Legalo

From STLab

Current revision

Legalo

Subgraphs identification algorithm (implemented by the Semantic Web property extractor)

   algorithm legalo is
       input G: a RDF graph  from FRED
       output S: a set of RDF triples expressing semantic relations between named entities recognized in G
   
       let E the set of named entities in G
       let C the set of all named entity pairs combinations including permutations
   
       for each named entity e1 in E
           for each named entity e2 in E
               if e1 != e2'
                   add <e1, e2'> to C
   
       let P a set of tuples <e1, e2, sub-graph>
       let DFS stand for depth first search
   
       for each pair <e1, e2> in C
           directPaths ← DFS in G retrieving the shortest direct paths between e1 and e2
           if directPaths is not empty
               for each directPath in directPaths
                   add <e1, e2, directPath> to P
           else  
               for each node of type dul:Event event in G
                   leftSubGraph ← DFS in G for retrieving the shortest direct paths connecting e1 and event
                   rightSubGraph ← DFS in G for finding the shortest direct paths connecting event and e2
                   if (leftSubGraph is not empty) && (rightSubGraph does not contain agentive roles)
                       add <e1, e2, rightSubGraph> to P
       
       let S a set of RDF triples
       
       for each tuple <e1, e2, sub-graph> in P
           property ← produce semantic web relation based on sub-graph
           add RDF triple e1 - property - e2 to S

Producing semantic web relations

Legalo produces a semantic web relation property based on each sub-graph included in P (cf. legalo algorithm). The label of property is generated by concatenating the labels included in the sub-graph according to a number of heuristics derived by incremental empirical observations. These heuristics are continuously evolving. Currently, the main ones are summarized as follows:

• Given a direct path between e₁ and e₂, the label of property will be produced by concatenating the arc labels in the path and the more general types' labels of the nodes in the path. See example below:

   e1 = dbpedia:Felix ; e2 = dbpedia:Ringo_Starr
   
   dbpedia:Felix :memberOf dbpedia:All_Star_Band
   dbpedia:All_Star_Band :of dbpedia:Ringo_Starr
   dbpedia:All_Star_Band rdf:type :Band
   
   would generate
   
   label(property) = memberOfBandOf
   
   and
   
   dbpedia:Felix :memberOfBandOf dbpedia:Ringo_Starr

• Given a sub-graph connecting e₁ and e₂ through a node event of type dul:Event, the label of property will be produced by concatenating the arc labels in the path and the more general types' labels of the nodes in the path, connecting event to e₂. The arc labels of events' thematic roles are not used as they show in the graph, but they are associated either with an empty label or with a lexicalization defined in a mapping table. See example below:

   e1 = dbpedia:John_McCarty ; e2 = dbpedia:LISP
   
   :invent_1 vn.role:Agent dbpedia:John_McCarty
   :invent_1 vn.role:Product dbpedia:Lisp
   :invent_1 rdf:type dul:Event
   dbpedia:Lisp rdf:type :Language 
   
   would generate
   
   dbpedia:John_McCarty :inventLanguage dbpedia:Lisp

• Examples of mappings between thematic roles labels and their lexicalizations:
  • vnrole:Beneficiary → for
  • vnrole:Instrument → with
  • vnrole:Destination → to
  • vnrole:Topic → about