====== Implementation of the DMV+CCM Parser ======
===== Introduction =====
This package includes implementations of the CCM, DMV and DMV+CCM parsers from Klein and Manning (2004), and code for testing them with the WSJ, Negra and Cast3LB corpuses (English, German and Spanish respectively). A detailed description of the parsers can be found in Klein (2005).
This work was done as part of the PhD in Computer Science I am doing at FaMAF, Universidad Nacional de Cordoba, Argentina, under the supervision of Gabriel Infante-Lopez, with a research fellowship from CONICET.
All the software is distributed under the [[http://www.gnu.org/copyleft/gpl.html|GNU GPL v3 license]].
===== About Version 0.2.0 =====
This version is aimed at reproducing some of the results from Klein and Manning (2004). The implemented DMV and DMV+CCM models are the versions with the one-side-first constraint. We couldn't reproduce yet the results for these models without the one-side-first constraint.
The following table shows the performance of the models over the WSJ10 corpus:
^ WSJ10 ^ UP ^ UR ^ UF1 ^ Dir ^ Undir ^ | DMV | 58.3 | 74.1 | 65.3 | 49.0 | 65.5 | | CCM | 64.3 | 81.6 | 71.9 | | DMV+CCM | 67.9 | 86.2 | 75.9 | 47.0 | 64.5 |
DMV reaches the given values at the 20th iteration. CCM converges to the given values since the 40th iteration. DMV+CCM's preformance starts to decrease at the 9th training iteration, reaching a peak of UF1 = 76.1%, Dir = 47.2%. The results from the table for DMV+CCM were arbitrarily picked from the 10th iteration.
===== Dependencies =====
* ntlk: http://www.nltk.org * lq-nlp-commons: http://www.cs.famaf.unc.edu.ar/~francolq.
===== Installation and Configuration =====
Extract lq-nlp-commons.tar.gz and lq-dmvccm.tar.gz into a folder and add the new folders to the Python PATH. For instance:
# tar -zxvf lq-nlp-commons.tar.gz # tar -zxvf lq-dmvccm.tar.gz # export PYTHONPATH=`pwd`/lq-nlp-commons-0.1.0/:`pwd`/lq-dmvccm-0.1.0/
===== Usage =====
==== Quickstart ====
To train (with 10 iterations) and test the CCM parser with the WSJ10 do the following (replace 'my_wsj_path' with the path to the WSJ corpus):
# python >>> import wsj10 >>> tb = wsj10.WSJ10(basedir='my_wsj_path') >>> from dmvccm import ccm >>> m = ccm.CCM(tb) >>> m.train(10) >>> m.test()
To parse a sentence with the resulting parser do something like this:
>>> s = 'DT NNP NN VBD DT VBZ DT JJ NN'.split() >>> (b, p) = m.parse(s) >>> b.brackets set([(6, 9), (1, 3), (4, 9), (4, 6), (3, 9), (0, 3), (7, 9)]) >>> t = b.treefy(s) >>> t.draw()
==== Parser Instantiation, Training and Testing ====
The WSJ10 corpus is used by default to train the parsers. The WSJ10 corpus is automatically extracted from the WSJ corpus. You must place the WSJ corpus in a folder named wsj_comb (or edit lq-nlp-commons/wsj.py, or read below).
For instance, to train (with 10 iterations) and test the CCM parser with the WSJ10 do the following:
# python >>> from dmvccm import ccm >>> m = ccm.CCM() >>> m.train(10) >>> m.test()
To give the treebanks explicitly:
>>> import wsj10, negra10, cast3lb10 >>> tb1 = wsj10.WSJ10() >>> tb2 = negra10.Negra10() >>> tb3 = cast3lb10.Cast3LB10() >>> m1 = ccm.CCM(tb1) >>> m2 = ccm.CCM(tb2) >>> m3 = ccm.CCM(tb3)
The Negra corpus must be in a file negra-corpus/negra-corpus2.penn (in Penn format), and the Cast3LB corpus in a folder named 3lb-cast.
To use alternative locations for the treebanks use the parameter basedir when creating the object. For instance:
>>> import wsj10 >>> tb = wsj10.WSJ10(basedir='my_wsj_path')
(similarly for Negra and Cast3LB)
When loaded for first time, the extracted corpuses are saved into files to avoid having to process the entire treebanks again. The files are saved in the NLTK data path (nltk.data.path), usually $HOME/nltk_data, to be loaded in future instantiations of the treebanks.
The DMV and DMV+CCM parsers are in the classes dmv.DMV and dmvccm.DMVCCM. The current implementation of DMV+CCM has the one-side-first constraint. They are used the same way CCM is. For instance:
>>> from dmvccm import dmv, dmvccm >>> m1 = dmv.DMV() >>> m1.train(10) >>> m1.test() >>> m2 = dmvccm.DMVCCM() >>> m2.train(10) >>> m2.test()
Take into account that training DMV and DMV+CCM is much slower than training CCM. A single training step can take more than 20 minutes.
==== Parser Usage ====
Once you have a trained instance of CCM, DMV or DMV+CCM, you can parse sentences with the parse() method. You may give a list of words or an instance of the sentence.Sentence class from lq-nlp-commons. The parse() method returns a pair (b, p), where b is an instance of bracketing.Bracketing and p is the probability of the bracketing. You can get the set of brackets from b.brackets or convert the bracketing to a tree with b.treefy().
>>> from dmvccm import ccm >>> m = ccm.CCM() >>> m.train(2) >>> s = 'DT NNP NN VBD DT VBZ DT JJ NN'.split() >>> (b, p) = m.parse(s) >>> b.brackets set([(6, 9), (1, 3), (4, 9), (4, 6), (3, 9), (0, 3), (7, 9)]) >>> t = b.treefy(s) >>> t.draw()
In the case of DMV and DMV+CCM you can also use the method dep_parse() to get the dependency structure parsed by these models (Klein, 2005).
>>> from dmvccm import dmv >>> m = dmv.DMV() >>> m.train(2) >>> s = 'DT NNP NN VBD DT VBZ DT JJ NN'.split() >>> (t, p) = m.dep_parse(s) >>> t.draw()
==== References ====
Franco M. Luque. Una Implementación del Modelo DMV+CCM para Parsing No Supervisado (2011). 2do Workshop Argentino en Procesamiento de Lenguaje Natural, Córdoba, Argentina.
Klein, D. and Manning, C. D. (2004). Corpus-based induction of syntactic structure: Models of dependency and constituency. In ACL, pages 478-485.
Klein, D. (2005). The Unsupervised Learning of Natural Language Structure. PhD thesis, Stanford University.
- Changes to previous version:
Initial Announcement on mloss.org.
No one has posted any comments yet. Perhaps you'd like to be the first?
Leave a comment
You must be logged in to post comments.