ACE Parse/Generation round-tripping "unknown" and synonyms

I am experimenting with a templating system that uses the ERG to generate phrases with the noun phrase replaced. Step 1 is to use ACE to parse a phrase and then see if ACE will regenerate the same phrase back.

I am testing using the “fict” (fiction) COCA corpus sample from here. My results so far with two questions below:

These all round-tripped properly:

The vampires didn’t get him.
FANS GATHERED TODAY AT the Los Angeles grave of dead rocker Tom Paine on the twenty-fourth anniversary of his death.
Paine, lead singer of the Seventies group Commonsense, has been called The First Goth.

All of those phrases generated at least one MRS that, when fed through ACE to generate, generated the original phrase (when capitalization was ignored).

Won’t Generate due to nns_u_unknown: The fans, who left black roses and paperback copies of Dracula, were a mix of aging baby boomers and newer acolytes, like this young gothette

_acolytes/nns_u_unknown is part of all the MRS alternatives so generation won’t work as discussed here in another post. The fix described there should address this.

None of the MRSs from these will generate:

Blond girl, black garb and eyeliner, long blond hair.
His death.
Pathetic and broke?
Of a heroin overdose in a cheap motel?
Johnny Goldberg ( former rhythm guitarist, Commonsense ) : We were so sincere then.

ACE won’t generate any of the MRS documents any of those phrases output. I believe that the unknown predication is what is preventing generation since I haven’t seen it work on anything.

Is this the same problem that is discussed in this post and I just need to use a different root for generating fragments?

Won’t Round-Trip: What is it about Tom Paine that attracts your generation?

ACE Gets close to generating the original text but never generates “that”, instead uses “which does”, so the best I get is:

MRS 1, GEN 1: What is it about Tom Paine which does attract your generation?

Is there a fix for getting the generator to generate the synonyms for things like “that” <–> “which does”?

I answered one of my questions: Yes, passing “-r root_frag” to ACE does fix the fragment generation as described in that post.

Still would love to know if there is a way to generate alternatives like “that” <–> “which does”, though

1 Like

This seems to be a deficiency in the ERG trigger rules – a question I hope @dan will pick up? I poked around a little bit and found that the semantically vacuous word that’s required – that_r – is indeed triggered, but with the wrong ARG0, causing it to be incompatible with the place it needs to fill in the sentence.

that_r is being triggered with that_rel_rule1, which is a spurious match: it thinks be_v_id is a relative clause modifying a noun whose primary EP is relative_mod, and as such it grabs the ARG0 of relative_mod as the ARG0 for that_r. Given the somewhat unusual scoping structure that was chosen for this analysis I won’t try to speculate on how to write an improved trigger rule.

1 Like