316 MT

Question 1

translation divergences

Answer 1

• differences between languages
• typology-how different languages have different affixes to them
  *structural- the structures in which sentences are set..SVO SOV
• lexical divergences- difference in ambiguity and grammar
• lexical gap- words that doesnt have a meaning in another language..rivere

Question 2

ambiguity resolution

Answer 2

words that are spelt the same in one language but has two separate words in another language..he plays the guitar..he plays soccer

Question 3

direct translation

Answer 3

morphological analysis,lexical transfer,local reordering, morphological generation.
* word to word translations.

Question 4

syntactic transfer
semantic transfer

Answer 4

analysis,shallow analysis
transfer(lexica,syntactic),transfer phase
generation,synthesis phase
requires both syntactic and direct translation
one makes possible parse trees then finds the most likely parse tree.

Question 5

a good translation should be:

Answer 5

faithful-correctly convery information and tone
fluent- grammatically well structured and readable

Question 6

phrase based translation

Answer 6

translation probability-
translation model-tells us what a setnence/phrase in a source language most likely translates to(faithfulness)

distortion probability- tells how likely a given sentence/phrase is(language model)(fluency)

Question 7

encoder and decoder

Answer 7

encoder is used for language understanding in source languae
decode is used for language generation in target language

Question 8

seq2seq encoder

Answer 8

the encoder usually uses an LSTm to generate context vectro from the output
*input tokes are read:
-one at time
-in reverse order
* the encoder conosists of stacked LSTMS
* column-timestep
row- single layer

Question 9

human rater evaluations

Answer 9

• fluency
  -asking raters to rate output interms of readability,clarity and naturalness
• cloze test:
  if theres a space within the output and someone can accurately depict the word that showed be there..then the machine is doing well
• fidelity:
  -adequacy-measures whether a translation contains the info that existed in the source
  -informativeness-text based evaluation if theres sufficient info in the output

Question 10

bleu

Answer 10

the closer machine translation is to a professional human translation,the better it is
-if results are needed urgently
* two constraints
a gram in the reference translation cannot be matched more than once
brevity penalty-very small sentences that would achieve a 1.0 precision is penalised