computing the model. Therefore, each query is a sample with 200 features. Feature weights (frequencies in the case of word-based features, RSV in the second case) were normalized using Z-transformation and the dimensionality were reduced by applying Singular Value Decomposition (SVD).
Run
MAP No expansion 0.2764 Always expand 0.2616
Expand sometimes 0.2672
Results are shown in Table 1. As we can see, there is a very small variation in MAP among di erent strategies. Nevertheless, this year our expansion approach did not improved over the base line. 4
The machine learning approach to determine when to perform expansion did not resulted in any improvement over our base line, that where no expansion was performed at all. It is clear that our expansion strategy should be studied in depth, as the behaviour of the expansion over queries is not consistent and seems not related to the gain in MAP.
This work has been partially supported by a grant from the Spanish Government, project TEXTCOOL 2.0 (TIN2009-13391-C04-02), a grant from the Andalusian Government, project GeOasis (P08-TIC-41999), and a grant from the University of Jaen, project RFC/PP2008/UJA-08-16-14 and project UJA2009/12/14.