Concept

Information access

• The freedom or ability to identify, obtain and make use of database or information effectively
• The most common current use of machine translation
• Improvements in machine translation can help reduce the digital divide in information access: that fact that much more information is available in English and other languages spoken in wealthy countries

Word Order Typology

1. SVO: e.g. German, French, English, Mandarin
2. SOV: e.g. Hindi, Japanese
3. VSO: e.g. Irish, Arabic

• SOV = Subject-Verb-Object
• Two languages that share their basic word order type often have other similarities
• e.g. VO languages generally have prepositions, whereas OV languages generally have postpositions

Lexical Divergences

• Languages differ in lexically dividing up the conceptual space, either one-to-many or even many-to-many translation
• Sometimes one language places more grammatical constraints on word choise than another
• One language may have a lexical gap, where no word or phrase, short of an explanatory footnote, can express the exact meaning of a word in the other language
• The field of MT and Word Sense Disambiguation are closely linked
• Languages differ in how the conceptual properties of an event are mapped onto specific words:
  • Verb-framed languages: e.g. Spanish
  • Satellite-framed languages: e.g. English

Morphological Typology

Morphologically, languages are often characterized along two dimensions of variation:

1. The number of morphemes per word
   1. Isolating languages: e.g. Vietnamese and Cantonese, in which each word generally has one morpheme
   2. Polysynthetic languages e.g. Siberian Yupik ("Eskimo"), in which a single word may have very many morphemes, corresponding to a whole sentence in English
2. The degree to which morphemes are segmentatble
   1. Agglutinative languages: e.g. Turkish, in which morphemes have relatively clean boundaries
   2. Fusion languages: e.g. Russian, in which a single affix may conflate multiple morphemes

• Translating between languages with rich morphology requires dealing with structure below the word level
• Thus generally use subword models like WordPiece or BPE

Referential Density

• The differences in frequencies of omission across pro-drop languages, e.g. Japanese and Chinese tend to omit far more than Spanish
• Pro-drop languages: languages that can omit pronouns
• Languages that tend to use more pronouns are more referentially dense than those that use more zeros
• Cold languages: referentially sparse languages, e.g. Chinese or Japanese, require the hearer to do more inferential work to recover antecedents
• Hot languages: languages that are more explicit and make it easier for the hearer

MT Evaluation

• chrF, BLEU