[developers] Docu of the MaxEnt .mem file structure?

[Woodley Packard]

Hi Alex,

There are some notes on training MaxEnt models using my alternate software from the summit, recorded here towards the bottom of the page:

http://moin.delph-in.net/IndraTreebanking

The link given there is for a Linux executable.  The software is available as source code too (svn checkout http://sweaglesw.org/svn/feature-forest/) but there is no documentation about how to navigate the dependencies.

Regards,
Woodley


> On Sep 22, 2017, at 7:40 AM, Stephan Oepen <oe at ifi.uio.no> wrote:
> 
> hi alex,
> 
> > (0) [1 (2) ^ aj-hd_scp_c av_-_s-cp-mc-pr_le "well"] 0.522613 {36392 447 36392 445} [0 1]
> 
> in a nutshell, the format is
> 
> (id) [tid (parameter*) symbol] weight {counts} [range]
> 
> the feature identiers are unique consecutive integers.
> 
> template identifiers are integers, specified through comments in the top of ’features.lisp‘ in [incr tsdb()].
> 
> each template can take zero to two parameters, e.g. the grandparenting level in type-1 features; see the above file and zhang, oepen, & carroll (2007; IWPT).
> 
> symbol is the representation of the feature proper, again interpreted specific to the template type, e.g. two grandparents in the above (of which one is the tree ’top sentinal‘), then the mother, then the sequence of daughters (singleton, in this case).
> 
> the counts and (minimum and maximum) value range information is there for bookkeeping purposes; MEM scoring does not depend on this.
> 
> MEM training is supported in [incr tsdb()] in the LOGON tree; documented to some degree on the wiki, i believe.  as of this summer, i think woodley offers an alternative full-forest trainer.  i recommend you look into that!
> 
> best wishes, oe
>