Apertium separable
Lttoolbox provides a module for reordering separable/discontiguous multiwords and processing them in the pipeline. Multiwords are manually written in an additional xml-format dictionary.
Installing
Prerequisites and compilation are the same as lttoolbox and apertium. See Installation. On Debian/Ubuntu derivatives, it is part of the nightly repo as apt-get install apertium-separable.
The code can be found at https://svn.code.sf.net/p/apertium/svn/branches/apertium-separable and instructions for compiling the module are:
./autogen.sh ./configure make make install
You'll need lttoolbox from SVN (or, greater than the current release 3.3.3) and associated libraries, and zlib (debian: zlib1g-dev).
It is not currently part of distributed Apertium binaries for other distros/OSs. It is now available via the nightly repositories as the apertium-separable module.
Lexical transfer in the pipeline
lsx-proc runs directly AFTER apertium-tagger and apertium-pretransfer:
(note: previously this page had said that lsx-proc runs between BETWEEN apertium-tagger and apertium-pretransfer. it has now been determined that it should run AFTER pretransfer.)
… | apertium-tagger -g en-es.prob | apertium-pretransfer | lsx-proc en-es.autoseq.bin | …
Usage
Creating the lsx-dictionary
The lsx dictionary format is largely similar to those of the morphological and bilingual dictionaries. (see also: Apertium_New_Language_Pair_HOWTO)
We begin with a declaration of the dictionary. There is currently nothing in it, only a declaration that we want to begin a new dictionary.
<dictionary type="separable"> </dictionary>
Then add the alphabet entry, this can be empty as the alphabet is only used for tokenisation and the lsx module comes after the text is tokenised. Now we have:
<dictionary type="separable">
<alphabet></alphabet>
</dictionary>
Next we need to add the symbol definitions, abbreviated to sdefs. These are the symbols that your words are tagged with, e.g. noun or verb or adj. Again, you should be able to just copy the sdef section from your language's monodix, and it should contain many more than in this basic example.
<dictionary type="separable">
<alphabet></alphabet>
<sdefs>
<sdef n="adj"/>
<sdef n="adv"/>
<sdef n="n"/>
<sdef n="sep"/>
<sdef n="vblex"/>
</sdefs>
</dictionary>
Now we need to add the paradigm definitions, abbreviated to pardefs. These represent patterns of word orders. The following example represents words tagged as adjective, noun, noun phrase, and frequency adjectives. See the note below about the tags <w/>, <t/>, <j/>. The lemma can be represented as anychars (<w/>, such as in adj and n below; or by typing out the word itself, such as in freq-adv below. Pardefs can be used to create other pardefs, such as in SN below. Adding paradigms into the dictionary, we get:
<dictionary type="separable">
<alphabet></alphabet>
<sdefs>
...
</sdefs>
<pardefs>
<pardef n="adj"> <!-- to represent all adjectives -->
<e><i><w/><s n="adj"/><j/></i></e> <!-- word only has the adj tag -->
<e><i><w/><s n="adj"/><t/><j/></i></e> <!-- word has the adj tag followed by one or more other tags -->
</pardef>
<pardef n="n"> #to represent all nouns
<e><i><w/><s n="n"/><t/><j/></i></e> <!-- word has the n tag followed by one or more other tags -->
</pardef>
<pardef n="SN"> #to represent all noun phrases
<e><par n="n"/></e>
<e><par n="adj"/><par n="n"/></e> <!-- word phrase is comprised of an adjective word followed by a noun word -->
<e><par n="adj"/><par n="adj"/><par n="n"/></e> <!-- word phrase is comprised of two adjectives followed by a noun -->
</pardef>
<pardef n="freq-adv">
<e><i>always<s n="adv"/><j/></i></e> <!-- i.e. ^always<adv>$ -->
<e><i>anually<s n="adv"/><j/></i></e>
<e><i>bianually<s n="adv"/><j/></i></e>
</pardef>
</pardefs>
</dictionary>
Finally, we add the main entries. Here is the final result of our small example dictionary:
<dictionary type="separable">
<alphabet></alphabet>
<sdefs>
<sdef n="adj"/>
<sdef n="adv"/>
<sdef n="n"/>
<sdef n="sep"/>
<sdef n="vblex"/>
</sdefs>
<pardefs>
<pardef n="adj">
<e><i><w/><s n="adj"/><j/></i></e>
<e><i><w/><s n="adj"/><t/><j/></i></e>
</pardef>
<pardef n="n">
<e><i><w/><s n="n"/><t/><j/></i></e>
</pardef>
<pardef n="SN">
<e><par n="n"/></e>
<e><par n="adj"/><par n="n"/></e>
<e><par n="adj"/><par n="adj"/><par n="n"/></e>
</pardef>
<pardef n="freq-adv">
<e><i>always<s n="adv"/><j/></i></e>
<e><i>anually<s n="adv"/><j/></i></e>
<e><i>bianually<s n="adv"/><j/></i></e>
</pardef>
</pardefs>
<section id="main" type="standard">
<e lm="be late" c="llegar tarde">
<p><l>be<s n="vbser"/></l><r>be<g><b/>late</g><s n="vbser"/></r></p><i><t/><j/></i>
<par n="SAdv"/><p><l>late<t/><j/></l><r></r></p>
</e>
<e lm="take away" c="sacar, quitar">
<p><l>take<s n="vblex"/></l><r>take<g><b/>away</g><s n="vblex"/></r></p><i><t/><j/></i>
<par n="SN"/><p><l>away<t/><j/></l><r></r></p>
</e>
</section>
</dictionary>
Note:
<w/>stands for one or more alphabetic symbols<t/>stands for one or more tags (multicharacter symbols).<j/>stands for the word boundary symbol $
i.e.
<e><w/>is equivalent to<t/><j/></e>any-one-or-more-chars<adj><required-anytag><...optional-anytag...><$>- ^tall<adj><sint><...>$
<e><w/>is equivalent to<j/></e>any-one-or-more-chars<adj><$>- ^tall<adj>$
A larger example dictionary can be found at https://svn.code.sf.net/p/apertium/svn/branches/apertium-separable/examples/apertium-eng-spa.eng-spa.lsx
The lsx dictionary file names are of the form apertium-A-B.A-B.lsx , where apertium-A-B is the name of the language pair. For example, file apertium-eng-cat.eng-cat.lsx is the lsx dictionary for the eng-cat pair. The names of the compiled binaries are of the form apertium-A-B.autoseq.bin . For example, eng-cat.autoseq.bin .
Compilation
Compilation into the binary format is achieved by means of the lsx-comp program.
$ lsx-comp apertium-eng-spa.eng-spa.lsx eng-spa.autoseq.bin main@standard 61 73
Processing
Processing can be done using the lsx-proc program.
The input to lsx-proc is the output of apertium-tagger and apertium-pretransfer ,
$ echo '^take<vblex><imp>$ ^prpers<prn><obj><p3><nt><sg>$ ^out of<pr>$ ^there<adv>$^.<sent>$' | lsx-proc eng-spa.autoseq.bin ^take# out<vblex><sep><imp>$ ^prpers<prn><obj><p3><nt><sg>$ ^of<pr>$ ^there<adv>$^.<sent>$
Example usages
Example #1: A sentence in plain text,
The Aragonese took Ramiro out of a monastery and made him king.
This is the output of feeding the sentence through apertium-tagger and then apertium-pretransfer :
^the<det><def><sp>$ ^Aragonese<n><sg>$ ^take<vblex><past>$ ^Ramiro<np><ant><m><sg>$ ^out of<pr>$ ^a<det><ind><sg>$ ^monastery<n><sg>$ ^and<cnjcoo>$ ^make<vblex><pp>$ ^prpers<prn><obj><p3><m><sg>$ ^king<n><sg>$^.<sent>$
This is the output of feeding the output above through lsx-proc with apertium-eng-spa.eng-spa.lsx:
^the<det><def><sp>$ ^Aragonese<n><sg>$ ^take# out<vblex><sep><past>$ ^Ramiro<np><ant><m><sg>$ ^of<pr>$ ^a<det><ind><sg>$ ^monastery<n><sg>$ ^and<cnjcoo>$ ^make<vblex><pp>$ ^prpers<prn><obj><p3><m><sg>$ ^king<n><sg>$^.<sent>$
Troubleshooting
Segmentation fault
Segmentation fault upon compilation or usage
The lsx-dictionary compiles fine with zero entries but gives a seg fault once entries are added
...no solution found yet
something is not updated or something in the makefile (?)
make sure that the makefile ...
Complaints about step_override()
svn update in lttoolbox (and do make, make install)
You'll need an up-to-date version of lttoolbox and associated libraries, and zlib (debian: zlib1g-dev).
Undefined symbol
In your dictionary you are probably using a symbol that you didn't define in the sdefs. Add the symbol to the sdefs.
Future work
Offloading multiwords from transducers to lsx
In theory we're offloading multiwords from the transducers to lsx. This leaves open some questions:
- how do we do N N compounds with lsx?
- how does translation to a multiword work? In theory it's possible to invert the transducer, but an attempt to try this results in a transducer that looks right but silently fails to apply to input. Also, it will need to be able to handle the output of transfer. —Firespeaker (talk) 00:02, 1 September 2017 (CEST)
Recycling dictionaries and/or paradigms
lsx-dictionaries are packaged in language pairs. the eng-spa lsx-dictionary can mostly be reaped by eng-cat. could we make use of the similarity?
Beta testing
Support for language pairs: we haven't gotten much extensive beta testing. The following are language pairs that have packaged the lsx-module:
- eng-cat
- eng-deu (?)
- kaz-kir
Beta test with more language pairs
Transfer-like super powers
- Transfer-like capabilities for the lexicon (super powers). E.g., gustar / like
Inheritance and clean up code
- [lsx-comp] was hitched from the lttoolbox/compiler. In the future we may want to integrate it back and have lsx-comp and lsx-proc inherit directly from lttoolbox (?)
The one-to-many bug
Given the following lsx file:
<dictionary type="sequential"> <alphabet>АӘБВГҒДЕЁЖЗИІЙКҚЛМНҢОӨПРСТУҰҮФХҺЦЧШЩЬЫЪЭЮЯаәбвгғдеёжзиійкқлмнңоөпрстуұүфхһцчшщьыъэюя</alphabet> <sdefs> <sdef n="adj"/> <sdef n="adv"/> <sdef n="n"/> <sdef n="nom"/> <sdef n="dat"/> <sdef n="v"/> </sdefs> <pardefs> <pardef n="adj"> <e><i><w/><s n="adj"/><j/></i></e> <e><i><w/><s n="adj"/><t/><j/></i></e> </pardef> <pardef n="n"> <e><i><w/><s n="n"/><t/><j/></i></e> </pardef> <pardef n="SN"> <e><par n="n"/></e> <e><par n="adj"/><par n="n"/></e> <e><par n="adj"/><par n="adj"/><par n="n"/></e> </pardef> </pardefs> <section id="main" type="standard"> <e lm="кабарда" c="хабар ет"> <p><l>хабар<b/>ет<s n="v"/></l> <r>хабар<s n="n"/><s n="nom"/><j/>ет<s n="v"/></r></p><i><t/><j/></i> </e> <e lm="абайла" c="абай бол"> <p><l>абай<b/>бол<s n="v"/></l> <r>абай<s n="adj"/><j/>бол<s n="v"/></r></p><i><t/><j/></i> </e> <e lm="абайла" c="абай бол"> <p><l>абай<b/>бол<s n="v"/></l> <r>абай<s n="adj"/><j/>бол<s n="v"/></r></p><i><t/>+ма<t/><j/></i> <!-- p><l>абай<s n="adj"/><j/>бол<s n="v"/><t/></l> <r>абай<b/>бол<s n="v"/><t/></r></p --> </e> <e lm="сууга түш" c="шомылда"> <p><l>сууга<b/>түш<s n="v"/></l> <r>суу<s n="n"/><s n="dat"/><j/>түш<s n="v"/></r></p><i><t/><j/></i> </e> </section> </dictionary>
and the following code to compile it (where $(PREFIX1) is kaz-kir and $(PREFIX2) is kir-kaz and $(BASENAME) is apertium-kaz-kir; the above file is apertium-kaz-kir.kir-kaz.lsx):
$(PREFIX1).autoseq.bin: $(BASENAME).$(PREFIX1).lsx lsx-comp $< $@ $(PREFIX2).autoseq.bin: $(BASENAME).$(PREFIX2).lsx lsx-comp $< $@ $(PREFIX1).revautoseq.bin: $(BASENAME).$(PREFIX1).lsx lt-print $(PREFIX1).autoseq.bin | sed 's/ /@_SPACE_@/g' > $(PREFIX1).autoseq.att hfst-txt2fst -e ε < $(PREFIX1).autoseq.att > $(PREFIX1).autoseq.hfst hfst-invert $(PREFIX1).autoseq.hfst | hfst-minimise > $(PREFIX1).revautoseq.hfst hfst-fst2txt $(PREFIX1).revautoseq.hfst | gzip -9 -c -n > $(PREFIX1).revautoseq.att.gz zcat < $(PREFIX1).revautoseq.att.gz > $(PREFIX1).revautoseq.att sed 's/@0@/ε/g' $(PREFIX1).revautoseq.att > $(PREFIX1).revautoseq.1.att lt-comp lr $(PREFIX1).revautoseq.1.att $@ $(PREFIX2).revautoseq.bin: $(BASENAME).$(PREFIX2).lsx lt-print $(PREFIX2).autoseq.bin | sed 's/ /@_SPACE_@/g' > $(PREFIX2).autoseq.att hfst-txt2fst -e ε < $(PREFIX2).autoseq.att > $(PREFIX2).autoseq.hfst hfst-invert $(PREFIX2).autoseq.hfst | hfst-minimise > $(PREFIX2).revautoseq.hfst hfst-fst2txt $(PREFIX2).revautoseq.hfst | gzip -9 -c -n > $(PREFIX2).revautoseq.att.gz zcat < $(PREFIX2).revautoseq.att.gz > $(PREFIX2).revautoseq.att sed 's/@0@/ε/g' $(PREFIX2).revautoseq.att > $(PREFIX2).revautoseq.1.att lt-comp lr $(PREFIX2).revautoseq.1.att $@
$ echo "^хабар ет<v><iv><ifi><p1><sg>$" | lsx-proc kaz-kir.revautoseq.bin
does not seem to work. The expected output is ^хабар<n><nom>$ ^ет<v><iv><ifi><p1><sg>$. The reverse does seem to work:
$ echo "^хабар<n><nom>$ ^ет<v><iv><ifi><p1><sg>$" | lsx-proc kaz-kir.autoseq.bin
Indeed outputs ^хабар ет<v><iv><ifi><p1><sg>$.
