In this paper we present the system used to solve the challenges of the Social Event Detection (SED) task at MediaEval 2011 challenge.
In a social event retrieval context, a query can be
splitted and mapped in three di erent parts according to the
general parameters characterizing an event: (
The queries are submitted to the search engine over the dataset. In our work, we use Solr 4 search engine to index the dataset and perform the search. The search is done as a 1http://dbpedia.org/ 2http://jena.sourceforge.net/ 3http://www.last.fm/api 4http://lucene.apache.org/solr/ mix of spatial (by using latitude and longitude values) and textual search. The data are indexed based on the textual metadata, including Title, Description and Tags. The search is then performed over all the three di erent metadata. In particular for Challenge 2 the queries are boolean queries with all the terms in AND, while in the Challenge 1 these conditions are more relaxed and the terms of each query are composed with the boolean operator OR. The reason for this is that in this challenge, a categorizer is provided as next step to lter out non-relevant retrieved occurrences.
The input of this block is a list of pictures with their metadata. This module is used only for Challenge 1 to extract pictures related to a soccer event. The categorization is performed over the three textual metadata for each picture, i.e., Title, Description and Tags. The di erent runs will exploit the descriptivity of each kind of metadata in the categorization process (see Section 3). To categorize the pictures the SemanticHacker API 5 over the di erent textual metadata was used. The categories produced are based on the Open Directory Project 6. The pictures are ltered by only keeping those categorized with a category that has radix Sports/Soccer.
The previous block returns a set of ltered pictures (Challenge 1 ) related to soccer events or pictures taken in the venues speci ed in the search step (Challenge 2 ) and grouped based on the venues. In this step the temporal information will be used to group the temporal related pictures. In that way the resulting clusters are nally grouped according to their temporal and locational information. To perform the clustering process the Quality Thresold Clustering (QT) algorithm is used [1]. This algorithm does not require to specify in advance the number of clusters and even it is computationally expensive, it is used only on retrieved documents. The resulting clusters may be semantically related and belonging to the same event. To merge semantically similar clusters a graph is built, where the nodes of the graph are the clusters, and two nodes are connected if they share at least a tag representing a named entity of an event or of an artist. To extract the named entities, we use the tags and submit them as queries to LastFM for the artist names and DBPedia for event names. Clusters are merged by nding the connected component as in [3].
The resulting clusters may be incomplete, i.e. the dataset
may contain other pictures related to the event clusters
extracted but not retrieved in the search step. The re
nement module is used here to query the dataset by using the
(
In this section, we present the di erent runs and their evaluation over di erent metrics. Table 1 provides a summary of our results.
Two di erent runs were performed in the rst challenge. In the rst run (Run 1 ) all the work ow of the system is performed excluding the re nement step and semantic merge between clusters. For Run 1 the categorization is performed by using only Tag metadata, while for the second run (Run 2 ), we also include the Title and Description metadata. From the results obtained (see Table 1) we can observe that including other metadata than tags resulted in a decrease of precision, probably due to the lack of descriptiveness of the other metadata.
For this challenge, three di erent runs were performed. The Run 1 (the baseline run) executed the algorithm without including the semantic merge and re nement steps. In both Run 2 and Run 3 the semantic merge and re nement steps were performed. Semantic merge was done by considering each cluster represented by the named entity representing events or artists. Moreover in Run 2, re nement is performed by querying the top-100 tags and the temporal range in which each cluster is closed. In Run 3, we used the entity names representing artists or events extracted from the set of tags of each cluster.
Precision
Recall F-Measure
NMI
We have presented a system to extract events for the given two challenges. As described in this paper, the best result in terms of precision was obtained in the rst challenge by using only the tags for the categorization step, while the other evaluation measure were better when using all the textual metadata. In the second challenge the best result was obtained using the complete work ow of the algorithm, i.e. using re nement step, in particular using the entity names in the re nement query for each cluster. Our future experiments, especially for the rst challenge, will include the use of the re nement step and semantic merge over the totality of the results (instead of applying it over groups of results coming from the query).