Machine Translation

Question 1

Machine Translation

Answer 1

Automatically translating from one language to another

Question 2

Why is machine translation difficult?

Answer 2

• Word order differences
• Vocabulary gaps
• Metaphor, idioms, collocations
• Cultural difference

Question 3

Vaquois triangle

Answer 3

Depth of linguistic analysis vs. distance to be covered by the translation

Furthest to closest:

• Word to word
• Syntax to syntax
• Semantic to semantic
• Interlingua

Question 4

How to build a translation model using a parallel corpus with word alignments

Answer 4

Learn the co-occurrence probabilities

Compute conditional probability as a relative frequency

P (witch | bruja)
= count (witch + bruja) / count(witch)

Question 5

How to build a translation model using a parallel corpus with translational probabilities

Answer 5

Compute the alignments that maximize the overall probabilities

Question 6

How to build a translation model from a parallel text using expectation maximization

Answer 6

Start by setting all translational probabilities to be uniform.

Compute word alignment probabilities from them.

Now recompute translation probabilities based on these word alignments.

Then recompute word alignments based on the new translational probabilities.

Repeat until convergence.

Question 7

Statistical approach to machine translation: two questions

Answer 7

Which words and phrases in the source language translate to which words and phrases in the target language?

How do I best phrase things (in a natural way) in the target language?

Question 8

Machine translation

Decoder: Given F, find most likely E

Answer 8

Language model: P(E)
Translation model: P(F | E)

P(E | F) = argmax P(F | E) * P(E)
(drop P(F) from Bayes rule because it is always the same)