Some thoughts on mapping MRSs to Wordnet sense (mainly illustrated with the ERG and PWN). Examples use a variant of the indexed MRS to save space (part of the work on Lexical Semantics).

Contents

1. Single Words
   1. Proper Nouns
   2. Decomposed nouns
   3. Superlatives
   4. Different POS
2. Multiple Words
   1. Compositional Compound Nouns
   2. Non Compositional Compound Nouns
   3. Idioms
3. Phrases
   1. Coordination
   2. Light Verbs

In straight forward cases like this, each open class predicate maps to a wordnet sense.

The cat_{cat_n1} ate_{eat_v1} a dog_{dog_n1}.
e3:
 _1:_the_q⟨0:3⟩[BV x6]
 x6:_cat_n_1⟨4:7⟩[]
 e3:_eat_v_1⟨8:11⟩[ARG1 x6, ARG2 x9]
 _2:_a_q⟨12:13⟩[BV x9]
 x9:_dog_n_1⟨14:18⟩[]

Mapping the predicates to lemmas gives us directly:

x6:_cat_n_1 = cat_n1
e3:_eat_v_1 = eat_v1
x9:_dog_n_1 = dog_n1

Note that quantifiers are sometimes in wordnet, sometimes not:

Each_{each_a1} cat_{cat_n1} ate_{eat_v1} a dog_{dog_n1}.
e3:
 _1:_each_q⟨0:4⟩[BV x5]
 x5:_cat_n_1⟨5:8⟩[]
 e3:_eat_v_1⟨9:12⟩[ARG1 x5, ARG2 x9]
 _2:_a_q⟨13:14⟩[BV x9]
 x9:_dog_n_1⟨15:19⟩[]

Here we also have:

_1:_each_q = each_a1

(with a non matching pos a <> q)

For proper nouns (and numbers and a few others), the predicate is an abstraction like named_rel, and the value is in the CARG:

Bast_{Bast_n1} ate_{eat_v1} a dog_{dog_n1}.
e3:
 _1:proper_q⟨0:4⟩[BV x6]
 x6:named⟨0:4⟩("Bast")[]
 e3:_eat_v_1⟨5:8⟩[ARG1 x6, ARG2 x9]
 _2:_a_q⟨9:10⟩[BV x9]
 x9:_dog_n_1⟨11:15⟩[]

We want:

x6:named = Bast_1

Some words are given complex semantics:

A cat_{cat_n1} ate_{eat_v1} here_{here_a1}
e3:
 _1:_a_q⟨0:1⟩[BV x6]
 x6:_cat_n_1⟨2:5⟩[]
 e3:_eat_v_1⟨6:9⟩[ARG1 x6]
 e10:loc_nonsp⟨10:15⟩[ARG1 e3, ARG2 x11]
 x11:place_n⟨10:15⟩[]
 _2:def_implicit_q⟨10:15⟩[BV x11]
 e16:_here_a_1⟨10:15⟩[ARG1 x11]

here is given semantics equivalent to "in this place". Ideally, we would like a mapping such as:

e10:loc_nonsp = here_a1
x11:place_n = here_n1

with "e16:_here_a_1" unmapped. "in this place" = here_a1 and "this place" = hear_n1

There are not so many of these, it should be possible to do them with exception handling

Wordnet has some superlatives (linked through domain usage to superlative_n_1): best, worst, least, ...

As far as I can tell, they are not actually linked to the relevant adjectives!

I think we should tag these with the relvant adjective (good, bad, less; , ...) and distinguish if need be by the presence of the superlative predicate.

ERG collapses many adjective/adverb distinctions: they are all 'a'. Wordnet often has them as different entries, linked with derivation links. I lean towards collapsing them :-).

Sometimes both the ERG and PWN treat a MWE as a single concept, and then it is easy.

The cat_{cat_n1} gobbled_{gobble up_v1} a dog_{dog_n1} up_{gobble up_v1}.
e3:
 _1:_the_q⟨0:3⟩[BV x6]
 x6:_cat_n_1⟨4:7⟩[]
 e3:_gobble_v_up⟨8:15⟩[ARG1 x6, ARG2 x9]
 _2:_a_q⟨16:17⟩[BV x9]
 x9:_dog_n_1⟨18:21⟩[]

The character mapping is a bit less direct, but the final mapping should be just:

x6:_cat_n_1 = cat_n1
e3:_gobble_v_up = gobble_up_v1
x9:_dog_n_1 = dog_n1

PROBLEM sometimes they will disagree. Postpone mapping for now.

The cat_{cat_n1} ate_{eat_v1} a guard_{guard_dog_n1} dog_{guard_dog_n1}.
e3:
 _1:_the_q⟨0:3⟩[BV x6]
 x6:_cat_n_1⟨4:7⟩[]
 e3:_eat_v_1⟨8:11⟩[ARG1 x6, ARG2 x9]
 _2:_a_q⟨12:13⟩[BV x9]
 e15:compound⟨14:23⟩[ARG1 x9, ARG2 x14]
 _3:udef_q⟨14:19⟩[BV x14]
 x14:_guard_n_1⟨14:19⟩[]
 x9:_dog_n_1⟨20:23⟩[]

NTU WN tags just MWE in this case, SemCor maps only the MWE I think we want:

x6:_cat_n_1 = cat_n1
e3:_eat_v_1 = eat_v1
x9:_dog_n_1 = guard_dog_n1
x14:_guard_n_1 = x

We get this in two steps. First we get the compositional reading from the ERG's treebanking (really we get guard_n_per, which also includes basketball guard, ...):

x6:_cat_n_1 = cat_n1
e3:_eat_v_1 = eat_v1
x9:_dog_n_1 = dog_n1
x14:_guard_n_1 = guard_n2 'a person who keeps watch over something or someone' 

Then we write an (optional) mtr, that rewrites the compound to a single noun (hopefully dealing with modifiers correctly).

<guard_n2, compound, dog_n1> => <guard_dog_n1>

It would be good to also link these in wordnet: we should have guard_dog_n1 is_a dog_n1, we also want 'guard_n2' internally modifies 'dog_n1' in 'guard_dog_n1'.

One could instead think of something like this:

x6:_cat_n_1 = cat_n1
e3:_eat_v_1 = eat_v1
x9:_dog_n_1 = dog_n1
x14:_guard_n_1 = guard_n1
e15:compound = guard_dog_n1

But this isn't quite right: too many predicates, compound is not a noun, ...

The cat_{cat_n1} ate_{eat_v1} a guard_{hot_dog_n1} dog_{hot_dog_n1}.
e3:
 _1:_the_q⟨0:3⟩[BV x6]
 x6:_cat_n_1⟨4:7⟩[]
 e3:_eat_v_1⟨8:11⟩[ARG1 x6, ARG2 x9]
 _2:_a_q⟨12:13⟩[BV x9]
 e14:_hot_a_1⟨14:17⟩[ARG1 x9]
 x9:_dog_n_1⟨18:21⟩[]

Here, we don't want the semantics "a dog that is hot", so:

x6:_cat_n_1 = cat_n1
e3:_eat_v_1 = eat_v1
x9:_dog_n_1 = hot+dog_n1

Ideally, the ERG should contain "hot dog" as a single entry, so that things map even better.

X keeps tabs on Y

We don't really know how to mark the whole idiom (although the ERG recognizes it)

But we can write a machine translation rule to rewrite it.

X doesn't know X's arse from X's elbow "X is an idiot." ?Postprocess

How should we show this?

grass and brown snakes grass_snakes and brown_snakes ?Postprocess

give a start give_start? Do we just make entries for all light verbs?

Many more corner cases to come :-): "Sleeping Beauty: is sleep v or n", more complex MWEs, ... .