2012-02-08
The Rosetta Stone
- Hieroglyphs/Egyptian Demotic/Ancient Greek
Challenges
1. Lexical ambiguity
- One wordform, many meanings:
- (en) Kill: (fr) Tuer, terminer;
- (en) Book: (fr) Livre, reserver;
- (en) Chair: (fr) Chaise, Président; ...
- (en) Buffalo: (fr) (ville de) Buffalo, NY; buffle; bison
2. Differing word orders
- (en) [Subject] – ((trans.) Verb) – {Object} – ("IBM bought Lotus", "[The dog] (bit) {the boy}")
- (jp) [Subject] – {Object} – ((trans.) Verb) – ("IBM Lotus bought", "[The dog] {the boy} (bit)")
- (en) [Det.] – (Adj) – {Noun} – ("[The] (red) {house}")
- (fr) [Det.] – {Noun} – (Adj) – ("[The] {house} (red)" "la maison rouge")
3. Unpreserved syntax
- La botella entró a la cuerva flotando
(the bottle entered the cave floating)
- The bottle floated into the cave
4. Syntactic ambiguity
- "Visiting relatives can be a nuisance."
- "Visiter la famille peut être ennuyant"
- "Les membres de la famille qui visitent peuvent être ennuyants"
5. Idiosyncracies
- "Burn the midnight oil"
- "Travailler tard" (work late)
- "Faire la guele" ("do the snout")
Classical Machine Translation: Dictionaries
- Few rules... if preceding word is ... then ...
Translate 'much' or 'many' into Russian
if preceding word is how return skol’ko
else if preceding word is as return stol’ko zhe
else if word is much
if preceding word is very return nil
else if following word is a noun return mnogo
else (word is many)
if preceding word is a preposition and next word is a noun
return mnogii
else return mnogo
- No analysis
- word order can seem funny in translation
- Part of speech
Classical Machine Translation: Transfer-based approach
- Transfer-based approach involves three phases:
- Analysis: e.g., build syntactic parse trees of the source sentence.
- Transfer: e.g., convert the source-language parse tree to a target-language parse tree.
- Generation: e.g., produce an output sentence from thetarget-language parse tree.
- These systems can involve fairly deep analysis, including semantic analysis.
Source
Jim says
I suggest that to really understand this stuff - and explain it to me - you try to code (likely in python) the simplest toy model you can invent to implement this (model1 -> model2 -> ...) chain of counting/probability calculations, as described in the "Knight: Statistical MT Workbook". (Note that there's an FAQ related document which looks like its actually a correction.)
