candidate_verbs.txt

Restricted:

very negative:
(happens when verb is common but particle rarely happens with it)
pour up
drop up
sign down
line down
tuck out
show out
jot up
stand back
check off
spit down
drum out
brush out
wipe up
follow down
wave up

very positive:
(happens when verb is rare so the particle is informative)
bandy about
piece together
fiddle around
boss around
sing along
leave behind
tear apart



Highly restricted:
very positive:
scrape together
piece together
knit together
bandy about
boss around
leave behind
tell apart
win over
mull over
gloss over


we have VP, V, and P
highest pmi when 
VP == V (the verb always appears with its particle, e.g. "bandy about"),
or VP == P (the particle only ever appears with this verb, e.g. "leave behind")

The real question is, given the verb, how much are you expecting the particle?
This should incorporate "null particles", so the VP==P high-pmi cases are dubious.
The VP==V ones are also scary,
because they just mean that the verb never appears with any other particle.

Very positive:
piece together
bandy about
boss around
leave behind	(VP=P)
gloss over   (gloss over the details, *gloss the details over, ?gloss it over)
fritter away
bowl over
goad on
hem in
rein in
cash in	(+on, cash checks)
tamp down     (+on)
wolf down
jot down
slow down	(+car)
fob off
wring out

Very negative:
show out
sign out
follow out
pick off
nail up
kick up


neutral:
get out
throw out
play down



What really matters isn't straight up pmi.
It's p(particle|true context) vs. p(particle|noisy channel context at distance d).
From that perspective, "out" in "show out" doesn't get monotonically better with distance.
You still need some licensor of "out".
We have p(out|show+particle) < p(out|V+particle).
We really want p(out|show+particle) vs. p(particle+out).
p(particle+out|V) will be extremely low. Many verbs don't ever predict a particle!
So this pmi will just about never be negative.








CONTROL FOR:
Animacy
Compositionality of particle and verb according to Lohse criteria
Frequency?
Which particle



show X out < show out X
bandy about Y < bandy Y about
wolf down Y < wolf Y down

boss X around
boss around X

animate object, high pmi
boss around
goad on
?slow down
egg on?
invite over
invite along

inanimate object, high pmi
bandy about
fob off
rein in
wolf down
slow down
wring out
tamp down
eke out?
mull over
gloss over
squirrel away

animate object, low pmi
sign out
show out
send up
dress down

inanimate object, low pmi
?pick off
nail up
kick up
set down
work up
sign over
make out?
smooth away
knock back



Look in Google Books too.

Lohse's compositionality criteria:
(10) Verb entailment test If[XVNPPt]entails[XVNP],thenassignVi.Ifnot,assignVd.
(11) Particle entailment test If[XVNPPt]entails[NPPredVPt],thenassignPti.Ifnot,assignPtd. PredV predicationverb(BE, BECOME, COME, GO, STAY)15


sparse distro over symbol sequences of length n
you can choose ordering
you can only keep b symbols in the context

for every symbol there are only b elements that influence its probability -- markov blanket

take lang opt view?
some orders have a simple policy