The free-form nature of triplestores offers a lot of flexibility for constructing databases, but that freedom can also make it less obvious how to find arbitrary data for retrieval, errorchecking, or general browsing. Gruff is a graphical triplestore browser that attempts to make data retrieval more pleasant and powerful with a variety of tools for laying out cyclical graphs, displaying tables of properties, managing queries, and building SPARQL and queries as visual diagrams.
have many different domains, DBPedia and Freebase are
good examples. The LODC vision is that the single data
representation for every type of knowledge (RDF) will
make it easy to combine datasets and dramatically increase
their value. In practice, the exploitation and exploration of
these data sources can be far from trivial. Two papers
recently made this point:
1. Heath [
But lets us go back to why it is hard to browse the LOD
cloud. To begin, most of the time an interesting problem
requires you to combine datasets. A typical example would
be from the biomedical domain. Say you take the RDF
version of Clinical Trials [
EXPLORATION AND DISCOVERY So what kinds of analysis would you do with these datasets: Example 1 - Automatic Discovery. Sometimes users just want to know if it would be possible to find any connection between two interesting instances from two different datasets. A good example would be to find all the links you can find between a particular target (Cytochrome c3) and say Morphine Sulfate given all the predicates or a subset of the predicates in the database. What is non-trivial about this? Well most Triplestores do not provide automatic discovery functions that would connect these instances or provide ways to specify which predicates to use while doing automatic discovery.
Example 2 - Write New Queries. Say a user has two instances from two different datasets (like Example 1) and you want to write a SPARQL query that links them together. Again, what is non trivial about this? In some sense a user has to know the datasets before they can write the queries, that is, you have to understand the predicates and the names of the classes that will link them together. Unfortunately the datasets from the example above lack domain and range restrictions so it not straight forward to explore the schema space. Additional usability problems are that you have to know the namespaces, the fragments of the namespaces, and very often the type of the objects. Example 3 - Basic Data Mining: Say you have an instance of a clinical trial and now you want to find all the clinical trials that discuss the same kind of drugs, diseases and targets. The non trivial part is that most triple stores do not offer you advanced ways to perform stored procedures on large datasets that could do this data mining for you. In order to make exploration of multiple datasets easier we developed Gruff, an advanced Graphical User Interface for working with RDF Data in a triplestore.
Gruff provides a full set of graph analysis capabilities to help users explore the LODC. In the most simple case a user selects two resources in Gruff, the predicates the graph algorithm needs to explore, and Gruff will find all connections between the two instances. The end result on the screen can even be saved as an NTriple file. The graph algorithms all take as input first class functions called generators. The contract for a generator is that it takes one node as an input, and will generate a set of nodes as output. Gruff will create these generators for you if you select the predicates that you want to explore.
In other cases the users want to investigate something more sophisticated, like conditionally exploring predicates per type of node (or resource), starting with the important predicates first. In this case the user needs to write some SPARQL or Prolog code and this can be facilitated with the visual query builder described below.
With Gruff you can first explore the graph over multiple link data sets completely visually, and once you find one pattern that you are interested in, you can easily build from that a visual query and find all the same patterns in the set of linked datasets.
Gruff provides several key capabilities for LODC exploration:
Triplestores specialize in relationships between things, and it can be difficult to get a feel for a tangled set of relationships.
Gruff's Graph View (figure 1) provides automatic layout of highly-cyclic graphs. Heuristics are used to make layouts of a modest number of nodes fast enough for interactive browsing. The algorithm specializes in keeping nodes from laying on top of unrelated link lines, even though all link lines are straight; those two features allow you to quickly spot all of the nodes that are linked with a particular node, and then to quickly spot all of the linked nodes of those nodes, and so on.
Once you have found one or more initial objects of interest, the familiar Table View (figure 2) can be used to jump to linked objects that are properties of other objects. The Graph View will add nodes and links that show a history of all of the paths that you traversed in the Table View. You can also incrementally add linked nodes directly in the Graph View. A triplestore can contain many objects that are not instances of a simple set of classes, and so it can be tricky to locate arbitrary objects of interest.
Gruff has a variety of ways for looking up arbitrary objects. These include (1) specifying an object textually as either its easy-to-type rdfs:label string, its full URI, or a brief URI that uses a namespace identifier; (2) selecting a node that has a selected rdf:type (selecting the type either from a single list or from an rdfs:subClassOf hierarchy); (3) finding all nodes that have property values containing arbitrary text, using free text lookup; and even (4) selecting from all sorted nodes.
When there are many choices to select from, Gruff manages the selection by using a series of pop-up menus that allow you to gradually hone in on a desired object. You use these menus by first selecting the starting character of an object's name, then the second character, and so on. Each menu displays only the choices that remain for the next character, with string completion for the remaining choices to the extent that's possible without ambiguity, allowing you to see increasingly recognizable choices.
Saving visual layouts and other views to file allows you to start back up quickly in a future Gruff session with a familiar set of "starter nodes", rather than specifying them again from scratch. The most recently saved or loaded views are always listed on the View menu for quick retrieval. You can also copy nodes from one view to another, for example, to avoid respecifying them.
While query languages like SPARQL are quite powerful, queries can become rather complex and therefore difficult to construct free of nesting mismatches and typos in URIs. Gruff's newest feature, the Graphical Query View (figure 3), allows creating queries as diagrams of nodes and links. A query diagram can include actual objects from the store, which you select as in other views, while other nodes and links represent query variables. Group graph patterns such as UNION and OPTIONAL groups can be laid out as grouper boxes that can be nested to any level, with proper nesting maintained automatically.
Once you've created a query diagram, Gruff will generate either SPARQL or Prolog code for the query. Seeing the generated query text (figure 4) can help you learn to write SPARQL or Prolog queries directly. You can also edit the generated text before performing the query. You can save queries either as graphical diagrams or as actual query text and load them any time later for modification into related queries. You can even generate a visual graph from query results, showing the network of all triples that are comprised of nodes and predicates that are in the query results and the query itself.
CONCLUSION In general, Gruff's tools are tightly integrated to allow browsing a triple-store in a variety of ways. Gruff is free and can be downloaded from www.franz.com/agraph/gruff/, where additional screen shots and demonstration videos can also be found.