Difference between revisions of "Ideas for Google Summer of Code"
(→List) |
(→List) |
||
Line 71: | Line 71: | ||
|- |
|- |
||
| '''Inferring transfer rules with active learning''' || 2. Hard || Re-working apertium-transfer-training-tools to get more general rules. The right level of generalisation can be achieved by asking non-expert users to validate examples. || Apertium-transfer-training-tools generates shallow-transfer rules from aligned parallel corpora. The corpus-extracted rules provide a translation quality comparable with the expert-written ones. However, human-written rules are more general and easier to mantain. The purpose of this task is to reduce the number of inferred rules while keeping translation quality. Gaps in the information elicited from the parallel corpus are filled with knowledge from non-expert users to achieve the right degree of generalisation. || C++, general knowledge of GIZA++, XML. || [[User:Vitaka|Víctor M. Sánchez-Cartagena]] |
| '''Inferring transfer rules with active learning''' || 2. Hard || Re-working apertium-transfer-training-tools to get more general rules. The right level of generalisation can be achieved by asking non-expert users to validate examples. || Apertium-transfer-training-tools generates shallow-transfer rules from aligned parallel corpora. The corpus-extracted rules provide a translation quality comparable with the expert-written ones. However, human-written rules are more general and easier to mantain. The purpose of this task is to reduce the number of inferred rules while keeping translation quality. Gaps in the information elicited from the parallel corpus are filled with knowledge from non-expert users to achieve the right degree of generalisation. || C++, general knowledge of GIZA++, XML. || [[User:Vitaka|Víctor M. Sánchez-Cartagena]] |
||
|- |
|||
| '''Active learning to choose among paradigms which share superficial forms''' || 2. Hard || Developing an active learning algorithm to choose the most appropriate paradigm (in a monolingual dictionary) for a new word. Previous research work allows us to reduce the problem to a set of paradigms which share superficial forms but not lexical forms. || Current research on the addition of entries to a monolingual dictionary by non-expert users has partially solved the problem of choosing the best paradigm for a new word. However, when different paradigms share all the lexical forms the problem becomes harder. In such case, it is necessary to ask the users to validate sentences in which the word to be added acts as different lexical forms. Such sentences may be extracted from a monolingual corpus. || XML, Java, Web technologies. || [[User:Vitaka|Víctor M. Sánchez-Cartagena]], Miquel Esplà-Gomis |
|||
|} |
|} |
||
Revision as of 19:40, 9 March 2011
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:
- the open bugs page on Bugzilla
- that link is broken.. this page seems to give open bugs though
- pages in the development category
- resources that could be converted or expanded in the incubator. Consider doing or improving a language pair (see incubator and nursery for pairs that need work)
- Unhammer's wishlist
- Top tips for GSOC applications
List
Note: The table below is sortable by column. Click on the little squares below or next to the headers.
Task | Difficulty | Description | Rationale | Requirements | Interested mentor |
---|---|---|---|---|---|
Adopt a language pair read more... |
4. Entry level | Take on an orphaned 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) | Apertium has a few pairs of languages (e.g. sh-mk, en-af, ga-gd, ur-hi, id-ms, 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 | XML, a scripting language (Python, Perl), good knowledge of the language pair adopted. | Francis Tyers, Jimregan, Kevin Scannell, Trondtr, Niunniminuni, Unhammer |
Discontiguous multiwords read more... | 3. Medium | 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'. | 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. | C++, Knowledge of FSTs | Francis Tyers, Jimregan |
Flag diacritics in lttoolbox read more... | 2. Hard | Adapt lttoolbox to elegantly use flag diacritics. Flag diacritics are a way of avoiding transducer size blow-up by discarding impossible paths at runtime as opposed to compile time. | This will involve designing some changes to our XML dictionary format (see lttoolbox, and implementing the associated changes in the FST compiling processing code. The reason behind this is that many languages have prefix inflection, and we cannot currently deal with this without either making paradigms useless, or overanalysing (e.g. returning analyses where none exist). Flag diacritics (or constraints) would allow us to restrict overanalysis without blowing up the size of our dictionaries. | C++ or Java, XML, Knowledge of FSTs | Francis Tyers (not Java), Jacob Nordfalk |
Linguistically-driven bilingual-phrase filtering for inferring transfer rules | 3. Medium | Re-working apertium-transfer-training-tools to filter the set of bilingual phrases automatically obtained from a word-aligned sentence pair by using linguistic criteria. | Apertium-transfer-training-tools is a cool piece of software that generates shallow-transfer rules from aligned parallel corpora. It could greatly speed up the creation of new language pairs by generating rules that would otherwise have to be written by human linguists | C++, general knowledge of GIZA++, Perl considered a plus. | Jimregan |
Context-dependent lexicalised categories for inferring transfer rules | 2. Hard | Re-working apertium-transfer-training-tools to use context-dependent lexicalised categories in the inference of shallow-transfer rules. | Apertium-transfer-training-tools generates shallow-transfer rules from aligned parallel corpora. It uses an small set of lexicalised categories, categories that are usually involved in lexical changes, such as prepositions, pronouns or auxiliary verbs. Lexicalised categories differentiate from the rest of categories because their lemmas are taken into account in the generation of rules. | C++, general knowledge of GIZA++, XML. | Jimregan |
Improve integration of lttoolbox in libvoikko read more... |
3. Medium | Dictionaries from lttoolbox can now be used for spellchecking directly with libvoikko (see Spell checking). The idea of this project is to improve the integration. Fix bugs, look at ways of codifying "standard"/"sub-standard" forms in our dictionaries. | Spell checkers can be useful, for languages other than English moreso. They are one of the "must have" items of language technology. If we can re-use Apertium data for this purpose it will help both the project (by making creating new language pairs more rewarding) and the language communities (by making more useful software). | XML, C++. | Francis Tyers |
Complex multiwords | 2. Hard | Write a bidirectional module for specifying complex multiword units, for example dirección general and zračna luka. See Multiwords for more information. | Although in the Romance languages it is not a big problem, as soon as you start to get to languages with cases (e.g. Serbo-Croatian, Slovenian, German, etc.) the problem comes that you can't define a multiword of adj nom because the adjective has a lot of inflection. |
Java or C++, XML | Jimregan |
Detect 'hidden' unknown words read more... | 3. Medium | The part-of-speech tagger of Apertium can be modified to work out the likelihood of each 'tag' in a certain context, this can be used to detect missing entries in the dictionary. | Apertium dictionaries may have incomplete entries, that is, surface forms (lexical units as they appear in running texts) for which the dictionary does not provided all the possible lexical forms (consisting of lemma, part-of-speech and morphological inflection information). As those surface form for which there is at least one lexical form cannot be considered unknown words, it is difficult to know whether all lexical forms for a given surface form have been included in the dictionaries or not. This feature will detect 'possible' missing lexical forms for those surface forms in the dictionaries. | C++ if you plan to modify the part-of-speech tagger; whatever if rewriting it from scratch. | Felipe Sánchez-Martínez |
Geriaoueg vocabulary assistant |
4. Entry level | Extend Geriaoueg so that it works more reliably with broken HTML and with any given language pair. | Geriaoueg is a program that provides "popup" vocabulary assistance, something like BBC Vocab or Lingro. Currently it only works with Breton--French, Welsh--English and Spanish--Breton. This task would be to develop it to work with any language in our SVN and fix problems with processing and displaying non-standard HTML. | PHP, C++, XML | Francis Tyers |
Corpus-assisted dictionary expansion | 4. Entry level | Semi-automatic bilingual word equivalence retrieval from a bitext and a monolingual word list. | Improve an existing Python script to retrieve the best translations (suggestions) of a word (typically an unknown word) given a particular parallel text corpus. Perhaps combine the result with automatic paradigm guessing (also suggestions) to improve the productivity of the lexical work for most contributors | Python, C/C++, AWK, Bash, perhaps web interface in PHP, Python, Ruby on Rails | Sortiz, Jimregan |
Improvements to the Advanced Web Interface | 4. Entry level | Improve the pre and post-editing Apertium translation web environment where the user has a range of tools available in order to modify the text before sending it to Apertium and after getting its translation. | Improve the existing interface on some of the following points: improving the system to extract information from the edition's logs, improving the cross-browser compatibility, increase the modularity of the code, increase the usability of the interface, integrating other useful tools like After The Deadline... | PHP and AJAX | Villarejo, Jimregan |
Improvements to target-language tagger training read more... | 2. Hard | Modify apertium-tagger-training-tools so that it can deals with n-stage transfer rules when segmenting the input source-language text, and applies a k-best viterbi pruning approach that does not require to compute the a-priori likelihood of every disambiguation path before pruning. | apertium-tagger-training-tools is a program for doing target-language tagger training, meaning it improves POS tagging performance specifically for the translation task, achieving a result for unsupervised training comparable with supervised training. This task would also require switching the default perl-based language model to either IRSTLM or RandLM (or both!). | C++, XML, XSLT | Felipe Sánchez-Martínez |
Accent and diacritic restoration read more... |
3. Medium | Create an optional module to restore diacritics and accents on input text, and integrate it into the Apertium pipeline. | Many languages use diacritics and accents in normal writing, and Apertium is designed to use these, however in some places, especially for example. instant messaging, irc, searching in the web etc. these are often not used or untyped. This causes problems as for the engine, traduccion is not the same as traducción. | C, C++, XML, familiarity with linguistic issues | Kevin Scannell, Trondtr |
VM for the transfer module read more... | 3. Medium | VM for the current transfer architecture of Apertium and for the future transfers, pure C++ | Define an instruction set for a virtual machine that processes transfer code, then implement a prototype in Python, then porting to C++. The rationale behind this is that XML tree-walking is quite slow and CPU intensive. In modern (3 or more stage) pairs, transfer takes up most of the CPU. There are other options, like Bytecode for transfer, but we would like something that does not require external libraries and is adapted specifically for Apertium. | Python, C/C++, XML, XSLT, code optimisation, JIT techniques, etc. | Sortiz |
Hybrid MT | 2. Hard | Building Apertium-Marclator rule-based/corpus-based hybrids | Both the rule-based machine translation system Apertium and the corpus-based machine translation system Marclator do some kind of chunking of the input as well as use a relatively straightforward left-to-right machine translation strategy. This has been explored before but there are other ways to organize hybridization which should be explored (the mentor is happy to discuss). Hybridization may make it easier to adapt Apertium to a particular corpus by using chunk pairs derived from it. | Knowledge of Java, C++, and scripting languages, and appreciation for research-like coding projects | Mlforcada, Jimregan |
Regular expressions in lt-tmxproc | 2. Hard | Adding regex support to lt-tmxproc would maximise the amount of translations we can get from an available TMX. | lt-tmxproc already includes some limited support for making translation units in a TMX file into something of a template, but only for digits. Gintrowicz and Jassem describe an interesting idea, using user-definable regular expressions, to turn items such as dates into templates. lttoolbox already has support for a subset of regular expressions; add a mechanism to allow the user to make use of this, and to include these regular expressions in processing. | Knowledge of C++ | Jimregan |
Quality control framework | 3. Medium | Write a unified testing framework for released language pairs in Apertium. The system should be able to track both regressions with respect to previous versions, and quality checks with respect to previous quality evaluations. | We are gradually improving our quality control, with (semi-)automated tests, but these are done on the Wiki on an ad-hoc basis. Having a unified testing framework would allow us to be able to more easily track quality improvements over all language pairs, and more easily deal with regressions. | PHP or Python | Francis Tyers |
Rule-based finite-state disambiguation | 2. Hard | Currently Apertium only has a bigram/trigram part-of-speech tagger. The objective of this task would be to implement a disambiguation framework for Apertium that can be expressed as a finite-state transducer. It might be a good idea to express this as constraint rules, in a novel XML-based file format. | For some languages, bigram/trigram POS disambiguation really doesn't work, especially when you want to disambiguate morphology (e.g. number, case) along with part-of-speech. So far we've been using constraint grammar for some of these languages. But although Constraint Grammar is great and powerful, it is also pretty slow. It would be a good idea to look at LanguageTool, and IceParser to get ideas on how this might be accomplished. | XML, C++ | Francis Tyers |
Dependency-based reordering | 2. Hard | Currently we have a problem with very distantly related languages that have long-distance constituent reordering. Some languages have dependency parsers which create graphs of words. | The purpose of this task would be to create a module that comes before or after apertium-transfer which reorders the graph. | XML, C++ | Francis Tyers |
Dictionary induction from wikis | 3. Medium | Extract dictionaries from linguistic wikis | Wiki dictionaries and encyclopedias (e.g. omegawiki, wiktionary, wikipedia, dbpedia) contain information (e.g. bilingual equivalences, morphological features, conjugations) that could be exploited to speed up the development of dictionaries for Apertium. This task aims at automatically building dictionaries by extracting different pieces of information from wiki structures such as interlingual links, infoboxes and/or from dbpedia RDF datasets. | MySQL and SPARQL, RDF, mediawiki syntax, perl, maybe C++ or Java | Atoral, Jimregan |
Advanced Wikipedia translation | 3. Medium | We have a tool for handling Wikipedia formatting; however, much of that formatting conveys meaningful information that could be 'translated'. | Infoboxes and other templates could be handled in a manner similarly to DBPedia's mappings wiki; categories can be inferred via DBPedia; interwiki links can be taken from the source page, but updated to add a link to the source; wiki links can be replaced via DBPedia with links appropriate to the target wiki (or discarded). This would be relatively easy to achieve, if based on the DBPedia extraction framework. | Java/Scala, RDF, SparQL, DBPedia | Jimregan |
Inferring transfer rules with active learning | 2. Hard | Re-working apertium-transfer-training-tools to get more general rules. The right level of generalisation can be achieved by asking non-expert users to validate examples. | Apertium-transfer-training-tools generates shallow-transfer rules from aligned parallel corpora. The corpus-extracted rules provide a translation quality comparable with the expert-written ones. However, human-written rules are more general and easier to mantain. The purpose of this task is to reduce the number of inferred rules while keeping translation quality. Gaps in the information elicited from the parallel corpus are filled with knowledge from non-expert users to achieve the right degree of generalisation. | C++, general knowledge of GIZA++, XML. | Víctor M. Sánchez-Cartagena |
Active learning to choose among paradigms which share superficial forms | 2. Hard | Developing an active learning algorithm to choose the most appropriate paradigm (in a monolingual dictionary) for a new word. Previous research work allows us to reduce the problem to a set of paradigms which share superficial forms but not lexical forms. | Current research on the addition of entries to a monolingual dictionary by non-expert users has partially solved the problem of choosing the best paradigm for a new word. However, when different paradigms share all the lexical forms the problem becomes harder. In such case, it is necessary to ask the users to validate sentences in which the word to be added acts as different lexical forms. Such sentences may be extracted from a monolingual corpus. | XML, Java, Web technologies. | Víctor M. Sánchez-Cartagena, Miquel Esplà-Gomis |
Notes
Further reading
- Transfer rule learning
- Sánchez-Martínez, F. and Forcada, M.L. (2009) "Inferring shallow-transfer machine translation rules from small parallel corpora" In Journal of Artificial Intelligence Research. volume 34, p. 605-635.
- Target-language driven part-of-speech tagger training
- Sánchez-Martínez, F.; Pérez-Ortiz, J.A. and Forcada, M.L. (2008) "Using target-language information to train part-of-speech taggers for machine translation". In Machine Translation, volume 22, numbers 1-2, p. 29-66.
- Regular expressions in lt-tmxproc
- Gintrowicz, J. and Jassem, K. (2007) "Using Regular Expressions in Translation Memories". In Proceedings of the International Multiconference on Computer Science and Information Technology, p. 87--92.