In general, the input to the semantic stage of analysis may be viewed as being a set of possible parses of the sentence, and information about the possible word meanings. The aim is to combine the word meanings, given knowledge of the sentence structure, to obtain an initial representation of the meaning of the whole sentence. The hard thing, in a sense, is to represent word meanings in such a way that they may be combined with other word meanings in a simple and general way.
Its not really possible even to present a simple approach to semantic analysis in less than half a lecture - I can only outline some of the problems and show what the output of this stage of analysis may be (though in general there are many different semantic theories and representations, just as there are many different grammar formalisms).
First, lets go back to our syntactically ambigous sentences and see how semantics could help:
If we have some representation of the meanings of the different words in the sentence we can probably rule out the silly parse. We might look up ``banana'' (maybe in some frame or semantic net system) and find that it is a fruit, and fruits generally don't fly. We might then the able to throw out the reading ``flies like a banana'' if we made sure that sentences which mean ``X does something like Y'' require that X and Y can do that thing!
Sometimes ambiguity is introduced at the stage of semantic analysis, for example:
Did John go to the river bank or the financial bank? We might want to
make this explicit in our semantic representation, but without contextual
knowledge we have no good way of choosing between them. This kind of
ambiguity occurs when a word has two possible meanings, but both of them
may, for example, be nouns.
To obtain a semantic representation it helps if you can combine the
meanings of the parts of the sentence in a simple way to get at the meaning
of the whole (The term
compositional semantics refers to this process). For those familiar
with lambda expressions, one way to do this is to represent word
meanings as complex lambda expressions, and just use function
application to combine them. To combine a noun phrase ``John'' and a verb
phrase ``sleeps'' we might have:
Verb phrase meaning:
Noun phrase meaning: john.
Apply VP meaning to NP meaning: =
The output of the semantic analysis stage may be anything from a semantic net to an expression in some complex logic. It will partially specifies the meaning of the sentence. From ``He went to the bank'' we might have two possible readings, represented in predicate logic as:
Good representations for sentence meaning tend to be much more complex than this, to properly capture tense, conditionals, etc etc, but this gives the general flavour.