= Machine Translation =

[[Image(/trac/research/raw-attachment/wiki/en/MachineTranslation/error.png)]]

== Topics ==
 * Statistical machine translation
 * Extension of translation memories
 * Domain-specic machine translation
 * Machine translation between close languages
 * Sub-word level machine translation

== Statistical machine translation ==

[[Image(/trac/research/raw-attachment/wiki/en/MachineTranslation/trans.png)]]

== Improving statistical machine translation ==
 * free state-of-the-art tools available (SRILM, Moses)
 * baseline SMT available for everyone
 * languages with high number of wordforms need special treatment
 * '''dělám, děláš, dělal, dělajícímu, dělaje, děláním, ...'''
   '''magas, magasabb, legmagasabb, legeslegmagasabb, ...'''
 * language models can be enriched with linguistic knowledge

[[Image(/trac/research/raw-attachment/wiki/en/MachineTranslation/kings.png)]]


== Word alignment matrix - from words to phrases ==

[[Image(/trac/research/raw-attachment/wiki/en/MachineTranslation/word_matrix1.png)]]


[[Image(/trac/research/raw-attachment/wiki/en/MachineTranslation/word_matrix2.png)]]


[[Image(/trac/research/raw-attachment/wiki/en/MachineTranslation/word_matrix3.png)]]


== Domain-specic machine translation ==
 * straightforward way of increasing quality of MT
 * domain-specic corpora can be downloaded on demand
 * separate models for each domain: ''' sports, cooking, gardening'''
 * one sense per domain: '''bat'''

[[Image(/trac/research/raw-attachment/wiki/en/MachineTranslation/bat.png)]]

 * translations of 
   * product details, product descriptions in e-shops,
   * manuals, warranty certicates,
   * user interface localizations, ...


== Machine translation between close languages ==
 * West and South Slavic languages: Czech, Slovak, Polish, Serbian, Croatian, Slovene
 * MT mainly on word level, structure is very similar
 * dierences can be described systematically by rules:
   '''hraje na klavíri''' <-> '''hraje na klavír'''
 * billion-word corpora available for these languages
 * dictionaries can be generated semi-automatically

 * -> searching for duplicates in close languages (reprinted news)

== MT quality, European languages ==

[[Image(/trac/research/raw-attachment/wiki/en/MachineTranslation/lang_matrix.png)]]


== Sub-word level machine translation ==
 * SMT principle applied on character level
 * translation on subword level (English -> Czech)
[[Image(/trac/research/raw-attachment/wiki/en/MachineTranslation/trans1.png)]]
 * translation across levels
[[Image(/trac/research/raw-attachment/wiki/en/MachineTranslation/trans2.png)]]

 * -> translation of out-of-dictionary words


== Conclusions ==
 * generating new segments for translation memories
 * domain-specic translation
 * translation between close languages
 * sub-word level translation