Ideas for Google Summer of Code

From Apertium
(Difference between revisions)
Jump to: navigation, search
(Improving Apertium language pairs mining Mediawiki Content Translation postediting information)
(Improving language pairs mining Mediawiki Content Translation postedits)
Line 221: Line 221:
 
* '''Difficulty''':<br><span style="background-color: #ffbdbd">1. Hard</span>
 
* '''Difficulty''':<br><span style="background-color: #ffbdbd">1. Hard</span>
 
* '''Required skills''':<br>Python, shell scripting, some statistics
 
* '''Required skills''':<br>Python, shell scripting, some statistics
* '''Description''':<br>Implement a toolkit that allows mining existing machine translation postediting data in [Mediawiki Content Translation https://www.mediawiki.org/wiki/Content_translation] to generate (as automatically as possible, and as complete as possible) monodix and bidix entries to improve the performance of an Apertium language pair. Data is available from Wikimedia content translation through an [API https://www.mediawiki.org/wiki/Content_translation/Published_translations#API]. This project is rather experimental and involves some research in addition to coding.
+
* '''Description''':<br>Implement a toolkit that allows mining existing machine translation postediting data in [Mediawiki Content Translation https://www.mediawiki.org/wiki/Content_translation] to generate (as automatically as possible, and as complete as possible) monodix and bidix entries to improve the performance of an Apertium language pair. Data is available from Wikimedia content translation through an [API https://www.mediawiki.org/wiki/Content_translation/Published_translations#API] or in the form of [Dumps https://dumps.wikimedia.org/other/contenttranslation/] available in JSON and TMX format. This project is rather experimental and involves some research in addition to coding.
 
* '''Rationale''':<br>Apertium is used to generate new Wikipedia content: machine-translated content is postedited (and perhaps adapted) before publishing. Postediting information may contain information that can be used to help improve the lexical components of an Apertium language pair.
 
* '''Rationale''':<br>Apertium is used to generate new Wikipedia content: machine-translated content is postedited (and perhaps adapted) before publishing. Postediting information may contain information that can be used to help improve the lexical components of an Apertium language pair.
 
* '''Mentors''':<br> [[User:Mlforcada|Mikel Forcada]], (more mentors to be added)
 
* '''Mentors''':<br> [[User:Mlforcada|Mikel Forcada]], (more mentors to be added)

Revision as of 17:27, 6 March 2017

Contents

This is the ideas page for Google Summer of Code, here you can find ideas on interesting projects that would make Apertium more useful for people and improve or expand our functionality. If you have an idea please add it below, if you think you could mentor someone in a particular area, add your name to "Interested mentors" using ~~~

The page is intended as an overview of the kind of projects we have in mind. If one of them particularly piques your interest, please come and discuss with us on #apertium on irc.freenode.net, mail the mailing list, or draw attention to yourself in some other way.

Note that, if you have an idea that isn't mentioned here, we would be very interested to hear about it.

Here are some more things you could look at:


List

Bring a released language pair up to state-of-the-art quality

  • Difficulty:
    2. Medium
  • Required skills:
    XML, a scripting language (Python, Perl), good knowledge of the language pair adopted.
  • Description:
    Take a released language pair, and drastically improve the performance both in terms of coverage, and in terms of translation quality. This will involve working with dictionaries, transfer rules, scripting, corpora. The objective is to make an Apertium language pair state-of-the-art, or close to state-of-the-art in terms of translation quality. This will involve improving coverage to 95-98% on a range of corpora and decreasing word error rate by 30-50%. For example if the current word error rate is 30%, then it should be reduced to 15-20%.
  • Rationale:
    Apertium has quite a broad coverage of language pairs, but few of these pairs offer state-of-the-art translation quality. We think broad is important, but deep coverage is important too.
  • Mentors:
    Francis Tyers, Mikel Forcada, Xavi Ivars
  • read more...

Adopt an unreleased language pair

  • Difficulty:
    3. Entry level
  • Required skills:
    XML, a scripting language (Python, Perl), good knowledge of the language pair adopted.
  • Description:
    Take on an orphaned unreleased language pair, and bring it up to release quality results. What this quality will be will depend on the language pair adopted, and will need to be discussed with the prospective mentor. This will involve writing linguistic data (including morphological rules and transfer rules — which are specified in a declarative language — and possibly Constraint Grammar rules if that is relevant)
  • Rationale:
    Apertium has a few pairs of languages (e.g. mt-he, ga-gd, ur-hi, pl-cs, sh-ru, etc...) that are orphaned, they don't have active maintainers. A lot of these pairs have a lot of work already put in, just need another few months to get them to release quality. See also Incubator
  • Mentors:
    Francis Tyers, Jimregan, Kevin Scannell, Trondtr, Unhammer, Darthxaher, Firespeaker, Hectoralos, Hrvoje Peradin, Jacob Nordfalk, Mikel Forcada, Vinit Ravishankar, Aida Sundetova, Xavi Ivars
  • read more...

Extend lttoolbox to have the power of HFST

  • Difficulty:
    1. Hard
  • Required skills:
    C++, XSLT, XML
  • Description:
    Extend lttoolbox (perhaps writing a preprocessor for it) so that it can be used to do the morphological transformations currently done with HFST. And yes, of course, writing something that translates the current HFST format to the new lttolbox format. Proof of concept: Come up with a new format that can express all of the features found in the Kazakh transducer; implement this format in Apertium; Implement the Kazakh transducer in this format and integrate it in the English--Kazakh pair.
  • Rationale:
    Some language pairs in Apertium use HFST where most language pairs use Apertium's own lttoolbox. This is due to the fact that writing morphologies for languages that have features such as the vowel harmony found in Turkic languages is very hard with the current format supported by lttoolbox. The mixture of HFST and lttoolbox makes it harder for people to develop some language pairs.
  • Mentors:
    Mikel Forcada, Tommi A Pirinen, User:Unhammer, Mikel Forcada, mentors wanted
  • read more


Discontiguous multiwords

  • Difficulty:
    2. Medium
  • Required skills:
    C++, Knowledge of FSTs
  • Description:
    The task will be to develop, or adapt a module to deal with these kind of contiguous multiword expressions, for example, taking 'liggja ekki fyrir' and reordering it as 'liggja# fyrir ekki'.
  • Rationale:
    In many languages, such as English, Norwegian and Icelandic, there are discontiguous multiwords, e.g. phrasal verbs, that we cannot easily support. For example 'liggja ekki fyrir' in Icelandic should be translated in English as 'to be not clear', but we cannot have 'liggja fyrir' as a traditional multiword because of the extra 'adverb', or it could even be a whole NP.
  • Mentors:
    Francis Tyers
  • read more...

Robust recursive transfer

  • Difficulty:
    1. Hard
  • Required skills:
    Python, XML, linguistics
  • Description:
    The purpose of this task would be to create a module to replace the apertium-transfer module(s) which will parse and allow transfer operations on an input.
  • Rationale:
    Currently we have a problem with very distantly related languages that have long-distance constituent reordering, because we can only do finite-state chunking.
  • Mentors:
    Francis Tyers, Sortiz, Mikel Forcada, Juan Antonio Pérez
  • read more...

Extend weighted transfer rules

  • Difficulty:
    1. Hard
  • Required skills:
    Python, C++, linguistics
  • Description:
    The purpose of this task is to extend weighted transfer rules to all transfer files and to allow conflicting rule patterns to be handled by combining (lexicalised) weights.
  • Rationale:
    Currently our transfer rules are applied longest-match left-to-right (LRLM). When two rule patterns conflict the first one is chosen. We have a prototype for selecting based on lexicalised weights, but it only applies to the first stage of transfer.
  • Mentors:
    Francis Tyers, Tommi Pirinen
  • read more...

Automatic blank handling

  • Difficulty:
    3. Hard
  • Required skills:
    C++, XML
  • Description:
    Let the C++ modules deal with formatting, treating some tags as word-bound.
  • Rationale:
    Our current handling of formatting/markup (html, odt, docx, latex) is brittle, requiring transfer rules to explicitly deal with blanks (e.g. markup), and some times inevitably outputting them in the wrong order. This project is implement automatic blank-handling, by treating some blanks as "glued to words" (e.g. italics, emphasis), and others as paragraph-level blanks.
  • Mentors:
    Unhammer, Tino Didriksen, Mikel Forcada
  • read more #1, read more #2

Improvements to the Apertium website

  • Difficulty:
    3. Entry level
  • Required skills:
    Python, HTML, JS
  • Description:
    Our web site is pretty cool already, but it's missing things like dictionary/synonym lookup, support for several variants of one language, reliability visualisation, (reliable) webpage translation, feedback, etc.
  • Rationale:
    https://apertium.org is what most people know us by, it should show off more of the things we are capable of :-)
  • Mentors:
    Jonathan
  • read more...

User-friendly lexical selection training

  • Difficulty:
    2. Medium
  • Required skills:
    Python, C++, shell scripting
  • Description:
    Make it so that training/inference of lexical selection rules is a more user-friendly process
  • Rationale:
    Our lexical selection module allows for inferring rules from corpora and word alignments, but the procedure is currently a bit messy, with various scripts involved that require lots of manual tweaking, and many third party tools to be installed. The goal of this task is to make the procedure as user-friendly as possible, so that ideally only a simple config file would be needed, and a driver script would take care of the rest.
  • Mentors:
    Unhammer, Francis Tyers, Mikel Forcada
  • read more...

Light alternative format for all XML files in an Apertium language pair

  • Difficulty:
    1. Hard
  • Required skills:
    Python, C++, shell scripting, XSLT, flex
  • Description:
    Make it possible to edit and develop language data using a format that is lighter than XML
  • Rationale:
    In most Apertium language pairs, monolingual dictionaries, bilingual dictionaries, post-generation rule files and structural transfer rule files are all written in XML. While XML is easy to process due to explicit tagging of every element, it is tedious to deal with, particularly when it comes to structural transfer rules. Apertium's precursor, interNOSTRUM, had lighter text based formats. The task involves: (a) designing and documenting an interNOSTRUM-style format for all of the XML language data files in a language pair; (b) writing converters to XML and from XML that are fully roundtrip-compliant: (c) designing a way to synchronize changes when both the XML and the non-XML format are used simultaneously in a specific language pair.
  • Mentors:
    Mikel Forcada, Juan Antonio Pérez, pair.
  • read more...

Eliminate dictionary trimming

  • Difficulty:
    0. Very Hard
  • Required skills:
    C++, Finite-State Transducers
  • Description:
    Eliminate the need for trimming the monolingual dictionaries, in order to preserve and take advantage of maximal source language analysis.
  • Rationale:
    Why we trim mentions several technical reasons for why trimming away monolingual information is currently needed. Unfortunately, this limitation means that a lot of useful contextual information is lost. It would be ideal if the source language could be fully analyzed independent of target language, with any untranslated part fed back into the source language generator.
  • Mentors:
    ? (Tommi Pirinen a.k.a. Flammie interested, but low on time if alone)
  • Work around everything in Why we trim

Bilingual dictionary enrichment via graph completion

  • Difficulty:
    0. Very hard
  • Required skills:
    shell scripting, python, XSLT, XML
  • Description:
    Generate new entries for existing or new bilingual dictionaries using graphic representations of bilingual correspondences as found in all existing dictionaries (note that this idea defines a rather open-ended task to be discussed in detail with mentors).
  • Rationale:
    Apertium bilingual dictionaries establish correspondences between lexical forms in a number of language pairs. Connections among them may be used to infer new entries for existing or new language pairs using graphs. The graphs may be directly generated from Apertium bidixes and exploiting using ideas that had already been proposed in Apertium or using existing RDF representations of parts of their content, which may benefit from the information coming from being linked to other resources.
  • Mentors:
    Mikel Forcada, Francis Tyers, Jorge Gracia
  • read more... read even more...


Transfer rule induction from comparable parsed corpora

  • Difficulty:
    1. Hard
  • Required skills:
    shell scripting, python, XSLT, XML
  • Description:
    A system to infer transfer rules from comparable corpora that have both been deeply parsed (with e.g. CG)
  • Rationale:
    Many languages have good CG's and fairly large monolingual corpora, but little parallel material. Given a small bidix, fairly large monolingual corpora and good analysers/CG's, we should be able to parse both corpora, translate lemmas and look for similar sentences, turning the differences in their parses into transfer rules.
  • Mentors:
    User:Unhammer
  • read more...

Back-off morphological generation with RNNs

  • Difficulty:
    1. Hard
  • Required skills:
    c++
  • Description:
    Write a pared-down RNN library to do morphological generation, add it as a backoff to lt-proc in case of a generation fail.
  • Rationale:
    One of the most frustrating things when developing a new language pair is that you have to get the tags just right in order to be able to generate, if they come in the wrong order, or if you're missing a tag or if you have too many tags, then you don't get any surface form generated. For many languages it is possible to train RNNs to a high-level of accuracy.
  • Mentors:
    User:Francis Tyers User:Mlforcada
  • read more...

UD and Apertium integration

  • Difficulty:
    3. Entry level
  • Required skills:
    python, javascript, HTML, (C++)
  • Description:
    Create a range of tools for making Apertium compatible with Universal Dependencies
  • Rationale:
    Universal dependencies is a fast growing project aimed at creating a unified annotation scheme for treebanks. This includes both part-of-speech and morphological features. Their annotated corpora could be extremely useful for Apertium for training models for translation. In addition, Apertium's rule-based morphological descriptions could be useful for software that relies on Universal dependencies.
  • Mentors:
    User:Francis Tyers User:Firespeaker
  • read more...

Add weights to lttoolbox

  • Difficulty:
    1. Hard
  • Required skills:
    c++
  • Description:
    Add support for weighted transducers to lttoolbox
  • Rationale:
    This will either involve implementing it from scratch or adding OpenFST as a backend. We would like to be able to use it both in the bilingual dictionaries, and in the morphological analysers, to be able to order analyses/translations by their probability/weight instead of by the random topological order.
  • Mentors:
    User:Francis Tyers User:Unhammer
  • read more...


Improving language pairs mining Mediawiki Content Translation postedits

  • Difficulty:
    1. Hard
  • Required skills:
    Python, shell scripting, some statistics
  • Description:
    Implement a toolkit that allows mining existing machine translation postediting data in [Mediawiki Content Translation https://www.mediawiki.org/wiki/Content_translation] to generate (as automatically as possible, and as complete as possible) monodix and bidix entries to improve the performance of an Apertium language pair. Data is available from Wikimedia content translation through an [API https://www.mediawiki.org/wiki/Content_translation/Published_translations#API] or in the form of [Dumps https://dumps.wikimedia.org/other/contenttranslation/] available in JSON and TMX format. This project is rather experimental and involves some research in addition to coding.
  • Rationale:
    Apertium is used to generate new Wikipedia content: machine-translated content is postedited (and perhaps adapted) before publishing. Postediting information may contain information that can be used to help improve the lexical components of an Apertium language pair.
  • Mentors:
    Mikel Forcada, (more mentors to be added)
  • (more soon)

Shallow-function labeller

  • Difficulty:
    2. Medium
  • Required skills:
    Python, shell scripting
  • Description:
    Implement a prototype shallow syntactic function labeller for Apertium
  • Rationale:
    In many pairs it is useful to know in addition to the morphological tags of a word, syntactic function tags in order to make an adequate translation. For instance, you might want to know in an ergative language if an absolutive is subject or object while translating. A shallow function labeller takes an annotated corpus and produces a model which can annotate new text.
  • Mentors:
    Unhammer, Francis Tyers, Mikel Forcada
  • read more...
Personal tools