Hi,

Here are a couple of thoughts.

1) you can write a custom flow controller.  See
http://uima.apache.org/d/uimaj-2.4.0/tutorials_and_users_guides.html#ugr.tug.fc

This would allow you to redirect CASes back from the end of the pipeline into
the front (or anywhere else you wanted), any number of times.

There are example flow controller implementations in the examples directory of
the distribution ( in src/main/java/org/apache/uima/examples/flow )

2) you could split the application as follows:

  a) a plain Java app which sets up the UIMA pipeline, and then creates & sends
CASes through the pipe line, as many times as you want.  Being that this is
plain old Java code, you can use any/ all of Java's capabilities to control 
this.

HTH -Marshall

On 6/17/2013 9:12 AM, Susanne Neumann wrote:
> Hi,
>
> I want to implement a bootstrapping algorithm using UIMA, which requires 
> processing a whole text collection several times. With each iteration, new 
> evidence based on the results of the previous runs on all the documents is 
> collected and applied. The number of iterations is determined at runtime.
>
> I planned to write a bootstrapping AE, but I can't figure out how to 
> iteratively process the collection with UIMA, because the process method 
> processes the text collection only once.
>
> As a workaround, I am considering to add the annotator several times to the 
> pipeline. However, it depends on the order in which the documents are 
> processed withing the pipeline to work as desired. In which order are 
> documents being processed in a pipeline? Does each component process the 
> whole text collection first before going to the next component, or is every 
> document of the collection being processed by each component first? In the 
> latter case, the workaround would not work. Another solution is, to run a 
> whole Pipeline containing the annotator several times.
>
> Is there any better way to iteratively process a text collection than those 
> workarounds? Which of them would work? Any hints on this are welcome.
>
> Thanks,
> Susanne
>

Reply via email to