These constraints serve two purposes: (i) to eliminate incoherent extractions, and (ii) to reduce uninformative extractions by capturing relations phrases expressed by a Verb-Noun combination, such as light verb constructions (LVCs).
There are some relation phrase datasets, such as Patty [19] and ReVerb [13] that can be used for this purpose.
Each relation phrase (rel_{overline{v_iv_j}}) (in edge (overline{v_iv_j}) of (Q^mathrm{S})) is also mapped to a list of candidate predicates and predicate paths.
(Semantic Relation) A semantic relation is a three-tuple (langle rel, arg1,aranglengle), where rel is a relation phrase in the paraphrase dictionary D, arg1 and arg2 are the two argument phrases.
In the running example of Fig. 9, (langle)"be married to," "who," "actor"(rangle) is a semantic relation, in which "be married to" is a relation phrase, "who" and "actor" are its associated arguments.
The goal of an Open IE system is to obtain the largest number of correct triples (arg1, rel, arg2) for any relation in the text, where arg1 and arg2 are the arguments of the relation and rel is a relation phrase.
(Semantic Query Graph) A semantic query graph, denoted as (Q^mathrm{S}), is a graph in which each vertex (v_i) is associated with an argument and each edge (overline{v_iv_j}) is associated with a relation phrase, (1le i,j le |V(Q^mathrm{S})|).
According to Etzioni et al. [7], incoherent extractions are 'the cases where extracted relation phrase has no meaningful interpretation,' while uninformative extractions are those which 'omit critical information.' Only results from the subset of the data where the judges agree were considered for results evaluation.
Table 6 Comparison between assessment rounds: triples and relation phrases First round Second round Triples 311 2,359 Relation phrases 119 87119
