In order to compute probabilities of Orange and live being related, you’d have to have a lot of data lying around (sort of like Google’s n-grams packages of up to 5 words).  That would never fit on a cellphone as a standalone chat engine.

I won’t pretend that I can exactly explain how e.g. an entity extraction engine works, but from playing with them I now they can identify entities e.g. cities even if they don’t have that city in their database.
From the limited understanding I have I thought based on the ontology in wordnet, an engine could identify what a word would be and in what kind of topic it should belong. E.g. cities or a persons name or in this case orange is a colour as it is an adjective etc.

OK. Entity extraction for a specific series of entities types can certainly be done (one could write a chatscript script to do it). However, they are almost always assuming the case of the incoming text is correct, which in chat is rarely true. And extracting an entity is not the same as deciding which meaning of orange to use given incoming text. I had thought you were seeking generalized resolution of which meaning of a word was being used, which is a much more complex problem than finding geographic names in some text.

Not to mention that determination of the semantics of a given word is pretty well exclusively based on context, which is actually beyond the scope of the ChatScript engine itself, and lies almost entirely on the botmaster to create the rules to determiner said context… Or should be, at least.