Difference between revisions of "User:Mathematic-alpha/gsoc-report"

From Apertium
Jump to navigation Jump to search
Line 97: Line 97:
= Project description =
= Project description =


A project on adoption of Guarani-Spanish language pair in Apertium had as its purpose a creation of machine translation system between Guarani and Spanish languages. As Guarani is one of the low-resource languages of the world the translation system is unlikely to be developed by other methods than Rule-Based Machine Translation. As the evidence of that we had only about 2800 texts from Wikipedia dumps [https://dumps.wikimedia.org/backup-index.html] and Guarani-Spanish aligned Bible [https://www.bible.com/es/versions/972-guabd-biblia-guarani-tumpa-inee] as a source.
A project on adoption of Medumba-Français language pair in Apertium had as its purpose a creation of machine translation system between Medumba and Français languages.
As Medumba is one of the low-resource languages of the world the translation system is unlikely to be developed by other methods than Rule-Based Machine Translation.
As the evidence of that we had only about 12 texts from Wikipedia incubator (links in Manifest) and quasi inexistent Medumba-Français aligned text sources as a source.


Generally project consisted of three main parts:
Generally project consisted of three main parts:


* Morphological analyzer for Guarani
* Morphological analyzer for Medumba
* Guarani-Spanish bilingual dictionary (bidix)
* Medumba-Français bilingual dictionary (bidix)
* Transfer rules
* Transfer rules


A detailed work plan for the project can be found [http://wiki.apertium.org/wiki/User:Anakuznetsova/Proposta here].
A detailed work plan for the project can be found [http://wiki.apertium.org/wiki/User:Mathematic-alpha/gsoc-proposal here].


==Morphological Analyzer ==
==Morphological Analyzer ==


⚫
We had to develop morphological analyzer almost from scratch.
We had to develop morphological analyzer almost from scratch. And the most challenging thing from the beginning was to find any properly organized lists of words or Guarani dictionaries. By the end of the Community Bonding period after 2 week of work we were able to analyze only 30% of words contained in wiki corpora. Coverage almost 90% by the end of Google Summer of Code program.
And the most challenging thing from the beginning was to find any properly organized lists of words or Medumba dictionaries.
By the end of the Community Bonding period after 2 week of work we were able to analyze only 20% of words contained in wiki corpora.
Coverage almost 90% by the end of Google Summer of Code program (if not of the bible text which is written in the old orthography).


Morphological analyzer is accompanied by twol file where 30 phonological rules are implemented.
<pre>
Twol is a kind of formalism used in Helsinki Finite State Technology (HFST).
GRN-Wiki
These rules resolve special cases of form transformations which could not be solved in morph analyzer, for example change of affixes after nasals etc.
To prevent our transducer from phonological errors we wrote tests for various cases (contained in "tests" directory).


⚫
In addition we have managed to do [https://github.com/ana-kuznetsova/apertium-BYV/blob/master/texts/eval1.txt syntactic tree annotation] for some of the texts from our wiki corpora, although it is on the initial stage and requires corrections.
dc ago 8 17:30:19 CEST 2018 109:13135 455256/508418 ~0.89543643222702579374
</pre>

<pre>
Bible

dc set 12 12:15:05 CEST 2018 109:13135 561107/623303 ~0.90021546503065122420
</pre>

⚫
All in all our morphological analyzer contains 12 638 lemmas:
* 4455 Nouns
* 2537 Verbs (divided in two groups by transitivity)
* 1668 Adjectives
* 457 Adverbs

This list does not include other lemma categories such as Proper Names (3052), different kind of pronouns, barbarisms (more frequently borrowed from Spanish), interjections, etc.

But we are able to analyze even more because some of the words have orthographical equivalents (as there are several traditions in written Guarani) or some forms found in corpora are just spelling errors.

Example of annotation made by morph analyzer:

<pre>
echo "Ou omba'apo hag̃ua" | lt-proc grn.automorf.bin

^Ou/Ou<v><iv><pres>/Ou<v><iv><p3><sg><pres>/Ou<v><iv><p3><pl><pres>$
^omba'apo/o<prn><pos><p3><sg>+mbaʼapo<n>$ ^hag̃ua/hag̃ua<post>$

Viene a trabajo.
</pre>

Although we did not do any disambiguation so far.

Morphological analyzer is accompanied by twol file where 30 phonological rules are implemented. Twol is a kind of formalism used in Helsinki Finite State Technology (HFST). These rules resolve special cases of form transformations which could not be solved in morph analyzer, for example change of affixes after nasals etc. To prevent our transducer from phonological errors we wrote tests for various cases (contained in "tests" directory).

⚫
In addition we have managed to do [https://github.com/ana-kuznetsova/apertium-grn/blob/master/texts/eval1.txt syntactic tree annotation] for some of the texts from our wiki corpora, although it is on the initial stage and requires corrections.




Line 155: Line 128:
Here are some sources that we used to construct morphological dictionary.
Here are some sources that we used to construct morphological dictionary.


# [http://descubrircorrientes.com.ar/2012/index.php/diccionario-guarani/1-guarani-espanol/1129-tai-c Discubrir Corrientes. La Enciclopedia Virtual Correntina] ;
# [http://descubrircorrientes.com.ar/2012/index.php/diccionario-Medumba/1-Medumba-espanol/1129-tai-c Discubrir Corrientes. La Enciclopedia Virtual Correntina] ;
# [https://github.com/LowResourceLanguages/hltdi-l3/tree/master/dicts Low Resource Language Dictionaries].
# [https://github.com/LowResourceLanguages/hltdi-l3/tree/master/dicts Low Resource Language Dictionaries].

'''Grammar reference'''
# Estigarribia, B. (2017). Guarani linguistics in the 21st century. 1st ed. BRILL, p.420.
# Krivoshein de Canese, N. and Decoud Larrosa, R. (1983). Gramatica de la lengua guarani. Asuncion: Nemity Krivoshein de Canese.


== Bilingual Dictionary ==
== Bilingual Dictionary ==
Bilingual dictionary (or bidix) was constructed from the lexicons used in morphological dictionaries mentioned above. Bidix entry looks the following way:
<pre>
<e><p><l>mbaʼapo<s n="n"/></l><r>tarea<s n="n"/><s n="f"/></r></p><par n="n_n"/></e>
</pre>

With this dictionary we can translate left to right from Guarani to Spanish.


The development of the bidix was halted for to the benefit of the morphological analyser for the Mə̀dʉ̂mbɑ̀.
<pre>
echo "buma mɛn" | apertium -d . byv-fra
#Venir #el tarea para
</pre>


== Transfer rules ==
== Transfer rules ==


We cannot reproduce correctly the structure of the Spanish sentence as well as word forms with right Spanish inflections as we still have few transfer rules. Transfer rules are developed only for nouns and some postpositions so far but we are planning to do it in the nearest future.
We cannot reproduce correctly the structure of the Français sentence as well as word forms with right Français inflections as we still have few transfer rules.
Transfer rules are developed only for nouns and some postpositions so far but we are planning to do it in the nearest future.


== Plans for future ==
== Plans for future ==
# Increase coverage as much as possible ;
# Increase coverage as much as possible ;
# Write transfer rules ;
# Write transfer rules ;
# Develop the interface ;


= Acknowledgements and impressions of GSoC program =
= Acknowledgements and impressions of GSoC program =


First of all I would like to express my gratitude to Francis Tyers, mentor, who has always been very helpful and supportive. In course of GSoC program he made me even more curious about linguistic technologies and Guarani language as well. He is highly qualified both in linguistics and technology. Hopefully this project will turn into larger academic work.
First of all I would like to express my total gratitude to Anastasia kuznetsova, mentor, who has always been very helpful and supportive (really).
In course of GSoC program she helped me when i was stuck and she was always available. She is really helpful. I also thank Jonathan Washington, mentor, who put in all his efforts to make this project a success.
Hopefully this project will motivate others and turn into larger academic work.

Revision as of 01:08, 26 August 2019

PROJECT: Adopt an unreleased language pair with a minimal user interface.

STATS and REPOs

Medumba progress
Date byv lexicon size Corpus size Coverage
12.06.2019 189 8674 ~0.37
20.06.2019 317 8674 ~0.39
05.07.2019 317 8674 ~0.56
10.07.2019 1186 6604 ~0.83
2019-07-21 1703 22627 ~65.21%
2019-07-25 1705 22637 ~69.19%
2019-07-25 17594 22624 ~77.77%
2019-08-05 27896 35035 ~79.62%
2019-08-21 12245 14823 ~90.01%
Medumba-French progress
Date byv-fra lexicon size Coverage WER, PER
12.06.2019 1592 ~0.10 Something
05.07.2019 1592 ~0.20 Something
10.07.2019 1592 ~0.35 Something


Monodix (Apertium-byv)
-My commits

Bidix (Apertium-byv-fra)
-My commits

Interface





Project description

A project on adoption of Medumba-Français language pair in Apertium had as its purpose a creation of machine translation system between Medumba and Français languages. As Medumba is one of the low-resource languages of the world the translation system is unlikely to be developed by other methods than Rule-Based Machine Translation. As the evidence of that we had only about 12 texts from Wikipedia incubator (links in Manifest) and quasi inexistent Medumba-Français aligned text sources as a source.

Generally project consisted of three main parts:

  • Morphological analyzer for Medumba
  • Medumba-Français bilingual dictionary (bidix)
  • Transfer rules

A detailed work plan for the project can be found here.

Morphological Analyzer

We had to develop morphological analyzer almost from scratch. And the most challenging thing from the beginning was to find any properly organized lists of words or Medumba dictionaries. By the end of the Community Bonding period after 2 week of work we were able to analyze only 20% of words contained in wiki corpora. Coverage almost 90% by the end of Google Summer of Code program (if not of the bible text which is written in the old orthography).

Morphological analyzer is accompanied by twol file where 30 phonological rules are implemented. Twol is a kind of formalism used in Helsinki Finite State Technology (HFST). These rules resolve special cases of form transformations which could not be solved in morph analyzer, for example change of affixes after nasals etc. To prevent our transducer from phonological errors we wrote tests for various cases (contained in "tests" directory).

In addition we have managed to do syntactic tree annotation for some of the texts from our wiki corpora, although it is on the initial stage and requires corrections.


Sources

Here are some sources that we used to construct morphological dictionary.

  1. Discubrir Corrientes. La Enciclopedia Virtual Correntina ;
  2. Low Resource Language Dictionaries.

Bilingual Dictionary

The development of the bidix was halted for to the benefit of the morphological analyser for the Mə̀dʉ̂mbɑ̀.

Transfer rules

We cannot reproduce correctly the structure of the Français sentence as well as word forms with right Français inflections as we still have few transfer rules. Transfer rules are developed only for nouns and some postpositions so far but we are planning to do it in the nearest future.

Plans for future

  1. Increase coverage as much as possible ;
  2. Write transfer rules ;
  3. Develop the interface ;

Acknowledgements and impressions of GSoC program

First of all I would like to express my total gratitude to Anastasia kuznetsova, mentor, who has always been very helpful and supportive (really). In course of GSoC program she helped me when i was stuck and she was always available. She is really helpful. I also thank Jonathan Washington, mentor, who put in all his efforts to make this project a success. Hopefully this project will motivate others and turn into larger academic work.

Retrieved from "https://wiki.apertium.org/w/index.php?title=User:Mathematic-alpha/gsoc-report&oldid=70371"