Here's how I think about UIMA types, versus rich data structures available in
modern programming languages.
UIMA was designed to facilitate combining more-or-less independently developed
How each primitive annotator is programmed, internally, is invisible to UIMA -
the annotator can make arbitrary use of any of modern programming language's
rich data structures. The UIMA types come into play at the annotator boundary -
they are there to specify how annotators send data from one annotator in the
pipeline to the next.
When writing a primitive annotator, developers should keep this distinction in
mind, and use appropriate internal data structures (in whatever language they're
writing in) for computation, and fetch and put data into CAS data structures
that are meant to be shared with other Annotators the CAS is sent to.
Data in the CAS can be accessed using JCasGen'ed classes. This is a bridge to
make accessing CAS data more Java -like and convenient. JCas is one of 2 main
"APIs" for a CAS. The other one is the non-JCas style, where you make use of
the CAS APIs directly, using UIMA Types and Features to create types and set and
get feature values. This style is available in the UIMA CPP implementation.
That implementation has no equivalent to the JCas style, and no equivalent to
The same principles apply, in that when writing a C/C++ annotator, you may use
all the modern data structures, etc. internally within the annotator, and at the
boundaries, where you want to share data with other annotators, you get or put
data into the CAS types and features.
Data in the CAS has to be serializable in standard ways. UIMA uses this
property to allow "remoting" annotators and connecting to them using a service
interface capability. You can see this in the UIMA-AS extension of UIMA, which
has deployment descriptors specifying how you want to deploy various annotators
that make up a pipeline.
There is one more capability UIMA provides for co-located Annotators in a
pipeline; these can be set up to share a common external resource. An external
resource can make use of any modern programming language constructs (e.g., a
HashMap). But it is not "serialized". It doesn't participate in the basic
architecture of sending a CAS through a pipeline of (independently developed)
annotators, and supporting the "remoting" of those annotators in a scaled-out
User augmentation of JCas classes is possible. But only the data that is in CAS
Types and features is transported when the CAS is sent to a remote annotator (or
from a Java Annotator to a C/C++ Annotator). As a limited side effect, when a
CAS is sent from a Java annotator to another Java annotator in a pipeline,
running on a single machine in a single JVM, the UIMA framework doesn't
serialize the data - it just passes things internally in memory. In this
instance, auxiliary data kept in a JCas class is maintained and can be used in
subsequent annotators. As you can see, extra data in the JCas, because it isn't
reliably available to other annotators except under some controlled situations,
should not generally be used for data to be shared among annotators; that's what
the CAS was designed to support. But it's fine to use this for data local to
one annotator. For that usage, though, it would be better to keep this data as
local data associated with the Annotator, rather than with some particular JCas
One other analogy for these modern times, might be useful. Think of the trend
today in microservices. The microservice, itself, can make use of arbitrarily
complex internal data structures. But its external interface is some fairly
simple set of APIs with fairly simple data.
The CAS is like those APIs, in its purpose - to provide a way for multiple
components (annotators) which might have arbitrarily complex internal data
structures, to communicate with independently developed other components
I hope this answers some of your questions.
On 11/30/2016 3:25 PM, David Fox wrote:
> Does the UIMA Java framework support modifying or extend the java class
> generated by JCasGen corresponding to a custom Type? If so, are there any
> common circumstances where this is necessary?
> I didn’t see anything in the examples or documentation about modifying the
> generated classes, but I also didn’t see anything saying you couldn’t. I
> suspect that this is not supported (and that otherwise you wouldn’t be able
> to pass a CAS between distributed UIMA AS components, or between a Java
> annotator and a C++ one). But it would be nice to know for certain.
> The reason I ask is that the set of data structures supported by UIMA types
> (individual FS references, FSList linked lists, and FSArray arrays) is
> fairly limited compared to modern programming languages, which often directly
> support associative arrays, trees, and graphs. I’m trying to understand
> whether this is a restriction on the implementation of custom types (which it
> would be if modifying/extending the generated class was not supported), or
> just on the public interface accessible via the UIMA API.