In this paper we introduce SemSor, a system developed especially for the analysis of emergency situations. It constantly collects information from sources of the Social Web, maps it to unique resources in the Semantic Web and uses the annotated information as basis for the situation analysis. If an emergency situation needs to get analyzed, four steps are required: First, all information that is already known about this situation must be entered in the SemSor-GUI. Second, the entered information needs to be mapped to resources in the Semantic Web. Third, using these resources as starting nodes, a spreading activation is applied along the relationships within the Semantic Web to nd relevant Social Web information. And fourth, the newly identi ed information is visualized according to di erent dimensions and can be ltered and explored by the user. In an iterative process, new insights can be used to re ne the query and thus improve the activated information until a comprehensive analysis of even complex situations is possible.
H.3.1 [Content Analysis and Indexing]: Indexing methods; H.3.3 [Information Search and Retrieval]: Information ltering, query formulation, relevance feedback; H.5.2 [User Interfaces]: Graphical user interfaces (GUI)
Analyzing emergency situations is di cult in cases where either the agent who does the analysis is not close by and thus is not able to directly hear or see what is going on, or the situation itself is distributed in space or time so that one single person alone has di culties in getting an overview. These cases hold true, e.g., in emergency operation centers (EOCs), where neither the agent who receives an emergency call is onside nor the person who has made the call has usually a comprehensive picture of the situation. However, having a comprehensive picture is especially important for the analysis of emergency situations in order to take the right actions and thus prevent all kinds of damage.
To overcome the di culties in analyzing emergency
situations, we propose an approach that combines the advantages
of the Social Web with those of the Semantic Web. The idea
is to scan Social Web entries, semantically annotate their
content and use spreading activation to nd exactly those
entries that are useful for the analysis of a speci c emergency
situation. The idea of combining Social and Semantic Web
to a Social Semantic Web has already been described e.g.
in [
The general idea of spreading activation in semantic graphs
has rst been introduced in [
In this paper we introduce an approach that applies
spreading activation in external semantic datasets and thus saves
storage space and calculating capacities. Datasets in the
LOD cloud [
In the following we rst describe the general SemSor architecture with all the components and steps that are required for the analysis of emergency situations and provide further details to each of the steps afterwards. This includes the crawling and annotating of Social Web entries, the initial query formulation, the spreading activation, the visualization and ltering, and the interactive query re nement. At the end of the paper a conclusion and an outlook on future work is given.
The SemSor System constantly scans Social Web sources,
e.g. Twitter, Flickr and YouTube, for new entries (Fig. 1,
A) and semantically annotates their textual content. Terms
with a distinct meaning, e.g. geographical or temporal
references, are therefore mapped to unique semantic resources
of datasets in the LOD cloud, e.g. DBpedia [
Once a certain situation needs to be analyzed (Fig. 1,
F), e.g. because of an incoming emergency call, everything
that is known about this situation must also be mapped
to unique semantic resources (Fig. 1, C). These resources
are then used as starting nodes for the spreading activation
[
In order to use information that is contained in Social
Web entries to support the analysis of emergency situations,
the entries rst have to be extracted and annotated by the
crawler component of SemSor. Even though the Social Web
contains a huge amount of data, only a minimum of this
information needs to get stored in the SemSor database.
In a rst step, a broad multitude of Social Web data gets
collected and evaluated according to a precon gured
metric that determines the a priori relevance of each individual
entry. Within this metric, di erent properties of an entry
like its source, the date and time of its creation as well as
location-based data get extracted and serve as basis to
calculate a weighted importance rating. The weights of the
metric can be con gured according to the individual needs
of its users (e.g. a speci c emergency response team) and
provide a basic means to decide, which entries should be
kept and which can be deleted if computational- or
storageresources become short. Following the collection and a priori
evaluation, the Social Web entries are analyzed and certain
terms in their textual contents are automatically assigned to
unique resources in datasets in the LOD cloud by using
services like e.g. OpenCalais [
The SemSor system supports the whole situation analysis process, including the initial query formulation, the search for relevant information via spreading activation, the visualization and ltering of the results as well as mechanisms to iteratively re ne the query. 4.1
The method of query is based on common question schemes of emergency calls according to relevant aspects of a situation: "What has happened?", "Where did it happen?" and "Who is involved?" (Fig. 2, B). The agent is supposed to provide approximate answers to at least some of these questions and can further substantiate his query by providing boundaries for the temporal and spatial extent of the situation. In this process, the agent is assisted by an adaptive autocomplete feature, which will try to interactively map given search terms to resources in the Semantic Web. Throughout this process the de nitions of proposed resources are provided in pop-up windows, e.g. corresponding Wikipedia articles (Fig. 2, A), to help users especially in the disambiguation of ambiguous input terms. Based on this procedure the system is able to get a reliable handle onto the relevant nodes in the Semantic Web. During an emergency situation, like e.g. the 2010 Haiti earthquake, disaster agents can obtain a general overview of the situation by performing a broad search on keywords like "Earthquake" and "Haiti". Based on the interactive mapping of search terms to semantic resources the query is annotated by SemSor and connected with resources in the Semantic Web. The nodes of this framework serve as the initially activated nodes in the spreading activation procedure. 4.2
Based on the initial activation of the user de ned staring nodes (Fig. 3, A), a homogeneous spreading activation is applied along the relationships between the resources in the datasets. In order to automatically activate semantically related resources that might also be relevant for the situation analysis, the activation happens along the instancerelationship layer (Fig. 3, B) as well as the class-relationship layer (Fig. 3, C). Thus the resulting set of Social Web entries is not limited to those containing references to one of the user de ned starting nodes only, but also includes entries referring to resources that are within a distinct semantic radius around the user de ned starting nodes (Fig. 3, D); this facilitates nding information that is relevant for the analysis of a certain emergency situation.
The spreading activation in SemSor is implemented mostly as a remote process. On the client side, the process is only triggered and controlled but is run completely within external datasets on server side. Therefore SPARQL queries are sent to the datasets to nd resources related to the starting nodes, which are then scored according to the semantical and topological properties of their relationships.
Related resources are found based on an approach
described in [
All those Social Web entries are then activated according to topological and semantical aspects and thereby scored. In our implementation the extent of activation depends on three aspects: 1) The length of the relationship, e.g. the length between the Flickr entry and the starting nodes is two, 2) the connection types within the relationships, e.g. "district of" or "capital", and 3) the classes of the intermediate resources, e.g. "Port-au-Prince" is a city. It is also possible to de ne certain connection types or classes that should not be used to spread the activation, which is useful if someone is not interested in relationships that contain certain instances or connections.
Based on their activation, the Social Web entries are visualized in the SemSor GUI. While the search still continues, all Social Web entries that were already discovered by the activation are presented through the result browser in different user-selectable views (e.g. tabular, map and statistic view). Every view provides the opportunity to obtain pictures, videos and other user generated content related to the situation (Fig. 2, D). The standard view is a listing of entries sorted by their individual semantic relevance (Fig. 2, C). In order to get an overview of the spatial distribution of possibly relevant entries, the agent can view them on a map that can also be used to formulate geographical constraints, e.g. to show only results that refer to a certain geographic region (Fig. 2, E). To further explore the set of results a time line o ers an overview over the temporal distribution of events (Fig. 2, G) and allows to formulate temporal constraints, e.g. to show only results that refer to a speci c period of time. To further explore the quality and diversity of results the agent has the opportunity to analyze diagrams that show the composition of chosen result-subsets by author, location or tag-categories. By acquiring these initial impressions, the agent can further asses the nature and extent of the situation and initiate subsequent steps. 4.4
At this point the advantages of the interactive features in SemSor come into play. If the resulting set generated by the initial spreading activation (Fig. 4, A and B) and diminished by the user de ned lters is yet not su cient, the agent can further re ne and expand his initial query by rating single result items on a continuous scale (Fig. 4, C) and by dragging additional tags from the items to the query elds. High ratings of some Social Web entries can then lead to the activation of semantic resources which are directly connected to those entries (Fig. 4, D). Thus after rating the items, the agent is given the possibility to restart his search, but this time, the spreading activation will execute starting also from the newly activated items (Fig. 4, E). Through this "pollination" and subsequent activation of remote nodes in the semantic graph the agent is given the chance to discover relevant regions and new Social Web entries that were not included or even near his initial query (Fig. 4, F). Based on the lters and the interaction procedure, the agent is able to cope with the enormous ood of Social Web data that can be found in connection with emergency situations and make bene cial use of them. 5.
In this paper, we described an approach that combines the Social and the Semantic Web in order to support the analysis of emergency situations. Exploiting information from the Social Web is especially useful when: (1) the situation is distributed in space or time, (2) rst responders are not on site, or (3) the situation cannot or only partly be observed; e.g. this can be the case if a situation takes place within a building or is hidden behind some obstacles.
The automatic semantic annotation of entries in the Social Web as well as the interactive mapping of entered search terms to unique resources in the Semantic Web allows information to be found not only by string matching but also according to its meaning. Therefore spreading activation is applied in external semantic datasets that o er both up-todate and comprehensive information on all kinds of topics.
Because of the outsourced spreading activation and the minimum information that is needed to store the annotated Social Web entries { an entry is represented by its URL and the URIs of the annotated resources only { SemSor is able to handle the huge amounts of available data and nd information relevant for the analysis of a certain emergency situation. The found information is visualized in multiple views and can be explored and ltered by the agent based on individual information needs. If the result set is not yet su cient, the query can interactively be expanded or narrowed down. Single entries can be rated as relevant or as irrelevant which changes the starting nodes and thus can result in other entries to be found by the spreading activation. The query can be re ned until a su cient analysis of the emergency situation is possible.
In its current implementation, the SemSor system is most suitable to analyze current or past emergency situations (cp. Fig. 2). Since many people use the Social Web to comment on emergency situations, relevant information is available even while a situation is happening. Together with the fact that only seconds are required from the time a new comment is uploaded to when it can be found in SemSor, it is already possible to facilitate the analysis of current and past emergency situations.
However, analyzing current or past emergency situations often cannot prevent them from happening. In order to prevent emergency situations a preventive analysis is required. First signs of a forthcoming emergency situation need to be detected and interpreted in the right way so that the right actions can be initiated. Besides the agent triggered search, this would require SemSor to automatically scan the Social Web entries for new topics and trends and iteratively produce an overview of the current situation. If certain topics get popular or unusual changes can be detected an alarm could be raced automatically that could force an agent to check the situation and decide on the right actions to prevent a possible emergency situation from happening. 6.
The prototypical implementation of our approach, the SemSor system, was developed within a student project by Clint Banzhaf, David Schmid, Dominik Jackle, Edwin Puttmann, Jochen Seitz, Johannes Dilli, Marijo Macet, Nico Ploner, Ste en Bold, Stephan Engelhardt and Thomas Michelbach. We would like to thank them for their excellent work. Furthermore we thank Bernhard Schmitz, Michael Worner and Thomas Schlegel for co-supervising the project. 7.