How morphologically complex languages are complex

• Nouns
  • Case (and pragmatic roles)
    wasi 'house', wasita, wasiq, wasipaq, wasiwan, wasipi, wasiman, wasimanta, wasipura, wasinpa, wasinta, wasinka, wasirayku, wasikama
  • Classes (gender, etc.)
    nyumba 'house', nyumba tatu 'three houses'; mtu 'person', watu watatu 'three persons'; mti 'tree', miti mitatu 'three trees'; kitanda 'bed', vitanda vitatu 'three beds'; neno 'word', maneno matatu 'three words'; ukuta 'wall', kuta mtatu 'three walls'
  • Possessor
    wasi 'house', wasiy, wasiyki, etc.
• Verbs
  • Subject agreement oí 'I heard', oíste 'you (sing.) heard', etc.
  • Object agreement ሰማሁ 'I heard', ሰማኋት 'I heard her', ሰማችኝ 'she heard me', ሰማንላቸው 'we heard for them'; ሰሙብን 'they heard "on" us' (179 combinations)
  • Tense/aspect/mood
  • Polarity
    ninaandika 'I write', siandiki 'I don't write', huandiki 'you don't write', niliandika 'I wrote' sikuandika 'I didn't write'
  • Other syntactic properties, especially relativization/nominalization
    nilikiandika 'I wrote it (book, etc.)', nilichokiandika 'which I wrote it'
  • Derivation: verbs from verbs, nouns, adjectives

Dealing with data sparsity

• Word tokens and word types
  Amharic: 4,846,999 tokens, 397,352 types
• Stemming and lemmatization
  ባይከፈትላቸውም 'even if it isn't opened for them'
  → -ከፈት- 'be opened (pres.)'
  → ተከፈተ 'be opened'
  → <kft> 'open'
• Morphological analysis
  ባይከፈትላቸውም →
  <kft>; [tam=past, vc=passive; sbj=[prs=3, num=sg, gen=ms]; iobj=[prs=3, num=pl, prp=l], +neg, cnj1='if', cnj2='also']
• Decoding and morphological generation
  <kft>; [tam=past, vc=pass; sbj=[prs=3, num=sg, gen=ms]; iobj=[prs=3, num=pl, prp=l], +neg, cnj1='if', cnj2='also']
  → ባይከፈትላቸውም

Redundant and non-redundant features

1. ህፃናት	መብቶቻቸው	እንዲጠበቁላቸውና	ጥቃት	እንዳይደርስባቸው
   infants	their-rights	so-that-they-are-kept-for-them-and	mistreatment	so-that-it-doesn't-arrive-on-them
   'so that infants' rights are maintained and they are not mistreated'

2. ልጆችዎን	በየጊዜው	ሐኪም	ዘንድ	ካሳዩ
   your-children-ACC	frequently	doctor	to	if-you-show
   'if you take your children to the doctor frequently'

   በትክክል	እያደጉ	መሆናቸውንና	ለተለያዩ	በሽታዎች	ያልተጋለጡ	መሆናቸውን
   properly	while-they-grow	their-being-ACC-and	to-various	diseases	that-they-haven't-been-exposed	their-being-ACC
   'that they are growing properly and that they haven't been exposed to various diseases'

   ሊያረጋግጡ	ይችላሉ
   that-you-make-sure	you-can
   'you can make sure'

From morphologically simple to morphologically complex languages

• Dealing with non-redundant features: phrasal/syntactic units
  ህፃናት ጥቃት እንዳይደርስባቸው
  infants mistreatment so that it doesn't arrive on them
• Dealing with redundant features: ??
  ህፃናት ጥቃት እንዳይደርስባቸው
  infants mistreatment so that it doesn't arrive on them

Enriching the source language (Avramidis and Koehn)

Another approach: generating features on the target end