Extracting Relations from Text: From Word Sequences to Dependency Paths - Natural Language Processing and Text Mining - page 36

targetpredicates P 1 ,P 2 , ..., P n such that e 1 is an argument of P 1 , e 2 is an argument
of P n , and any twoconsecutive predicates P i and P i +1 share a commonargument
(where by “argument” we mean both arguments and complements).
(1) He had no regrets for his actions in Brcko .
his → actions ← in ← Brcko
(2) U.S. troops today actedforthefirsttime tocapture an alleged
Bosnian war criminal, rushing from unmarkedvans parkedinthe
northern Serb-dominated city ofBijeljina.
troops → rushing ← from ← vans → parked ← in ← city
(3) Jelisiccreatedanatmosphereofterroratthe camp by killing,
abusing and threatening the detainees .
detainees → killing ← Jelisic → created ← at ← camp
detainees → abusing ← Jelisic → created ← at ← camp
detainees → threatning ← Jelisic → created ← at ← camp
detainees → killing → by → created ← at ← camp
detainees → abusing → by → created ← at ← camp
detainees → threatening → by → created ← at ← camp
Fig. 3.5. Relation examples.
3.3.2 Learning with Dependency Paths
Theshortest path betweentwoentities in a dependency graph offers avery con-
densedrepresentation ofthe informationneededto assess their relationship. Ade-
pendency path is representedas a sequence ofwords interspersedwitharrows that
indicate theorientation of eachdependency, as illustratedinTable 3.1. These paths,
however, arecompletely lexicalized and consequentlytheir performance will belim-
ited by data sparsity. Thesolutionis to allow paths to use both words and their
word classes, similar with the approachtakenforthesubsequence patterns in Sec-
tion3.2.1.
Theset offeatures can thenbe definedas a Cartesian product overwords and
word classes, as illustratedinFigure 3.6forthe dependency path between'protesters'
and 'station' in sentence S 1 .Inthis representation,sparse or contiguoussubse-
quences ofnodes along thelexicalizeddependency path (i.e., path fragments) are
includedas features simplybyreplacing the rest ofthe nodes with their correspond-
inggeneralizations.
Examples offeatures generatedbyFigure 3.6are “protesters → seized ← sta-
tions,” “Noun → Verb ← Noun,” “Person → seized ← Facility,” or“Person
→ Verb ← Facility.” The total number offeatures generatedbythis dependency
path is 4 × 1 × 3 × 1 × 4.