kawu / crf-chain1-constrained Goto Github PK

The R0 construction method supplied to the training function can be of the type

SentL a b -> Set b

instead of:

[SentL a b] -> Set b.

Then we can union results gathered for individual sentences. The same applies to the feature selection.

It has a dummy mkWordL = id implementation right now.

It appears an update of one of the dependencies broke the build. See http://hydra.cryp.to/build/574078/log/raw for a complete log that shows the error.

The (Feature -> Ix) map is used during the training process, but it is not needed for label tagging. It would be great if we could (optionally) get rid of the map, because it probably takes a lot of memory. To consider:

• Is it possible to represent the use of the map on the type level?
• Is it possible to restore the map on the basis of the rest of the model representation?

See theFIXME label in the source code.

Constraints can be shared between different words during the training process. Pros:

• Easier interface, since there would be no reason to treat unknown words in any special way.
• Sharing would give higher memory usage reduction -- now only constraints imposed over unknown words are shared.