Target-language tagger training
The package apertium-tagger-training-tools trains taggers based on both source and target language information. The resulting probability files are as good as supervised training for machine translation purposes, but much quicker to produce, and with less effort (you don't have to manually tag a corpus). In this description, a part-of-speech tagger for the source language (SL) will be trained using information from the target language (TL).
Language pair[edit]
This example presumes that you want to train a tagger for the Occitan ←→ Catalan (apertium-oc-ca) language pair in the Occitan → Catalan (oc-ca) direction. You will need to substitute values that refer to that pair with those for your chosen language pair.
You will need to download and install the language pair in question either from SVN or from a package. The method implemented in this package is appropriate for those language whose part-of-speech tagger was trained in an unsupervised manner.
To prepare and compile the required language-pair data follow the instruction provided at the linguistic package itself. Usually you only need to type ./configure and make.
Building a target language model[edit]
If you're using apertium-trigrams-langmodel, then follow this section, if not, continue to the next.
Requirements: A raw corpus of the target language (ca.corpus.txt). If you are generating oc→ca, this would be ca, if you're generating a tagger for ca→es this would be es. The corpus should be around 0.5 million words, bigger than this does not result in significant improvement in accuracy.
$ apertium-trigrams-langmodel -t -i ca.corpus.txt > catalan.lm
The following output should appear:
LOCALE: en_GB.UTF-8 Training........................................................................ ................................................................................ ................................................................................ ................................................................................ ................................................................................ ................................................................................ ................................................................................ ...................... 2868350 processed words. 1643982 distinct trigrams found. 833143 distinct bigrams found. 153714 distinct monograms found. Writing word2id information ... Writing Simple Good-Turing model for 1-grams ... Writing 1-grams ... Writing 2-grams... Writing 3-grams...
The dots will continue while the language model is being constructed, and the number will vary.
Preparing the source-language data[edit]
Requirements:
- A raw corpus of the source language (corpus.txt — for example
oc.corpus.txt) - A source dictionary (dic.dix — e.g.
apertium-oc-ca.oc.dix) - A compiled dictionary of the source language (dic.bin — e.g.
oc-ca.automorf.bin) - A tagger definition file (tagger.tsx — e.g.
apertium-oc-ca.oc.tsx) - The file with the transfer rules to be used (trules.xml or .t1x file — e.g.
apertium-oc-ca.oc-ca.t1x)
Note: dic.bin was generated when preparing the language pair data. dic.dix, tagger.tsx and trules.xml are provided with the language-pair package.
Generate the corpus file[edit]
$ apertium-tagger-gen-crp-file oc.corpus.txt dic.bin > lang.crp
Should give the output:
Generating crp file This may take some time. Please, take a cup of coffee and come back later.
Generate the dic file[edit]
$ apertium-tagger-gen-dic-file dic.dix dic.bin tagger.tsx > lang.dic
Should give the output:
Generating dic file This may take some time. Please, take a cup of coffee and come back later.
Extract regex rules[edit]
$ apertium-xtract-regex-trules trules.xml > regexp-trules.txt
Should give no output, but leave the file regexp-trules.txt
Preparing the translation script[edit]
At this point you should have the following files:
$ ls -1 -sh total 132M 78M corpus.crp 38M ca.corpus.lm 17M corpus.txt 16K oc.dic 8.0K regex-trules.txt
The TL-driven training algorithm needs to be provided with a translation script conveniently configured to translate a given disambiguation hypothesis into the TL.
Together with this packages you can find the script 'translation-script-es-ca.sh'. It is configured to translates Spanish disambiguation hypothesis into Catalan.
Copy this file from the example/ directory and then edit this file and change the DATA and DIRECTION variables. DATA must point to the folder holding the language-pair data previously prepared; DIRECTION must store the translation direction.
Assuming you are in the oc-tagger-data directory:
$ cp apertium-tagger-training-tools/example/translation-script-es-ca-batch-mode.sh . $ mv translation-script-es-ca-batch.sh translation-script-oc-ca-batch-mode.sh
Then edit the translation-script-oc-ca-batch-mode.sh and change the above variables.
Three-stage transfer[edit]
Note: The apertium-tagger-training-tools package does not currently work properly with three-stage transfer. It is possible to use it, by making the following changes to the translation-script, but the segmentation will be wrong and this is likely to effect the final quality of the tagger.
Change the line:
apertium-transfer $DATA/trules-$DIRECTION.xml $DATA/trules-$DIRECTION.bin $AUTOBIL |\
To:
apertium-transfer $DATA/$DIRECTION.t1x $DATA/$DIRECTION.t1x.bin $DATA/$DIRECTION.autobil.bin |\ apertium-interchunk $DATA/$DIRECTION.t2x $DATA/$DIRECTION.t2x.bin |\ apertium-postchunk $DATA/$DIRECTION.t3x $DATA/$DIRECTION.t3x.bin |\
Preparing the likelihood script[edit]
To estimate the likelihood of each translation the TL-driven algorithm is provided with an script. In this package you can find an example of this script called 'likelihood-script-catalan.sh'. It uses the apertium-trigrams-langmodel package to calculate the likelihood of each input string.
Again, assuming you're in the oc-tagger-data directory:
$ cp ../../apertium-tagger-training-tools/example/likelihood-script-catalan-batch-mode.sh . $ mv likelihood-script-catalan.sh likelihood-script-occitan-batch-mode.sh
Change this script to use the desired data or to use another language model. You'll need to change the LMDATA variable to be ca.corpus.lm. Keep in mind that the TL-driven algorithm will provide an input TL string to the script and that it expects a likelihood, i. e. a double value, conveniently formated using the appropriate locale.
Training through the TL-driven algorithm[edit]
After you have all of that collected you can generate the .prob file using the following commands:
Warning: Some language pairs perform some orthographical operations after the transfer module. In those cases is a good idea to provide the superficial forms (words) involved in those operations through the --supforms parameter.
In the following examples the --file argument specifies the prefix of the main files used, so for example if you have: corpus.crp, corpus.dic, and corpus.txt the prefix will need to be 'corpus'.
Commands[edit]
- Without disambiguation hypothesis pruning
$ apertium-tagger-tl-trainer --train 500000 \
--tsxfile tagger.tsx \
--file <prefix> \
--tscript ./translation-script.sh \
--lscript ./likelihood-script.sh \
--trules regexp-trules.txt
- With disambiguation hypothesis pruning
To do this you will need an intial model, initialmodel.prob, estimated through another training method (Kupiec, Baum-Welch, ...) — creating this initial model is described in the article about tagger training.
$ apertium-tagger-tl-trainer --tsxfile tagger.tsx \
--train 500000 \
--prune 1 1000 0.6 1 \
--initprob initialmodel.prob \
--file <prefix> \
--tscript ./translation-script.sh \
--lscript ./likelihood-script.sh \
--trules regexp-trules.txt
Output[edit]
This is some example output from the training process. If what you've got looks like this, then you're on the right track!
Command line: apertium-tagger-tl-trainer --train 300000 --tsxfile ../apertium-oc-ca.oc.tsx \ --file corpus --tscript ./translation-script-oc-ca-batch.sh --lscript ./likelihood-script-catalan-batch.sh \ --trules regex-trules.txt --norules --gen-paths oc.PATHS Reading transfer rules from file 'regex-trules.txt' done. Calculating ambiguity classes ... 92 states and 283 ambiguity classes Target-language driven HMM-based part-of-speech training method....... Training corpus will be processed for 500000 words HMM parameters will be calculated each time 0 words are processed Calculated parameter will be saved in files with the name 'corpus.N.prob' Are fobidden and enforce rules going to be used? 1 Translation script is: './translation-script-oc-ca-batch.sh' Likelihood estimation script is: './likelihood-script-catalan-batch.sh' Ready for training...... go! Initialising allowed bigrams ... done. Error: conversion error Error: conversion error Error: There is a path with a null translation: ^se<prn><pro><ref><p3><mf><sp>$ ^èsser<vbser><pri><p3><sg>$ ^pas<adv>$ ^que<cnjsub>$ ^lo<det><def><m><sg>$ ^miralh<n><m><sg>$ ^fisèl<adj><m><sg>$ ^de<pr>$ SEGMENT: s' es pas que lo miralh fisèl de Warning: This segment has no OK translations. Skipping SEGMENT: s' es pas que lo miralh fisèl de Warning: This segment has no translations into TL 1. Skipping SEGMENT: un parlaire adult Warning: This segment has no OK translations. Skipping SEGMENT: a de Error: There is a path with a null translation: ^çò<detnt>$ ^de<pr>$ SEGMENT: çò d' Warning: This segment has no OK translations. Skipping SEGMENT: çò d' Warning: This segment has no OK translations. Skipping SEGMENT: que pòdon instituir la lenga . Warning: This segment has no OK translations. Skipping SEGMENT: a l' un còp l' instituir e
Troubleshooting[edit]
Error: pcre_compile missing )
This does not mean that pcre_compile is missing, but that pcre_compile didn't find a ) that it expected.
Errors like this can pop up if your trules.xml have unescaped #'s in them. Use \# instead.
