DBMT

Czech-English Dependency-based Machine Translation – Čmejrek, Cuřín, and Havelka 03
PCEDT BLEU dtest/etest: 0.1974 0.1704

je to Magenta pipeline, jen generovani je rule-based (misto statistikeho tree-to-tree transducing a pak LM)

na českém prekladu Penn Treebanku

1. tokenizace a tagging [Hajic 98]
2. parsing do a_trees [Hajic 98, Charniak 99]
3. afun assigment [ZZ 02]
4. a_tree → t_tree [Bohmova 01]
5. func assigment C4.5 [ZZ 02]
6. slovnik pomoci GIZA++ [Och and Nay 02] one most probable translation, 1-2 as 1-1 multiword
   - generator
   
   ====== Generator ======
   dostane TGTS bez tfa, a co koreference
   
   == 1. determining contextual boundness ==
   povazuji v CZ od slovesa vlevo jako CB → definite article
   od slovesa vpravo nezapojene → indefinite article
   
   
   == 2. reordering of constituents ==
   podle CB se z ACT|PAT|ADDR vyberou Sb
   declarative sentence: CB adjuncts + Sb + V + direct/indirect Obj + UB adjuncts
   
   == 3. generation of verb forms ==
   pasivum/aktivum se dela podle Sb funktoru
   
   == 4. insertion of prepositions and articles ==
   preps: podle české a podle EN nounu
   
   articles: definite při postmodified NP, premodified by superlative or ordinal num
   article prevented: uncountable + proper nouns nebo predetermination by possessive a demonstrative pronouns
   
   == 5. morphology ==
   asi ne morpha
   hledaji v tabulce
   ^ word form ^ morphological tag ^ lemma ^
   kdyz nenajdou tak somple rules
   taky vokalizace pro indefinite article