Lookup Based Greedy Decoding for Machine Translation

1
Look-up Based Greedy Decoding for Machine
Translation

• Tony Zhang
• Steven Bills

2
Dictionary-based look-up

• IBM Models 1 and 2 generate translation
  probabilities statistically using parallel text
• We generate probabilities by looking up
  translations in freely available online lexical
  resources Wikipedia and Wiktionary
• These probabilities are often significantly more
  accurate than those generated by the IBM models

3
The translation process

• Our system translates from English to French
• We apply chunking, reordering and stemming in a
  pre-processing phase
• We generate an initial gloss of the pre-processed
  English sentence by using the most likely
  translation of each word
• We greedily apply mutations to the French
  translation until no superior mutations are left
• We apply post-processing to the resulting
  translation to remove duplicate words and perform
  contractions

4
Example mutations

• the last election
• Initial gloss le élection durer (the election
  to last)
• Retranslation mutation le élection dernier
  (correct)
• parliament is
• Initial gloss parlement est (needs an
  article)
• Insertion mutation le parlement est (correct
  article)
• the speaker wants to try
• Initial gloss le président veut à essayer
• Deletion mutation le président veut essayer

5
Processing

• Chunking
• Translate expressions as a whole rather than
  word-by-word when they have entries in our
  dictionary
• someone else ? quelquun dautre
• Reordering
• Swap nouns and adjectives
• useful organization ? organisation utile
• Conjugation
• Conjugate infinitives to agree with the subject
• the men walk ? les hommes marcher
  (infinitive)
• marcher ? marchent (3rd person plural
  conjugation)

6
Results

• Demonstrated the feasibility of building
  translation probabilities from online lexical
  resources
• Mutations and pre- and post-processing fixed many
  of the problems associated with word-by-word
  replacement
• Intelligibility and semantic closeness of
  translations drastically improved over IBM Models
  1 and 2