maxent libraries?

[Daniël de Kok]

On May 23, 2011, at 3:01 AM, Sasha Rush wrote:
> liblbfgs is in C and has an mit license. A bit bare bones compared to something like mallet, but you can use that to train a maxent model.

Indeed, once you have an implementation of an optimization method, making a maxent parameter estimator is fairly easy. I wrote a maxent estimator for ranking tasks (parse disambiguation, fluency ranking) based on liblbfgs:

https://github.com/danieldk/tinyest

It can also be used for building classifiers, but the data representation is not optimal (you'll end up enumerating the same set of features for every class in a given context).

I started with an Haskell implementation of the very same program using the lbfgs module, but for large data sets it's not very memory-efficient if you want Haskell'ish data structures. It's far nicer to use FFI, and write an enumerator to construct data in C from a nice Haskell representation.

-- Daniël
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://projects.haskell.org/pipermail/nlp/attachments/20110523/f030ce25/attachment.htm>