Development of a German-English Translator

1
Development of a German-English Translator

• Felix Zhang
• TJHSST Computer Systems Research Lab 2007-2008
• Period 5

2
Summary of previous quarters

• Developed a functional translator for simple
  German sentences to simple English sentences
• All rule-based, no statistical methods, yet
• Very specifically geared towards German
  language dependent

3
Scope for 4th quarter

• Statistical methods
• Part-of-speech tagging
• Morphological analysis
• Information available in TIGER corpus
• Accuracy testing
• Reliability of statistical methods
• Vs. Rule-based

4
Finishing Rule-Based Translation

• Convert translation into user-readable format
• Punctuation and capitalization
• fzhang_at_ltsp1 /research python proj.py
• Part of speech tags 'den', 'art', 'kurzen',
  'adj', 'Mann', 'nou', 'machen', 'ver',
  'die', 'art', 'kleinen', 'adj', 'Kinder',
  'nou'
• Morphological analysis 'kurzen', 'adj',
  'akk', 'mas', 'dat', 'pl', 'Mann',
  'nou', 'akk', 'mas', 'dat', 'pl',
  'machen', 'ver', '1', 'pl', '3', 'pl',
  'pres', 'kleinen', 'adj', 'nom', 'pl',
  'akk', 'pl', 'Kinder', 'nou', 'nom',
  'pl', 'akk', 'pl'
• Disambiguated after noun-verb agreement
  'kurzen', 'adj', 'akk', 'mas', 'dat',
  'pl', 'Mann', 'nou', 'akk', 'mas',
  'dat', 'pl', 'machen', 'ver', '3',
  'pl', 'pres', 'kleinen', 'adj', 'nom',
  'pl', 'akk', 'pl', 'Kinder', 'nou',
  'nom', 'pl'
• Lemmatized 'kurzen', 'kurz', 'Mann',
  'Mann', 'Man', 'machen', 'machen',
  'kleinen', 'klein', 'Kinder', 'Kind'
• Root translated 'den', 'the', 'kurzen',
  'short', 'Mann', 'man', 'machen', 'make',
  'die', 'the', 'kleinen', 'small', 'Kinder',
  'child'
• NP Chunked English 'the', 'art', 'short',
  'adj', 'man', 'nou', 'akk', 'mas', 'dat',
  'pl', 'make', 'ver', '3', 'pl', 'pres',
  'the', 'art', 'small', 'adj', 'child',
  'nou', 'nom', 'pl'
• Assigned an element type
• 'the', 'art', 'short', 'adj', 'man',
  'nou', 'akk', 'mas', 'dat', 'pl', 'dobj',
  'make', 'ver', '3', 'pl', 'pres', 'mverb',
  'the', 'art', 'small', 'adj', 'child',
  'nou', 'nom', 'pl', 'sub'
• Assigned priority
• '5', 'the', 'art', 'short', 'adj', 'man',
  'nou', 'akk', 'mas', 'dat', 'pl', 'dobj',
  '2', 'make', 'ver', '3', 'pl', 'pres',
  'mverb', '1', 'the', 'art', 'small', 'adj',
  'child', 'nou', 'nom', 'pl', 'sub'
• Rearranged to English structure
• '1', 'the', 'art', 'small', 'adj',
  'child', 'nou', 'nom', 'pl', 'sub', '2',
  'make', 'ver', '3', 'pl', 'pres', 'mverb',
  '5', 'the', 'art', 'short', 'adj', 'man',
  'nou', 'akk', 'mas', 'dat', 'pl', 'dobj'
• Inflected
• The small childs make the short man.

5
Statistical Methods

• Trivial methods
• Part of speech tagging, morphological analysis
• Tags based on most frequently occurring tag with
  the word in corpus
• Theoretically, should still achieve reasonable
  levels of accuracy

6
Testing

• Check all 746,660 words in TIGER corpus
• Problem Running time
• Match Actual part of speech / linguistic
  properties of the word in the corpus part of
  speech, etc. that program assigns based on
  maximum likelihood
• Predictions from research 90 accuracy for part
  of speech tags

7
Accuracy Part of speech

• Reaches about 90 percent accuracy, as predicted
• Match NE NE Nato
• 656882
• Match ADV ADV allein
• 656883
• Match VAFIN VAFIN sein
• 656884
• Match PIAT PIAT kein
• 656885
• Match NN NN Zukunftskonzept
• 656886
• Match APPR APPR für
• 656887
• Match ART ART der
• 656888
• Match NN NN Sicherheit
• 656889
• Match APPR APPR in
• 656890

8
Accuracy Morphological Analysis

• Lower accuracy Properties are more
  context-dependent
• Match -- -- allein
• 550403
• Match 3.Sg.Pres.Ind 3.Sg.Pres.Ind ist
• 550404
• Match Nom.Sg.Neut Nom.Sg.Neut Zukunftskonzept
• 550405
• Match -- -- für
• 550406
• Match Acc.Sg.Fem Acc.Sg.Fem die
• 550407
• Match Acc.Sg.Fem Acc.Sg.Fem Sicherheit
• 550408
• Match -- -- in
• 550409
• Match Dat.Sg.Neut Dat.Sg.Neut Europa
• 550410
• Total matches 550410
• Total words 746660

9
Problems

• No good way to compare effectiveness of
  rule-based vs. statistical translations
• Accuracy not high enough
• 10 error too large in a 700,000 word corpus
• 27 even worse
• Slow run times

10
Future Research

• Combine statistical and rule-based methods
  together in hybrid program
• Find way to compare accuracy of two techniques
• More advanced (and more accurate) statistical
  methods Context-based