The following two sentences are synonymous:
Bill went into the house. Bill entered the house.
This means that they encode the same underlying conceptual structure:
go(e,bill,t) to(t,i) in(i,h) house(h)
This is equivalent to Jackendoff's notation, where the referential indices are left implicit:
[GO([BILL],[TO([IN([HOUSE])])])]
The mapping between sentences and concepts is ensured by the following CCG lexicon:
went :- S1\NP2/PP3 : go(1,2,3) entered :- S1\NP2/NP3 : go(1,2,4), to(4,5), in(5,3) Bill :- NPbill into :- PP1/NP2 : to(1,3), in(3,2) the house :- NP1 : house(1)
The conceptual predicates are organised into an ontology:
entity / / \ \ / / \ \ / / \ \ thing place path event | | | | | | | | house in to go
This is to be understood as follows:
- Every entity belongs to exactly one of the following four categories: thing, place, path or event.
- Every 'house' is also a 'thing'.
- Every 'in' is also a 'place'.
- Every 'to' is also a 'path'.
- Every 'go' is also an 'event'.
More formally:
- ∀x. thing(x) xor place(x) xor path(x) xor event(x)
- ∀x. house(x) -> thing(x)
- ∀x,y. in(x,y) -> place(x) and thing(y)
- ∀x,y. to(x,y) -> path(x) and place(y)
- ∀x,y,z. go(x,y,z) -> event(x) and thing(y) and path(z)
Note that we do NOT NEED to decompose lexical concepts like 'into' or 'enter':
- ∀x,y. into(x,y) -> ∃z. to(x,z) and in(z,y)
- ∀x,y,z. to(x,y) and in(y,z) -> into(x,z)
- ∀x,y,z. enter(x,y,z) -> ∃w. go(x,y,w) and into(w,z)
- ∀x,y,z,w. go(x,y,z) and into(z,w) -> enter(x,y,w)
In other words, 'into' is a subtype of 'in', and 'enter' is a subtype of 'go'.
The semantic interpretation process takes place as follows:
- The input sentence S is parsed using the lexicon/grammar, yielding a lexico-syntactic conceptual structure C (or a set of these if S is ambiguous).
- C is tested for 'consistency' with the ontology.
No comments:
Post a Comment