Extension of MOLTO's MT approach to deal with non-restricted language

Exploration and development of translation engines especialised for patents

Patents MT&IR prototype, translation tools

By GF's robustness

• Robust parsing

By hybrid SMT-GF techniques

• GF grammar for patents
  
• SMT trained on patents
  
• On top SMT decoding
  

Standard SMT system with

• Corpus: MAREC corpus (biomedical domain, EN-DE-FR)
  • In-domain tokeniser
• Language model: 5-gram interpolated Kneser-Ney discounting, SRILM Toolkit
• Alignments: GIZA++ Toolkit
• Translation model: Moses package
• Weights optimization: MERT against BLEU
• Decoder: Moses

Built on the Resource Grammar

Extensions

• 4 categories for chunk parsing and 3 for agreement
• 8 functions that deal with agreements and parsing chunks
• 9 extensions to the resource grammar dealing with general-purpose constructions
• 13 constructions to deal with claim-specific data
• 8 structural words typical for patent domain case

Core lexicon

• Most frequent words, SMT translation, one-to-one correspondance

Static lexicon

• SMT lexical/translation tables, one-to-many correspondance

Dictionary of compounds (German)

• SMT tables, dictionary, many-to-one correspondance

Safe building

• Unknown lexical items are looked up in the monolingual dictionaries
• Two levels of confidence: unambiguous vs. ambiguos

Unsafe building

• Uses smart paradigms to build items absent in the dictionary
• Three levels of confidence: unambiguous, ambiguos and unsafe

SMT & GF integration led by GF

• GF grammar with SMT built lexicon and disambiguation by frequency counts
• Robust parsing with statistical models for searching the space and for disambiguation

SMT & GF integration led by SMT

• Additional SMT decoding on top of GF and SMT to choose the best translation options

Hard Integration

• GF phrases are forced to appear
• SMT complements
• Top SMT reorders

Soft Integration

• GF and SMT phrases interact
• Top SMT reorders and chooses the best option
• LM plays an important role in choosing

Characteristics and options

• Static vs. dynamic lexicon (two types)
• Base vs. extended lexicons
• Single vs. multiple GF translations available
• Hard vs. soft integration
• Integration at decoding time vs. tuning

Number of phrases from every system choosen at the end

Automatic evaluation, En2Fr

Automatic evaluation, En2De

Manual evaluation, En2Fr & EnDe

Manual evaluation, En2Fr & EnDe

Some improvements to do after manual evaluation. Still,

• Exact matches with the reference translation got sometimes ranked as 1 or 2
• Even the published human translations are evaluated as unpublishable
• In some cases very low BLEU scoring sentences got a 4

Perl script that runs the translator

 molto-server:hybrid$ perl H1PTrad.pl

 Usage: perl H1PTrad.pl -v # -m [runtime|unsafe|demo] <input> [src2trg] 
 -v: verbosity [0,1,2]
 -m: mode [runtime|unsafe|demo]
 input: file to translate
 src2trg: language pair

 Ex: 
 perl H1PTrad.pl -v 1 -m demo /Users/systems/input/patsA61P.test.en en2fr

Prototype available in EN, DE and FR

Architecture

• SMT-based pipeline for automatic translation of annotated documents

• Multilingual document retrieval (query flagship)

• GF-based querying subsystem for automatic translation of CNL queries to SPARQL (query flagship)

• User Interface (query flagship)