Alex Fraser
NSSNLP, Kathmandu University

In this short assignment we will again look at Google Translate for
the language you used for assignment 1.  We will then take a quick
look at the Indian Parallel Corpora for that same language.

1) Take the 5 sentences for which you got bad output from Google
Translate. Translate them again, preferably from a different IP
address (for instance, if you translated them in the lab, try
translating them from KUIC).

1a) Do you get the same output as before? Were your corrections
partially or fully adopted?

1b) Go back to the bad output you obtained previously. Now that you
know about phrase-based SMT learned from word-aligned parallel
corpora, can you provide a different analysis of the bad output?

2) Take the first 5 sentences of the *training* data for the language
you are working with from the Indian Parallel Corpora. Try these
sentences in Google Translate (in the Hindi/Tamil/Urdu to English
direction, as you did in assignment 1). Compare the outputs with the
English parallel sentences (recall that these sentences were created
using Amazon Mechanical Turk). What observations can you make about
these two different English translations? Is the Google Translate
output correct? Can you analyze any issues with it which are obviously
caused by phrase-based SMT?