Welcome to SwiRL.
SwiRL is free GPL software. Please see the file COPYING for details.
==============================================================================
What is it?
-----------
SwiRL is a Semantic Role Labeling (SRL) system constructed on top of the full
syntactic analysis of text. The syntactic analysis is performed using Eugene
Charniak's parser (included in this package). SwiRL trains one classifier for
each argument label using a rich set of syntactic and semantic features.
The classifiers are learned using one-vs-all AdaBoost classifiers, using
Xavier Carreras' AdaBoost software (included in this package).
When tested on the CoNLL evaluation data, SwiRL obtains an F1 measure of
77.44 on the WSJ testing section and of 66.65 on the Brown testing section.
The execution time is slightly less than 3 seconds/sentence (includes
full parsing and SRL).
Author
------
Mihai Surdeanu, but this package includes also Eugene Charniak's parser and
Xavier Carreras' AdaBoost software.
Installation
------------
Note of advice: this code is known to work on many flavors of Linux (Slack,
Debian, and variants of these) using gcc >= 3.0 AND gcc <= 4.1.0. With anything
else, you're on your own...
The usual ./configure; make; make install. See INSTALL for more details on this
process. Additionally, after completing this process, copy the required model
directories (model_charniak for the parser and model_swirl for the SRL part)
wherever you prefer.
Optional step:
Before running make install, but after make, it is a good idea to run
make test_brown. This tests SwiRL on the Brown testing section from the CoNLL
evaluation data. Make sure you obtain the same F1 measure as the one reported
in the NEWS file. If you want to check SwiRL on the WSJ testing section run
make test_wsj. Note: make test_brown takes about 30 minutes, make test_wsj
takes about 2 hrs.
Command line
------------
(a) Parsing a complete file:
To parse a complete file the command line is the following:
swirl_parse_classify \
<SRL model directory> \
<charniak's parser model directory> \
<file to parse>
For example, if you would like to run SwiRL on the WSJ test section, use:
swirl_parse_classify model_swirl model_charniak testing/test.wsj.input
The SRL model directory contains the set of SRL classifiers (one per argument
label). The parser model directory contains the data required by Charniak's
parser.
SwiRL outputs the processed propositions at the standard output in CoNLL
format.
The input file accepts four possible formats, depending on the amount of
information available:
(a) 0 (word ne pred)+
where each token has lexical and NE information, plus a boolean flag
to indicate if it is (1) or is not (0) a predicate. See
testing/test.wsj.input for an example of this format.
(b) 1 (word pos ne)+
where each token has lexical, POS and NE information. Predicates are
detected on the fly.
(c) 2 (word ne)+
where each token has lexical and NE information. POS tags and predicates
are detected on the fly. This is probably the most common format to be
used through the API.
(d) 3 (word)+
where each token has only lexical information. POS tags and predicates
are detected on the fly. NE information is not used, so the results using
this format will be slightly worse than usual.
(b) Running the interactive shell:
If you would like to test SwiRL without generating the input file, you can
with the following command:
swirl_parse_classify \
<SRL model directory> \
<charniak's parser model directory>
This opens an interactive shell, where you can type your sentence.
Note: the input format must be in one of the above four classes.
(c) Retraining SwiRL:
make train MODEL_DIR=<directory where to store your model files>
Alternatively, if you want to train on a different corpus, run:
make train \
MODEL_DIR=<directory where to store your model files>
WORD_FILE=<CoNLL-like word file> \
NE_FILE=<CoNLL-like NE file> \
CHARNIAK_FILE=<CoNLL-like file with Charniak-generated syntax> \
PROP_FILE=<CoNLL-like file with propositions>
Note: training on the full CoNLL corpus takes up to 2-3 days!
Note: The AdaBoost learner is configured by default with the best parameters
I observed on the CoNLL development data: 1000 rounds of boosting and decision
trees of depth 3. Feel free to play with these parameters if you believe they
should be changed.
Note: In the current configuration SwiRL requires about 5GB of disk space
and 4GB of RAM to train. If you don't have these resources available you
can reduce the number of positive and negative examples to be used by the
trainer in src/lib/Constants.h: POSITIVE_EXAMPLES_MAX_COUNT and
NEGATIVE_EXAMPLES_MAX_COUNT.
(d) Using the API:
The API is defined in the file src/lib/Swirl.h (or
${prefix}/include/swirl/Swirl.h after make install). Two methods are important:
Swirl::initialize(), which fully initializes SwiRL, and Swirl::parse(), which
parses an input sentence. The output is a Tree object, which is basically
a fully-parsed sentence, where every syntactic node has attached a list of
semantic roles for the various predicates in the sentence. The argument list
for a given node can be accessed with Tree::getPredictedArguments().
You can take a look at src/bin/swirlParseAndClassify.cc to see how the API
is to be used.