Original issue 45 created by ClearTK on 2009-02-06T16:23:21.000Z:

I want a factory method that takes a Feature and creates a new feature with
an updated name. I am going to add a method with the following feature:

public static Feature createFeature(String namePrefix, Feature feature)

which returns:

return new Feature(createName(namePrefix, feature.name), feature.value);

Original issue 31 created by ClearTK on 2009-01-19T18:18:07.000Z:

we need to solidify the part-of-speech tagging functionality so that you
don't have to resort to using the example code or having to do it all
yourself if you want a specialized part-of-speech tagger. I have a bunch
of code that I wrote for an old version of ClearTK that I had in a separate
project that supported the results in our LREC workshop paper. I am in the
process of refactoring this code and adding it to the ClearTK code base.

Original issue 12 created by ClearTK on 2008-12-05T19:28:49.000Z:

Please use labels and text to provide additional information.

There are a number of use cases for creating analysis engines, collection
readers, etc. outside of unit testing. In fact, any time you are going to
create and run components without using descriptor files - these methods
will come in handy.

We should put the following methods into UIMAUtil:

• getAnalysisEngine
• getCollectionReader
• getTypeSystem
• setConfigurationParameters

Original issue 7 created by ClearTK on 2008-12-05T18:21:14.000Z:

The following unit tests have data in the text of the .java file that
should not be redistributed (e.g. song lyrics, treebank data, etc.) These
unit tests are only available on the old repository and either need to be
refactored or replaced.

Parent directory
/ClearTK/test/src/

Affected Files
org/cleartk/classifier/encoder/features/string/StringFeatureEncoderTests.java

Original issue 11 created by ClearTK on 2008-12-05T19:20:24.000Z:

Please use labels and text to provide additional information.

We need to document our expectations for what goes into _InitialView:
basically plain text with no markup, no escaping, etc. It might also be
useful to include an overview of what other views we include, e.g.
TreebankView, PropbankView, etc.

Original issue 50 created by ClearTK on 2009-02-09T23:10:58.000Z:

The method TreebankFormatParser.parseDocument calls splitSentences which
attempts to break down the document into single sentence parses. This
method assumes that sentences are delimited by "( (S" as they are in the
PTB. I have some data that has one parsed sentence per line where each
sentence start with "(S1 (S". The method splitSentences is not splitting
these files up into sentence-level parse strings. Instead it passes in the
whole file to parse(String, String, int) which returns a single
TopTreebankNode with child nodes corresponding to the parse of the last
sentence of the file.

I think this is a fairly common format that we should handle.

Original issue 30 created by ClearTK on 2009-01-15T17:32:46.000Z:

When you are going to use a non-sequential tagger for a sequential tagging
task (e.g. BIO chunking) then you are going to want to use previous labels
as features to pass into the classifier. This precludes using
consumeAll(List<Instance>) because you need to classify each instance one
at a time so that feature extraction has access to labels already assigned.
This is in general not necessary for data writers and more generally it
should not be part of the instance consumer interface. I would, however,
like to be able to cast my instance consumer and get isSequential.

Original issue 36 created by ClearTK on 2009-01-23T19:03:55.000Z:

Please use labels and text to provide additional information.

We should have a component that can calculate inverse document frequencies
(IDFs) over a corpus.

Original issue 47 created by ClearTK on 2009-02-06T21:48:42.000Z:

As of r224 the following descriptors are not being tested anywhere:

"org.cleartk.corpus.ace2005.Ace2005GoldAnnotator"
"org.cleartk.corpus.ace2005.Ace2005GoldReader"
"org.cleartk.corpus.ace2005.Ace2005Writer"
"org.cleartk.corpus.conll2003.Conll2003GoldReader"
"org.cleartk.corpus.penntreebank.PennTreebankReader"
"org.cleartk.srl.conll2005.Conll2005GoldAnnotator"
"org.cleartk.srl.conll2005.Conll2005GoldReader"
"org.cleartk.srl.propbank.PropbankGoldAnnotator"
"org.cleartk.srl.propbank.PropbankGoldReader"
"org.cleartk.srl.propbank.TreebankPropbankGoldAnnotator"

See the output of DescriptorCoverageTests.

Original issue 16 created by ClearTK on 2008-12-05T20:01:32.000Z:

Please use labels and text to provide additional information.

We currently test the linear and RBF kernels for SVMlight. Since we're
using Philipp's implementations, we should also test the polynomial kernel
and the sigmoid kernel.

Original issue 1 created by ClearTK on 2008-12-05T17:15:36.000Z:

What steps will reproduce the problem?

1. ReflectionUtil.getTypeParameterClass(Class) is used.
2. It only works for classes of the form class X extends Y<Z>

What is the expected output? What do you see instead?

ReflectionUtil.getTypeParameterClass(Class, Class) should be used everywhere.

Please use labels and text to provide additional information.

We need to write some tests where we mix a Classifier and a
ClassifierAnnotator with different parameter types, and make sure we're
throwing exceptions correctly.

Original issue 29 created by ClearTK on 2009-01-09T20:29:59.000Z:

I just noticed a descriptor file (SentencesAndTokens.xml) that has an
out-of-date license statement that points to the "for research only"
license. This needs to be replaced.

Original issue 14 created by ClearTK on 2008-12-05T19:52:00.000Z:

Please use labels and text to provide additional information.

We should remove our dependency on the Julie type system - we should be
able to exercise TypePathExtractor (the only test that uses the Julie type
system) well enough using just the ClearTK types.

Original issue 32 created by ClearTK on 2009-01-20T18:38:47.000Z:

I have a use case where I want something other than blank lines betwixt my
blocks when using LineWriter. In my case I want to print out the document
id at the beginning. I propose to add a BlockWriter interface that allows
one to write out whatever one wants at the top of each block. The default
behavior will still be to print out a newline at each block.

Original issue 28 created by ClearTK on 2008-12-22T18:07:09.000Z:

we have the model checked in and it works but there are not instructions on
how to recreate the model if we ever needed to. Basically, the parsed
sentence is provided in the unit tests that make use of the model. This
needs to be put in a separate data file repeated N times and then passed
into the appropriate model builder - probably a direct call to a main
method of an OpenNLP class.

Original issue 38 created by ClearTK on 2009-01-30T22:06:22.000Z:

I received the following feedback from a developer new to UIMA, NLP, and
machine learning - though otherwise very sharp. I think it would be
worthwhile to address all of the points that he makes.

<begin-message>
I looked at the wiki earlier and was a bit overwhelmed
and abandoned it as my first source of information.
After a short conversation with Kevin, it makes a bit
more sense, but I think it would benefit from a paragraph
or two of introduction. I may be on the edge of the expected
audience, but following are some questions I have after looking
at the main page and the main wiki page. A better introduction
on googleCode may not answer them all.

BTW, Do we have a book in the lab library that introduces machine
learning in the context of NLP? I've read Jackson and Moulinier.

http://code.google.com/p/cleartk/

"...feature extraction library" ...like what? POS, named entity,
misc relationships?

" ...wrappers..." UIMA wrappers?
I'd like to learn more about maximum entropy, support vector machines
and conditional random fields, but wouldn't expect that from a ClearTK
intro.

...also sequential taggers, chunkers, role labelling and temporal
resolution.

• Where does the name come from? (certainly not Tcl/Tk)

 http://code.google.com/p/cleartk/w/list

• What's a classifier? ...I'm guessing you could use one
  to do tagging in UIMA.

-What's the Maxent classifier and how is it different
than the POS tagger?

-What's a chunk tokenizer and how is that different from
other kinds of tokenizers.

• How is ClearTK both a pos tagger and these other things?

Original issue 26 created by ClearTK on 2008-12-12T18:42:29.000Z:

creating a JCas is fairly expensive (my tests indicate about .1 seconds).
So, each unit test should create at most one jcas and then call reset
instead of calling newJCas.

In fact, it may be possible to create a single JCas that gets kept across
all unit tests using a static ThreadLocal - one for each type system used
in the unit tests.

Original issue 19 created by ClearTK on 2008-12-05T21:35:12.000Z:

Here is a quick sketch of what we expect the window extractor to do:

Given a focus annotation, extract features on either side of or within the
focus annotation such that you can specify (optionally) a boundary past
which the extractor will not go. This will be referred to as boundary
conditions below. Additionally, you should be able to specify which
annotations to examine inside the boundary (all, 0-3, 2-5, etc.) This will
be referred to as start and end annotations below. The following outlines
the different possibilities:

Start annotation:

• not specified (default 0)
• specified (e.g. 1, 2, 3...)

End annotation:

• not specified (examine all elgible annotations past the start annotation)
• specified (e.g. 1, 2, 3, ...)

Boundary conditions:

• not specified (this condition will allow WindowExtractor to be a
  SimpleFeatureExtractor by default - default values will be 0 and
  document.size()
• character offsets
• a single annotation (e.g. a sentence, paragraph, etc.)
• multiple annotations (e.g. current sentence + preceding sentence)

The current implementation requires that you specify the start and end
annotation and use a single annotation as the boundary condition.

Similar issues almost certainly exist for WindowNGramExtractor

Original issue 2 created by ClearTK on 2008-12-05T17:34:20.000Z:

What steps will reproduce the problem?

1. Anything using the constituentParse attribute of the Sentence type

What is the expected output? What do you see instead?

The corresponding TopTreebankNode should be selected using
AnnotationRetrieval instead.

Please use labels and text to provide additional information.

We should keep the requirements of our Sentence type as minimal as possible
so that in the not so distant future, we can allow people to use their own
Sentence types.

Original issue 10 created by ClearTK on 2008-12-05T19:13:19.000Z:

Please use labels and text to provide additional information.

We need to document the names of all views that each AnalysisEngine reads
from or writes to. For example, TreebankGoldAnnotator reads from
"TreebankView" and writes to "_InitialView", but doesn't document this
anywhere. This is crucial for users that want to use our annotators with
other views, e.g. through view mapping:

http://incubator.apache.org/uima/downloads/releaseDocs/2.2.2-incubating/docs/html/tutorials_and_users_guides/tutorials_and_users_guides.html#ugr.tug.aae.logging

Original issue 9 created by ClearTK on 2008-12-05T18:52:31.000Z:

There is deprecated code in
/ClearTK/test/src/org/cleartk/classifier/feature/proliferate/FeatureProliferatorTests.java
that should be removed. The warnings are currently suppressed.

Original issue 42 created by ClearTK on 2009-02-06T15:51:34.000Z:

We have introduced a nasty and unnecessary dependency on our type system by
using the DocumentUtil class in all of our collection readers and various
places elsewhere. I noticed this most recently when I was trying to use
the XWriter and XReader in a different project and was not wanting anything
to do with the ClearTK type system. We should probably re-engineer this
such that one could implement their own createDocument and getDocument
methods that are specific to a different type system.

Original issue 22 created by ClearTK on 2008-12-05T22:08:46.000Z:

What is the expected output? What do you see instead?

The build.xml currently compiles into build/bin, while Eclipse compiles into bin. This makes the
interactions a bit more complicated when one uses both ant and Eclipse. Building in ant and
building in Eclipse should be equivalent.

Original issue 23 created by ClearTK on 2008-12-05T22:42:37.000Z:

Please use labels and text to provide additional information.

Descriptors should sit beside the corresponding java classes in the
hierarchy. They can then be imported directly from the .jar file.

Original issue 39 created by ClearTK on 2009-02-04T19:16:05.000Z:

ClearTK descriptors can be used from a .jar with XML like:

<import name="org.cleartk.util.PlainTextCollectionReader"/>

But this only works if we put PlainTextCollectionReader.xml in the
appropriate package folder. Philipp has moved some of the descriptor files
from "desc" to their appropriate packages, but we should do it for all of
them (e.g. PlainTextCollectionReader).

In the end, "desc" should be empty except for CPEs. All other descriptor
files should be in an appropriate package.

Original issue 3 created by ClearTK on 2008-12-05T17:48:34.000Z:

Please use labels and text to provide additional information.

There is currently no documentation for the different feature extractors.
We need to both complete the JavaDocs, and, more importantly, give a more
tutorial-like overview of what is available and how it can be used.

Original issue 4 created by ClearTK on 2008-12-05T17:55:10.000Z:

What steps will reproduce the problem?

1. Use a LineWriter to write out tokens that were created by chunk-based
   tokenizer

What is the expected output? What do you see instead?

There are blank lines in the output file where there shouldn't be.

Please use labels and text to provide additional information.

Running the same code using the default penn tokenizer, there are no blank
lines. Is the chunk-based tokenizer creating tokens that contain newlines?

Original issue 41 created by ClearTK on 2009-02-06T15:04:00.000Z:

I have started integrating the opennlp.uima wrapper around the "chunking"
constituent parser so that it works with our type system. The class
opennlp.uima.parser.chunking.Parser provides a method createAnnotation
which provides an OpenNLP Parse object which must be converted into
annotations which can be posted to the CAS. I have written a first pass at
a conversion between the opennlp.tools.parser.Parse object and our type
system in org.cleartk.syntax.treebank.util.TreebankNodeUtility (it probably
belongs elsewhere).

One problem with this approach is that opennlp.uima.parser.chunking.Parser
has been declared "final" which means we either need to modify the code for
our own purposes or we need to convince the authors that this class should
not be final (in progress).

http://opennlp.cvs.sourceforge.net/viewvc/opennlp/opennlp.uima/

Original issue 15 created by ClearTK on 2008-12-05T19:55:39.000Z:

Please use labels and text to provide additional information.

We need to explain how Train and BuildJar can be used at the command line
(and perhaps what that's actually doing in the code).

Original issue 48 created by ClearTK on 2009-02-08T22:24:17.000Z:

Right now, there's no way short of writing your own new EncoderFactory to
use a RowNormalizingFeaturesEncoder. SVMEncoderFactory should do what
ContextValueEncoderFactory and look at the UimaContext to determine whether
to use normalization or not.

Original issue 27 created by ClearTK on 2008-12-15T22:03:18.000Z:

I added a new method AnnotationRetrieval.getOverlappingAnnotations() which
seems to work for some code I am writing for my project - but it deserves
some javadoc comments and some unit tests.

Original issue 49 created by ClearTK on 2009-02-09T19:50:21.000Z:

I happened to notice two type system descriptor files that don't seem to
have any purpose in the org.cleartk.syntax.treebank.type package. My guess
is that they are there by mistake and can be safely removed.
TreebankCoreference isn't even in the TypeSystem.xml file

Original issue 5 created by ClearTK on 2008-12-05T18:14:52.000Z:

Please use labels and text to provide additional information.

To be more compatible with UIMA, we should use Loggers everywhere instead
of printing to System.out or System.err. Some information about how to do
this is here:

http://incubator.apache.org/uima/downloads/releaseDocs/2.2.2-incubating/docs/html/tutorials_and_users_guides/tutorials_and_users_guides.html#ugr.tug.aae.logging

Original issue 17 created by ClearTK on 2008-12-05T20:25:38.000Z:

Please use labels and text to provide additional information.

We should write a tutorial on how to write a new feature extractor class,
demonstrating when a FeatureProliferator might be useful, etc.

Original issue 51 created by ClearTK on 2009-02-10T00:26:44.000Z:

Take TreebankGoldAnnotator.xml as an example - it takes in one view
(TreebankView) that has treebank parses in it and puts the parsed treebank
node objects into another view (_InitialView). These need to be listed in
the capabilities section of the descriptor file. This makes them much more
flexible wrt to aggregate analysis engines where you are trying to map one
view to another.

Original issue 33 created by ClearTK on 2009-01-20T20:12:25.000Z:

When doing things like cross-validation, it is nice to be able to point the
PlainTextCollectionReader at a directory and provide a list of file names
that should be run (or several lists).

Original issue 43 created by ClearTK on 2009-02-06T16:07:59.000Z:

I have been using aggregate analysis engines a fair bit recently and I've
noticed that when you change the view of a single-view analysis engine to
something other than _InitialView - you end up posting Document annotations
to the other view when you probably still (or also) want it in the initial
view.

I haven't thought this through carefully - but I think the solution is to
make sure that whenever createDocument is called that the Document
annotation gets added to the _InitialView. I'm not sure whether it should
also get added to the mapped view or not - I think putting it in both might
be preferred and shouldn't cause any problems.

Original issue 24 created by ClearTK on 2008-12-06T00:42:43.000Z:

There are a number of tests that fail because we do not have (yet)
permission to redistribute the data that the tests rely on.

GeniaPOSParserTests
GeniaPosGoldReaderTests
TimeMLGoldAnnotatorTests
TreebankAligningAnnotatorTests

Original issue 20 created by ClearTK on 2008-12-05T21:49:49.000Z:

Please use labels and text to provide additional information.

Wrappers for OpenNLP are already provided with UIMA in the
examples/opennlp_wrappers directory. We should use these instead of
maintaining our own wrappers. We may still want to distribute descriptors
that use our type system.

Original issue 35 created by ClearTK on 2009-01-21T18:20:01.000Z:

The class:

org.cleartk.syntax.treebank.util.TreebankFormatParser

contains the following regex:

private static final Pattern nonwhiteSpaceCharPattern =
Pattern.compile("[^\s+]");

It looks like the + sign inside the square brackets is misplaced and is
causing problems with treebank data that used to parse. When I remove the
plus sign (I don't think it needs to be there as a quantifier either) then
my data parses (i.e. it doesn't barf the first time a token beginning with
a + sign appears.)

Original issue 34 created by ClearTK on 2009-01-20T20:43:32.000Z:

We should establish a best-practice that requires all configuration
parameters to be read using either
UIMAUtil.getDefaultingConfigParameterValue() or
UIMAUtil.getRequiredConfigParameterValue().

This makes the code more consistent and will avoid heart-breaking mishaps
like reading in an empty string when you are expecting a null (for example).

We should do a pass over all initialize methods and replace all occurrences
of context.getConfigParameterValue.

Original issue 25 created by ClearTK on 2008-12-12T17:47:30.000Z:

the default jar file created by the build script should not include the
large models provided by opennlp. otherwise, you end up with a >40MB jar file.

Original issue 44 created by ClearTK on 2009-02-06T16:16:06.000Z:

I was looking through someone else's UIMA code and noticed that their
parameter names were "somewhat-qualified" wrt the package name you might
find them in. I would like to suggest that we name our parameters with
fully qualified names wrt the class they are defined (and documented) in.
For example,

"ViewName" defined by PlainTextCollectionReader.PARAM_VIEW_NAME should be
named: "org.cleartk.util.PlainTextCollectionReader:ViewName" - or something
simlar.

This will ensure that our parameter names our never ambiguous and provides
a clear way to find where the parameter is defined and documented.

Original issue 6 created by ClearTK on 2008-12-05T18:18:59.000Z:

Please use labels and text to provide additional information.

We need a tutorial/overview of how FeaturesEncoder, OutcomeEncoder,
EncoderFactory, etc. work together.

Original issue 18 created by ClearTK on 2008-12-05T20:35:05.000Z:

Please use labels and text to provide additional information.

The snowball stemmer annotator should grow extra parameters that specify
the token type and the slot where to store the stem. Right now it only
works with the default token type, assuming the "stem" slot.

Original issue 40 created by ClearTK on 2009-02-06T14:56:57.000Z:

The uima wrappers being developed by the opennlp folks are really nice.
Basically, we can rip out our lame postagger wrapper and replace it with a
descriptor file that specifies our type system.

See http://opennlp.cvs.sourceforge.net/viewvc/opennlp/opennlp.uima/

We will likely need to build the code ourselves for now keeping careful
track of which revision we are using and any changes we've made.

Original issue 21 created by ClearTK on 2008-12-05T21:53:49.000Z:

Please use labels and text to provide additional information.

We should only ever use UIMAUtil.getDefaultingConfigParameterValue() or
UIMAUtil.getRequiredConfigParameterValue() when getting component
parameters. This way if we decide we need to do extra checking (e.g. for
empty strings), we can do it all in one place.

Original issue 46 created by ClearTK on 2009-02-06T17:52:09.000Z:

for clarity - the TreebankGoldAnnotator should post treebank annotations to
a "GoldView" rather than the "_InitialView". This makes the names of the
views clearer and simplifies the understanding of the sofa mappings in
aggregate analysis engine descriptors. Generally, if you are using gold
standard data to evaluate system output - you shouldn't assume that the
gold-standard data is in the _InitialView - but rather in some special view

• a "GoldView".

Original issue 8 created by ClearTK on 2008-12-05T18:26:15.000Z:

Please use labels and text to provide additional information.

For bad examples, see:

• ClassifierAnnotator.PARAM_CLASSIFIER_JAR
• DelegatingDataWriter.PARAM_DATA_WRITER
• DataWriter_ImplBase
• InstanceConsumer_ImplBase
• MaxentDataWriter

For good examples see:

• ChunkerHandler

Original issue 13 created by ClearTK on 2008-12-05T19:32:53.000Z:

Please use labels and text to provide additional information.

To train models the following binaries need to be on your path:

• LIBSVM: svm-train
• LIBLINEAR: train
• SVMlight: svm_learn

We should document this somewhere.