On 7/11/11 6:03 PM, Jörn Kottmann wrote:
The configuration parameter part can done better if the code
is part of the core framework. So working together here, would
be a win for UIMA and uimaFIT.
Over at OpenNLP we have a lot of code which just deals with that,
I believe such a configuration framework needs to solve three issues:
- Defining the mandatory parameters, which are currently specified
in the descriptor.
This will greatly reduce the risk of running an AE with a non-compatible
descriptor. The mandatory parameter checks can do their work more
reliable.
In the current implementation an AE might need to declare a new
mandatory parameter,
but a user runs it with an old descriptor (where it is not declared),
in this case
the built-in checks cannot detect that a parameter is missing.
But when the class itself defines the mandatory parameters it will.
- It should be able to inject parameter values into instance variables.
- Type system mapping should be supported by the framework itself,
e.g. declaring new value types, which is a type system Type and Feature.
OpenNLP also needs to load resources, but this cannot be improved by
annotations,
because the initialize method needs to create a "wrapper" object per
loaded resource,
the loaded resource is hidden behind the wrapper.
For example, it declares a NameFinderModelResource, which has a method
getModel(),
this model is then used to instantiated a NameFinderME object, which is
written into
an instance field and then called from the process method.
Jörn
