It only dawned on me as I was writing that we’re facing many of the same problems in our named-entity detection and database linkage, which are just a kind of syntax and semantics.

The tie-in is that developing an annotation standard for a machine learning task is tantamount to developing a traditional kind of syntactic and semantic theory.

I’ve been meaning to write a longer post on this topic — corpus annotation coding standards are the new home of the kind of linguistics I always liked, which is more example driven than theory driven.

I didn’t even go into the fact that Wallace and Gromit are fictional, which in itself presents a non-trivial semantic puzzle.

[Punctuation and Semantics Exercise: Explain the alternate hyphenations of "old school" and the quotes in the first paragraph of this response.]

This may be absurdly old-school, but it seems to me that to have a named entity to detect in the first place you need three things.

- a name (we can safely assume that this is a subsequence of the document
it is in)
- an entity (what’s that?)
- the appropriate relation between the name and the entity (what’s that?)

Thus, in my view, any system that does named entity detection is in effect taking some
kind of position on the second two aspects. Either it works with an explicit
theory of these things (in which case, don’t forget to consult Montague and his successors), or it’s theory is implicit, and must be deduced from its
behaviour before it can be evaluated. In either case the question of exactly what we want from an NE detector seems worthwhile. And that, I expect, will
be deeply and irrevocably application dependent.

My interest and comment was in the context of NE detection, which is what the original article was about. From that point of view, I suspect any current algorithm would treat the text as two references and any one->two mapping would be non-trivial on or post co-reference stage.

They could have written

draft resolution A/56/L.28 as modified by Add.1

instead of

draft resolution A/56/L.28 and Add.1

but they didn’t have to, because shared knowledge of the pragmatics of the
situation gives “and” the right force anyway.

Domain knowledge says that Add.1 is probably an addendum
or addition. But Add.1 could easily be something that is totally dependent on
the content of A/56/L.28, like:

“remove Wensleydale from the list of cheeses in paragraph 4″

In that case, the effective draft resolution is indeed A/56/L.28 and Add.1, and
it would be coherent for people to say stuff like “I would be in favour of A/56/L.28, which isn’t on the table, but must vote against A/56/L.28 and Add.1, which is “. As usual, the deal with named entities is that people are using language to talk about the things they care about, and there is some slop.

The question of how a bureaucracy formulates and uses the rules for building resolutions and voting on them is probably a research topic for someone, but might be a tough sell as a contribution to linguistic semantics. My sense is that there are two semi-formal systems operating here: the linguistic system that defines how the mapping to meaning usually works and the legal framework that explains how you resolve disputes what was voted on. There are no strong reasons to expect that these will align exactly. To understand a usage, however, you will sometimes have to consider both.

OK. Now I’ll read what Alexandre says

I would love to see your take on this expression. I wrote down my understanding of it on my own blog: http://blog.outerthoughts.com/2009/03/conjunctions-in-named-entities/