1. Introduction
2. Translator's tools
3. Lexicon extraction
   1. Background
   2. Lexicon sources
   3. Showcases
   4. TermFactory demo
4. Dissemination

Translator's tools

• GF Cloud Simple Translation Tool (joint work with WP2)
• MOLTO translations integrated into Pootle (WP3)

Lexicon extraction

• Core technology for large-scale translation

Simple Translation Tool: translation editor / bilingual document authoring tool

Simple Translation Tool: translation editor / bilingual document authoring tool

MOLTO translations integrated to Pootle: traditional translation workflow

Demo installations at tfs.cc/pootle/ and cloud.grammaticalframework.org/translator

Demo video of Pootle

Pootle source code available as git repository at tfs.cc/git/pootle.git

Motivations

• Bigger grammars, new domains
• First experiments on free translation
• Especially WP5, Statistical and Robust Parsing

Scenarios

• Converting existing resources to GF lexicons
• Searching and managing terms in TermFactory

Types of lexicons

I. Monolingual

II. Multilingual

A. Uni-sense

B. Multi-sense

• Extracted from a monolingual resource
• Idiomatic, tailored for each language
• Used as a resource or for a monolingual application

DictEng.gf:

    a_priori_Adv = mkAdv "a priori";
    aardvark_N = mkN "aardvark" ;
    ab_initio_Adv = mkAdv "ab initio";
    aback_Adv = mkAdv "aback";
    abandon_V2 = mkV2 (mkV "abandon");

DictGer.gf:

    a_priori_Adv = mkAdv "a priori" ;
    aachener_N = reg2N  "Aachener" "Aachener" masculine ;
    aalfang_N = reg2N  "Aalfang" "Aalfänge" masculine ;
    aasvogel_N = reg2N  "Aasvogel" "Aasvögel" masculine ;
    abaenderbar_A = regA "abänderbar" ;

• Common abstract syntax, concrete syntaxes are translations of it
• Used for multilingual applications

DictEngGer.gf:

    abandon_V2 = dirV2 (mkV "verlassen" "verlasst" 
                            "verließ" "verließe" "verlassen" );
    abase_V2 = dirV2 (mkV "erniedrigen" "erniedrigt"
                          "erniedrigte" "erniedrigte" "erniedrigt");
    abasement_N = mkN "Erniedrigung";

• Uni-sense
  • Entries by lemmas, not word senses
• Multi-sense
  • Every distinct word sense gets an entry
  • Possible to combine with external word sense disambiguation tool
  • Example:

    LinkedDictGer:
        brother_08111676_N = reg2N "Bruder" "Brüder" masculine ;
        brother_08112052_N = reg2N "Bruder" "Brüder" masculine ;
        brother_08112265_N = reg2N "Kamerad" "Kameraden" masculine ;
    

  • Format: lemma_senseNumber_POS
    • Sense numbers independent of language

GF lexicon is a part of grammar: we need

• baseform: the uninflected form of the word (car, walk as opposed to cars, walking)
• inflection paradigm: cat : cats, child : children
• valency frame: word's arguments and their behaviour

Smart paradigms: infer the lower-level paradigm from the baseform

• Predictability does not decrease when the complexity of morphology grows (Détrez and Ranta, EACL 2012)

Annotated data

• TermFactory RDF
  • Includes necessary information for GF grammars
  • Conversion from TF RDF to GF included in the platform
• WordNet
• Morphological lexicons
  • Bulgarian: OpenOffice spellcheck (~60,000 words)
  • Finnish: Nykysuomen sanalista (~94,000 words)
  • French: Morphalou (~68,000 words)
  • Swedish: SALDO (~120,000 words)

Unannotated data

• Domain ontologies
  • ~18,000 concepts extracted for patent case
• Phrase tables
  • Phrase table: Entries of (source phrase, target phrase, probability) for a SMT system
  • Learned from parallel text data
  • ~4,000 words extracted for French, ~8,000 for German patent grammars
• Unannotated word lists
  • Bulgarian: Apertium dictionary (~18,000 words)
  • Urdu: Waseem Siddiqi’s English-Urdu dictionary (~27,000 words)
  • Various languages: Wiktionary

• TermFactory RDF format supports GF conversion
  • Different rules for each language, manual work
• Other formats: scripts written for each source

1. Linked lexicons from WordNet, used for translation with word sense disambiguation
2. TermFactory: term extraction from various sources, conversion to GF

• Multilingual multi-sense lexicon from WordNets (Hindi and Universal WordNet)
• External word sense disambiguation tool (IMS) in parsing
• First results: slight improvement for Hindi and German
• Results and documentation in Virk, 2013: Computational Linguistics Resources for Indo-Iranian Languages (PhD thesis, University of Gothenburg)

• TermFactory: Terminology management platform created in UHEL
• Collaborative work
• XHTML based editing
• RDF format with GF-specific predicates
• Conversion to GF lexicon

• Conference papers

  • Grégoire Détrez, Aarne Ranta, 2012. Smart Paradigms and the Predictability and Complexity of Inflectional Morphology. EACL 2012.
  • Ramona Enache, Cristina España-Bonet, Aarne Ranta, and Lluís Màrquez, 2012. A Hybrid System for Patent Translation. EAMT 2012.
  • Meritxell Gonzàlez, Maria Mateva, Ramona Enache, Cristina España-Bonet, and Lluís Màrquez, 2013. MT Techniques in a Retrieval System of Semantically Enriched Patents. Accepted for MTSummit 2013.

• Theses

  • Inari Listenmaa, 2012. Ontology-based lexicon management in a multilingual translation system – a survey of use cases. Master's thesis, University of Helsinki.
  • Shafqat Virk, 2013. Computational Linguistics Resources for Indo-Iranian Languages. PhD thesis, Chalmers University of Technology/University of Gothenburg.