"Lee is a beautiful singer" interpretations

I’m trying to get my head around how to map the two interpretations of
“Lee is a beautiful singer”

  1. Lee sings beautifully
  2. Lee is beautiful and sings

onto the two well-formed alternatives I generate from the MRS:

_a_q(x9, 
    and(_beautiful_a_1(e14, x9), _singer_n_1(x9)), 
    proper_q(x3, 
         named(Lee, x3), 
         _be_v_id(e2, x3, x9)))

proper_q(x3, 
    named(Lee, x3), 
    _a_q(x9, 
        and(_beautiful_a_1(e14, x9), _singer_n_1(x9)), 
        _be_v_id(e2, x3, x9)))

Which of these is applying beautiful to just singer and not lee? (and why?)

Maybe I still haven’t gotten the well-formed MRS structures right and there are more alternatives I’m not seeing?

MRS treats both of those as the same. If you want more, you need internal structure for singer (something like: singer(x) = sing(e,x) & person(x), in the style of the Generative Lexicon). This would be interesting to do, and perhaps could be done semi automatically with wordnet, but is a non-trivial extension.

2 Likes

Thanks @bond. I guess the good news is I wasn’t missing some subtle logic interpretation.

To give a sense of why this would be non-trivial (as @bond has said), consider “beautiful musician”. There is no verb “music” (at least, in standard English), so it’s not obvious how to decompose “musician” – we could invent an abstract music_v predicate which doesn’t correspond to a real verb, or we could use a real verbal predicate even though the verb doesn’t appear here (like _play_v_1).

The more general point is that the ERG only gives you semantic distinctions which can be seen in the syntax. While it is tempting to give a more fine-grained semantics, it is difficult to do this consistently at the scale of a whole language. Copestake (2009, section 3) discusses this general point, but using a different example: labelling the arguments of predicates.

2 Likes

Thanks for the detail on this @guyemerson.

I still haven’t gotten enough background swapped in to understand why introducing an abstract music_v predicate is too far from the syntax to introduce but something like loc_nonsp is ok to introduce. At my current level of understanding, these both look like “inducing some abstract thing that wasn’t in the syntax”, but I know my understanding is still really shallow and probably wrong.

Currently working through Representation and Inference for Natural Language, and hoping I’ll get a bit more insight…

I didn’t understand the point from @bond

MRS treats both of those as the same

The online demo of ERG gives me two very similar MRS for “Lee is a beautiful singer” (that I interpreted as “Lee is beautiful and Lee is a singer”) and two very distinct MRS for examples 1 and 2.

@guyemerson thank you for the example. It is somehow related to the example from https://www.goodreads.com/en/book/show/1041621.Natural_Language_Understanding. Terms like father or singer have a more clear semantics but terms like dog or musician may have always an incomplete semantics.

The ERG has a relatively small number of abstract predicates like loc_nonsp. If we allow things like music_v, we are on a slippery slope and will probably end up with more abstract predicates than real predicates. Other cases are more difficult still. What about “beautiful poet”? Is “poem” or “poet” a more basic word, when neither is obviously derived from the other, and “poe” is not a word? We can of course come up with a solution for any particular word, but it is a challenge to come up with a solution that can be consistently applied at the scale of a whole language.

loc_nonsp is also anchored in real predicates – it corresponds to prepositions, but is more general (e.g. it could mean “in” or “at” or “on”). Where it’s used in the grammar, there will be a corresponding construction using a real predicate.

@bond meant that the two suggested readings of “Lee is a beautiful singer” (which @EricZinda has glossed with 1 and 2) are not distinguished in MRS. 1 and 2 were not meant to be taken as example sentences.

1 Like

G’day,


arademaker

    February 17

I didn’t understand the point from @bond

MRS treats both of those as the same

The online demo of ERG gives me two very similar MRS for “Lee is a beautiful singer” (that I interpreted as “Lee is beautiful and Lee is a singer”) and two very distinct MRS for examples 1 and 2.

I think the two MRSs for “Lee is a beautiful singer” are those where singer is a proper noun (like the sewing machine) and when it is a common noun. They do not correspond to the two interpretations “Lee is a person who sings beautifully” and “Lee is a beautiful person who sings”.

Ahh, one can not model in MRS the two different readings of the sentence suggested in [1,2] … MRS doesn’t have expressivity to distinguish the two readings? Or MRS as representation language does have the expressivity but the process of generate the MRS is the limitation?

Thank you @bond for making it clear. Thank you @guyemerson for your attention.

The limitation is writing a grammar to generate the MRS, if we want to deal with all related examples as well.

There’s a general preference for underspecifying MRS, unless there’s a clear distinction in the syntax – for example, if readings 1 and 2 required different morphology on the adjective, we would want to distinguish them.

The design principles articulated here might be helpful in this context:
http://moin.delph-in.net/WeSearch/DesignPrinciples

2 Likes