13
16
inside of this directory.
17
To use this language-pair package with Apertium YOU DO NOT NEED TO
18
RETRAIN THE TAGGER. Probabilities and auxiliary data are provided for
19
both the oc-ca and the ca-oc translation directions which should be
20
acceptable for most applications, and should work even if you change
21
the dictionaries in a reasonably way.
23
If for some reason you need to retrain the tagger (for example, you
24
have made really extensive changes to the dictionaries such as
25
creating new lexical categories), you have three alternatives:
27
* To perform a supervised training:
29
To this end you need the files specified in the README file inside
30
oc-tagger-data and ca-tagger-data which are not provided. When performing
31
a supervised training, tagged corpora(oc-tagger-data/oc.tagged and
32
ca-tagger-data/ca.tagged) could be obsolete for some words. If this is the
33
case, the tagger training program will show you where the problems are and
34
you will need to solve them by hand. Be sure to solve the problems by
35
modifying ONLY the .tagged file, NEVER the .untagged file that is
36
automatically generated.
38
The supervised training is done by typing:
39
make -f oc-ca-supervised.make (for the Occitan part-of-speech tagger)
40
make -f ca-oc-supervised.make (for the Catalan part-of-speech tagger)
42
This is the training method followed to train the Catalan
43
part-of-speech tagger.
45
* To perform a classical (expectation-maximization) unsupervised training:
47
For this purpose you will need to assemble a large (hundreds of
48
thousand of words) plain-text corpus for each language (for example,
49
using a robot to harvest text from online newspapers) and put them in
50
the proper place, for instance oc-tagger-data/oc.crp.txt and
51
ca-tagger-data/ca.crp.txt. This type of training does not need human
52
intervention but, as expected, results will be less adequate than
53
those obtained with the supervised training.
55
The unsupervised training is done through the iterative Baum-Welch
56
algorithm. By default the number of iterations is set to 8, but you
57
can change this value by editing the Makefile and changing the
58
value of TAGGER_UNSUPERVISED_ITERATIONS.
60
The unsupervised training is done by typing:
61
make -f oc-ca-unsupervised.make (for the Occitan part-of-speech tagger)
62
make -f ca-oc-unsupervised.make (for the Catalan part-of-speech tagger)
64
* To perform an unsupervised training by using target-language
65
information and the rest of the modules of the Apertium MT engine:
67
To do so you need large plain-text corpora on both languages. Please
68
download the apertium-tagger-training-tools package and follow the
69
instructions provided there. This is the training method followed to
70
train the Occitan part-of-speech tagger.
18
===================================================================
20
More information about this module, and others can be found on
21
the Apertium: Wiki, http://wiki.apertium.org