Extracting Relations from Text: From Word Sequences to Dependency Paths - Natural Language Processing and Text Mining - page 35

Table 3.1. ShortestPathrepresentation ofrelations.
Relation Instance Shortest Path in UndirectedDependency Graph
S 1 :protesters AT stations protesters → seized ← stations
S 1 :workers AT stations workers → holding ← protesters → seized ← stations
S 2 :troops AT churches troops → raided ← churches
S 2 :ministers AT churches ministers → warning ← troops → raided ← churches
graph toestablishing the relationship R ( e 1 ,e 2 ) is almost exclusively concentrated
in theshortest path between e 1 and e 2 in the undirectedversion ofthe dependency
graph.
If entities e 1 and e 2 are argumentsofthesame predicate, thentheshortest path
betweenthemwill pass throughthe predicate, whichmaybeconnecteddirectlyto
the twoentities, or indirectlythroughprepositions. If e 1 and e 2 belong to different
predicate-argument structures that share a commonargument, thentheshortest
path will pass throughthis argument. This is thecase with theshortest path be-
tween'stations' and 'workers' in Figure 3.4, passing through'protesters,' whichis
an argument commonto both predicates 'holding' and 'seized'. InTable 3.1, we
showthe pathscorresponding to the four relationinstances encodedinthe ACE
corpus forthe twosentences from Figure 3.4. All these pathssupport the Located
relationship.Forthefirst path, it is reasonable to infer that if a Person entity
(e.g., 'protesters') is doingsome action (e.g., 'seized') to a Facility entity (e.g.,
'station'), thenthe Person entity is Located at that Facility entity. The second
path captures the fact that thesame Person entity (e.g., 'protesters') is doing two
actions (e.g., 'holding' and 'seized') , one actionto a Person entity (e.g., 'workers'),
and theotheractionto a Facility entity (e.g., 'station'). Areasonable inference in
thiscase is that the 'workers' are Located at the 'station'.
In Figure 3.5, weshowthree moreexamples ofthe Located (At) relationship
as dependency pathscreatedfrom oneortwo predicate-argument structures. The
second example is an interestingcase, as it illustrates how annotationdecisions
are accommodatedinour approach. Using areasoningsimilar with that fromthe
previous paragraph, it is reasonable to infer that 'troops' are Located in 'vans,' and
that 'vans' are Located in 'city'. However, because 'vans' is not an ACE markable,
it cannot participate in an annotatedrelationship. Therefore, 'troops' is annotated
as being Located in 'city,' whichmakes sense due to the transitivity ofthe relation
Located.In our approach, thisleads toshortest paths that pass throughtwoor
more predicate-argument structures.
Thelastrelation example is a case where thereexistmultiple shortest paths
in the dependency graph betweenthesame twoentities - there are actuallytwo
different paths, with each path replicatedinto three similar paths due tocoordina-
tion. Our current approach considersonly oneoftheshortest paths, nevertheless it
seems reasonable to investigate using all ofthemas multiple sources of evidence for
relation extraction.
There may becases where e 1 and e 2 belong to predicate-argument structures that
have no argument in common. However, because the dependency graph is always
connected, we areguaranteedtofind a shortest path betweenthe twoentities. In
general, weshall find a shortest sequence ofpredicate-argument structures with