[UIMA 3.0] Holding to CASes

[Pablo Duboue <pa...@gmail.com>]

Certain use cases of NLP work (multidocument summarization comes to
mind) require extensive work out-of-framework to render in UIMA:
selected sentences and segments had to be moved out of the CAS and
kept in non-UIMA data structures until all the documents of interest
(e.g., the document cluster) has been processed. This was a big
deterrent for using UIMA in such cases.

Does UIMA 3.0 now supports having 1000 CASes floating around for a few
annotations? Or, if users were to do that, would they blow up their
RAM usage (and blame the framework for their own, unsupported case).

In the past, holding to CASes was a no-no. I want to know whether UIMA
3.0 changes that.

Re: [UIMA 3.0] Holding to CASes

[Marshall Schor <ms...@schor.com>]

On 11/19/2016 1:09 PM, Pablo Duboue wrote:
> Certain use cases of NLP work (multidocument summarization comes to
> mind) require extensive work out-of-framework to render in UIMA:
> selected sentences and segments had to be moved out of the CAS and
> kept in non-UIMA data structures until all the documents of interest
> (e.g., the document cluster) has been processed. This was a big
> deterrent for using UIMA in such cases.
>
> Does UIMA 3.0 now supports having 1000 CASes floating around for a few
> annotations? Or, if users were to do that, would they blow up their
> RAM usage (and blame the framework for their own, unsupported case).
In v2, the "Default" settings for a CAS were tuned for large-ish CASes.  But
these defaults can be overridden and very small CASes created.

In v3, the representation of Feature Structures directly as Java objects
eliminates one of the big consumers of space in the default setting.  So it's
more convenient.  But you can achieve the same result (small CASes)  in v2 by
overriding the defaults.

-Marshall
> In the past, holding to CASes was a no-no. I want to know whether UIMA
> 3.0 changes that.
>