blob: 2ff8504959834705d6d708d43370ab01b5654495 [file] [log] [blame]
Contents
- Listing of README's in this project
- Introduction
- Building a model
- Building a tag dictionary
- Running the POSTagger analysis engine
- POSTagger.xml
- POSTaggerAggregate.xml
- POSTaggerCPE.xml
- Evaluating a POS tagger
###################################
Listing of README's in this project
###################################
- data/pos/training/README - describes how to create training data for the
part-of-speech (POS) tagger.
- resources/models/README - describes how the pos-tagging models were created
- test/data/README - describes data used for the unit tests
############
Introduction
############
This project provides a UIMA wrapper around the popular OpenNLP part-of-speech
tagger. The UIMA examples project provides a default wrapper from which we have
borrowed liberally. We have created our own wrapper so that it will work better
with our type system and to add features and supporting components.
Additionally, both the OpenNLP package and the UIMA examples OpenNLP wrappers
lack documentation for how to do things like generate training data, build a
part-of-speech tagging model, and build a tag dictionary. The latter in
particular can be very confusing if you are new to OpenNLP. We have
attempted to provide all of the necessary documentation here.
A part-of-speech tagging model is included with this project.
The model derives from a combination of GENIA, Penn Treebank (Wall Street Journal) and anonymized
clinical data per Safe Harbor HIPAA guidelines. Prior to model building, the clinical data was
deidentified for patient names to preserve patient confidentiality. Any person name in the model
will originate from non-patient data sources.
################
Building a model
################
If you wish to build your own part-of-speech model that works with the OpenNLP
part-of-speech tagger you will need to follow these steps:
1) obtain training data - see data/pos/training/README
2) build a model using the training data - see resources/models/README
#########################
Building a tag dictionary
#########################
One thing that can be confusing about the OpenNLP part-of-speech tagger is that
there are two data structures with similar sounding names - Dictionary and
TagDictionary. In short, the Dictionary construct is one that can and should
be ignored while the TagDictionary is one that needs a bit of attention.
A tag dictionary is used when tagging text, not during the training of a POS model.
Unfortunately, OpenNLP does not provide a mechanism for creating a tag
dictionary, so we have provided one. It can be run with the following command:
java org.apache.ctakes.postagger.TagDictionaryCreator <training-data>
<tag-dictionary>
<case-sensitive>
where <training-data> is a file containing pos-of-speech tagged training data
as described in data/pos/training/README
where <tag-dictionary> is the file that will be created, and where the
tag dictionary will be written to
where <case-sensitive> is either 'true' or 'false' depending on whether the
tag dictionary should be case sensitive or not.
For relevant material about the difference between Dictionary and TagDictionary
see the following:
https://sourceforge.net/forum/forum.php?thread_id=1720863&forum_id=9943
https://sourceforge.net/forum/forum.php?thread_id=1894043&forum_id=9943
DefaultPOSContextGenerator.getContext(int, Object[], String[])
OpenNLP provides a default tag dictionary for the English part-of-speech model
called tag.bin.gz which can be downloaded from:
http://opennlp.sourceforge.net/models/english/parser/tagdict
You should use this tag dictionary only if you are using the model from:
http://opennlp.sourceforge.net/models/english/parser/tag.bin.gz
If we want to use the tag dictionary in a case insensitive way, then entries
in the tag dictionary which are not all lowercased will be ignored because
the tag dictionary fails to lowercase entries read in from the file.
It only lowercases the words that are compared against the dictionary when
"CaseSensitive" is set to false. Therefore, if you want the tag dictionary
to be used in a case insensitive way, be sure to build the tag dictionary
using 'false' as the third argument (as described above).
#####################################
Running the POSTagger analysis engine
#####################################
%%%%%%%%%%%%%
POSTagger.xml
The file desc/POSTagger.xml provides a descriptor for the POSTagger analysis
engine which is the UIMA component we have written that wraps the OpenNLP
part-of-speech tagger. Open this file using the Component Descriptor Editor as
described in the tutorial. Click on the tab labeled "Overview" to observe that
the class called by this descriptor is "org.apache.ctakes.postagger.POSTagger".
Click on the tab labeled "Parameter Settings" to view the parameters required
by the POSTagger component. The descriptor file does not document the
parameters because they are documented in the api javadocs for the class
org.apache.ctakes.postagger.POSTagger. Please consult that documentation for
additional details. The parameters are:
- PosModelFile - the file that contains the part-of-speech tagging model
- TagDictionary - the file that contains the tag dictionary (if available)
- CaseSensitive - determines whether to use the TagDictionary in a case
sensitive way or not.
%%%%%%%%%%%%%%%%%%%%%%
POSTaggerAggregate.xml
The descriptor desc/POSTaggerAggregate.xml defines a pipeline for part-of-speech
tagging that creates all the necessary inputs (e.g. token and sentence
annotations).
Open it using the Component Descriptor Editor as described in the tutorial.
Click on the tab labeled "Overview" to observe that the engine type is
"Aggregate".
Click on the tab labeled "Aggregate" to see the components that need to be run
before the POSTagger can run.
Click on the tab labeled "Parameter Settings" to see that the same three
parameters need to be set from the POSTagger.xml file. If you set these
parameters to acceptable values, you can open and run
desc/POSTaggerAggregate.xml using the CAS Visual Debugger.
%%%%%%%%%%%%%%%%
POSTaggerCPE.xml
The file desc/POSTaggerCPE.xml provides an xml-specification of a component
processing engine (CPE) which can be opened, edited, and run using the UIMA
CPE GUI as described in the tutorial. Open this file using the UIMA CPE GUI
and set the parameters for the collection reader to point to a local collection
of files that you want part-of-speech tagged. Set the parameters for the
POSTagger as appropriate for your environment and, finally, set the output
directory of the XCAS Writer CAS Consumer.
The results of running the pipeline are written to the output directory
as XCAS files.
These files can be viewed in the CAS Visual Debugger.
#######################
Evaluating a POS tagger
#######################
There are two ways a POS tagger should be evaluated:
1) Use gold standard tokens
Run the POS tagger using gold standard tokens and calculate the percentage of
part-of-speech labels that have been correctly assigned.
If this is gold standard sentence:
The_DT major_JJ inducible_JJ protein_NN complex_NN that_WDT binds_VBZ ._.
And if this is the output for that sentence:
The_DT major_JJ inducible_NN protein_NN complex_NN that_WDT binds_VBD ._.
The accuracy should be 6/8 = 75%.
2) Use tokenizer generated tokens
Run the tokenizer and use this as input to the POS tagger. In this scenario, we calculate
f-measure in the following way:
true positive - a token that has the correct boundary and part-of-speech label
false positive - a tagged token that does not have the correct boundary and/or
part-of-speech label
false negative - a token in the gold standard data that was not correctly generated by
the tokenizer/POS tagger
If this is gold standard sentence:
This_DT complex_NN is_VBZ not_RB cyclosporin_JJ -sensitive_JJ ._.
And if this is the output for that sentence:
This_DT complex_JJ is_VBZ not_RB cyclosporin-sensitive_JJ ._.
true positives = 4
false positives = 2
false negatives = 3
f-measure = (2 * recall * precision) / (precision + recall)
= (2*TP) / (2*TP + FP + FN)
f-measure = (2*4) / (2*4 + 2 + 3) = 8 / 13 = .615
In fact, if you do the evaluation this way for the "gold standard tokens" evaluation, then
you will get the same answer as the accuracy calculation given above.