s collection DBpedia. Since the metadata is short for retrieval by query words, we decided to expand the metadata using a typical query expansion method. In our experiments, we use the Rocchio algorithm for document expansion. Our best run is in the 26th rank of all 57 runs which is under our expectation, and we think that the main reason is that our document expansion method uses all the words from the metadata documents which contain words which are unrelated to the content of the images. Compared with our text retrieval baseline, our best document expansion run improves MAP by 11.17%. As one of our conclusions, we think that the document expansion can play an e ective factor in the image metadata retrieval task. Our content-based image retrieval uses the same approach as in our participation in ImageCLEF 2008.
This is DCU's rst participation in the WikipediaMM task of CLEF. This task aims to nd relevant images based on the query text and image. In the image collection, every image is associated with a metadata le. The le usually contains the description, copyright, author, camera parameters, date and location of the image; for the topics, the data consists of the image and the query text. So this task can be performed by text retrieval or content-based image retrieval or by a combination of these two methods. Our main research e orts are on the text retrieval. Since the useful information in the metadata is usually very short, this text retrieval task is di erent from the ad-hoc retrieval for news or articles, and we call it short-length documents retrieval.
Since the metadata les contain only very few words to describe the content of the images,
we decided to expand the metadata document from an external resource, the Wikipedia abstracts
collection DBpedia1. Document expansion is quite similar to query expansion in information
retrieval research, and we use the Rocchio algorithm as our document expansion method [
In our approach, we use the Lemur toolkit2 as the retrieval system. We have tested all di erent retrieval models in the Lemur toolkit and found that the tf-idf model performs well in this task. Our formal runs employ Lemur's tf-idf method as the retrieval model.
We parse the metadata to be used for indexing and document expansion. For the text part of the topics, we directly extract all words to form a query. The system overview is shown in Figure 1.
For the queries, we choose to expand queries based on external resources or not which could lead di erent results; for the metadata document, we will expand it when it's length is less than a threshold L < 200. We found long-length documents in the metadata documents contain more noise to get useful relevant feedback terms.
In the following parts, we will describe the preprocessing, the retrieval model used in the task, and the document expansion algorithm. 2.1
For WikipediaMM 2009, we use the following data: the topics, the metadata collection and DBpedia. All these collections are preprocessed to be used in our task. For the topics, we select the title part as the query; for the metadata collection, the text is selected as the query to perform the document expansion. An example is shown in Figure 2: 1http://dbpedia.org 2http://www.lemurproject.org/
For every metadata document, after parsing we set the remaining text as the query.
For the DBpedia, we use the text as index and remove the stop words computed from itself. An example document from DBpedia is: <DOC> <DOCNO>1969 Paris Open</DOCNO> <TEXT>The 1969 Paris Open was a professional tennis tournament played on indoor carpet courts. It was the 2nd edition of the Paris Open (later known as the Paris Masters). </TEXT> </DOC>
The English DBpedia includes 2,787,499 documents corresponding to a brief form of a Wikipedia article. We compute 500 stop words from DBPedia and remove all the stop words before indexing it. 2.2
After testing di erent information retrieval models on the text based image retrieval task, we found that the tf-idf model outperforms state-of-art models such as Okapi BM25 or language modeling in Lemur toolkit. So we choose the tf-idf model as our baseline retrieval model in this task. The document term frequency (tf ) weight we use in tf-idf model is: tf (qi; D) =
k1 f (qi; D) f (qi; D) + k1 (1 b + b lldc ) (1) f (qi; D) is the frequency of query term qi in Document D, ld is the length of document D, lc is the average document length of the collection, and k1 and b are parameters set to 1.2 and 0.75 respectively. The idf of a term is given by log(N=nt). N is number of documents in the collection and nt is the number of documents containing term t.
The query tf function (qtf ) is de ned similarly with a parameter representing average query length. The score of document D against query Q is given by:
i=1 s(D; Q) = X tf(qi; D) qtf(qi; Q) idf(qi)2 n (2) qtf is the tf for a term in queries and it's computed using the same method with the tf in documents. 2.3
Our document expansion method is similar to a typical query expansion process. We use the
pseudo-relevance feedback as our document expansion method with Rocchio's algorithm [
For every metadata document, after preprocessing we use the remaining text as the query. We retrieve the top 100 documents as the assumed relevant documents. With all the words from the returned top 100 documents we rst remove all the stop words. The stop words list is produced from the DBpedia document collection, and we compute the term frequency from the DBpedia collection and set the top 500 words as the stop words. For the top 100 relevant documents in DBpedia, we compute a word frequency list and remove the stop words and the original words from the query. We select the top ve words as the document expansion words.
The number of relevant documents for document expansion is higher than normal because the Wikipedia abstract corpus usually has very short documents. If we only used 10 or 20 as the assumed relevant documents, it would be very hard for us to get useful feedback terms from the relevant documents. Furthermore, the original metadata documents are short so we only select the top 5 terms as the feedback terms. Then the expanded terms will be added into the metadata document and the index is rebuilt. 3
For content-based image retrieval we make use of the following six global visual features de ned in the MPEG-7 speci cation:
Scalable Colour (SC): derived from a colour histogram de ned in the HSV colour space. It uses a Haar transform coe cient encoding, allowing scalable representation. Colour Structure (CS): based on colour histograms, the feature represents an image by both the color distribution (similar to a color histogram) and the local spatial structure of the colour.
Colour Layout (CL): compact descriptor which captures the spatial layout of the representative colours on a grid superimposed on an image.
Colour Moments (CM): similar to Colour Layout, this descriptor divides an image into 4x4 subimages and for each subimage the mean and the variance on each LUV color space component is computed.
Edge Histogram (EH): represents the spatial distribution of edges in an image. Edges are categorized into ve types: vertical, horizontal, 45 degrees diagonal, 135 degrees diagonal and non directional.
Homogeneous Texture (HT): provides a quantitative representation using 62 numbers, consisting of the mean energy and the energy deviation from a set of frequency channels.
Our work for visual querying was the same approach as undertaken in ImageCLEF 2008. For
a visual query, we take the topic images and extract from each their representation of the image
by each of the six features above. For each feature we query its associated retrieval expert (i.e.
visual index and ranking function) to produce a ranked list. The ranking metric for each feature
is as speci ed by MPEG-7 and is typically a variation on Euclidian distance. For our experiments
we kept the top 1000 results. Each ranked list was then weighted and the results from all ranked
lists are normalized using MinMax [
The weighting we employed was linear, using an approach where the weights are determined
at query-time. This approach is the same as used in our previous ImageCLEF experiments, with
an explanation found in [
We have submitted 5 runs to the WikipediaMM task including 4 runs by the text retrieval system and 1 run by image retrieval system. The main technique used in the text retrieval is: query expansion (QE), query expansion from external resource (QEE), document expansion from external resource (DEE). Our 4 runs are combinations of these techniques and the baseline run uses only the tf-idf IR model without additions. The results for English monolingual WikipediaMM 2009 task are shown in Table 1.
Run dcut df-baseline dcut df-dbpedia-qe dcut df-dbpediametadata-dbpediaqe dcut df-dbpediametadata-qe dcuimg
TXT IMG
BASELINE
DEE QEE+DEE+QE
DEE+QE BASELINE
MAP 0.1576 0.1685 0.1641 0.1752 0.0079
P@10 0.2600 0.2600 0.2378 0.2578 0.0244
Our best result ranks in the middle of all the o cial runs in the WikipediaMM 2009 task. Compared with our baseline result, document expansion plays an important role in our best result. Document expansion can improve the MAP from 0.1576 to 0.1685, but query expansion from external resource in combination with our methods does not show much improvement. Our image run gets bad results due to some computation error which will xed in the future research. 5
We presented our system for the WikipediaMM task of CLEF 2009 focusing on document
expansion. Document expansion has not been thoroughly researched for information retrieval. From the
past research, whether the document expansion can improve the retrieval e ectiveness or how to
improve it is not obvious [
From this task, our main nding is that the document expansion can improve the retrieval e ectiveness much when the document length is short. On the other hand, query expansion from the external resource does not improve performance since the query text usually is very accurate and does not need to be expanded with more words. 6
This research is supported by the Science Foundation Ireland (Grant 07/CE/I1142) as part of the Centre for Next Generation Localisation (CNGL) project.