Assignment 8: Machine Translation

Natural Language Processing:
Assignment 8: Machine Translation

Introduction
As always, check out the Github repository with the course homework templates:
git://github.com/ezubaric/cl1-hw.git
The code for this homework is in the hw8 directory. The goal of this
homework is for you to build a machine translation scoring function based
on IBM Model 1.
Data
The data are gzipped samples from Europarl. And can be found in the
GitHub directory with the code.
Also in the directory are two text files with lists of words. These can help
you monitor the progress of the algorithm to see if the lexical translation
probabilities look like what they should.
What to Do
You have been given a partial EM implementation of IBM Model 1 translating foreign (f) from English (e).1 The maximization function is complete,
but the expectation is not, nor is the function to score a complete translation
pair. You need to fill in two functions.
1This is different from what we initially discussed in class; we’re applying the noisychannel model here: p(f|e)p(e), which translates into translating from English to foreign
mathematically, even if the system is to translate from foreign to English.
1
Generating Counts (20 points)
The first function you need to fill in is sentence counts, which should
return an iterator over english and foreign words pairs along with their
expected count. The expected count is the expected number of times that
word pair was translated in the sentence, given by the equation
c(f|e; ~e, ~f) = X
a
p(a|~e, ~f)
X
lf
j=1
δ(f, fj )(e, ea(j)
). (1)
Scoring (10 points)
The second function you need to fill in is the noisy channel scoring method
translate score, which is the translation probability (given by Model 1)
times the probability of the English output under the language model.
Running the Model (10 points)
Run the model and produce the lexical translation table for the development
words. Don’t leave this until the last moment, because this can take a while.
How to solve the Problem
Don’t start using the big data immediately. Start with the small data. If
you run a correct implementation on the toy data, you should get something
like the following:
What to turn in
You will turn in your complete ibm trans.py file and the word translations
for the supplied test file (devwords.txt) after three (3) iterations of EM.
Extra Credit (5 points)
If you would like extra credit, add an additional function that computes the
best alignment between a pair of sentences.
2
Listing 1: Successful output from running on toy data ✞ ☎
1 python ibm_trans.py 5 toy toy_test.txt
2 Corpus <__main__.ToyCorpus instance at 0x27d08c0, 5 m1 iters
3 Model1 Iteration 0 ...
4 Sentence 0
6 blind:blind:0.250000 an:0.138889 hier:0.138889 von:0.138889
7 bist:0.111111 du:0.111111 nicht:0.111111 irish:du:0.166667
8 irisch:0.166667 kind:0.166667 mein:0.166667 wilest:0.166667
9 wo:0.166667 is:ist:0.181818 der:0.090909 fahrrad:0.090909
10 himmel:0.090909 mein:0.090909 nicht:0.090909 noch:0.090909
11 rosa:0.090909 siebte:0.090909 weit:0.090909 cries:auch:0.200000
12 der:0.200000 himmel:0.200000 weint:0.200000 wo:0.200000
13 are:du:0.206897 bist:0.120690 blind:0.120690 nicht:0.120690
14 irisch:0.086207 kind:0.086207 mein:0.086207 wilest:0.086207
15 wo:0.086207
17 done
19 ...
21 ========COMPUTING WORD TRANSLATIONS=======
22 balloons 99 0.203438
23 not blind 0.028217
24 blind blind 0.693760
25 jets 99 0.341654
27 LM Sentence 0
28 ========COMPUTING SENTENCE TRANSLATIONS=======
29 ENGLISH: my bike cries 99 fighter jets
30 FOREIGN: mein fahrrad weint 99 dusenjaeger
31 SCORE: −30.100041
33 ENGLISH: the heaven is pink about now
34 FOREIGN: von hier an bleibe der himmel rosa
35 SCORE: −44.224041
✡✝ ✆
3
Questions
1. My code is really slow; is it okay if I don’t run my code on the entire
dataset? (by using the limit argument)
Yes, but if your numbers look too far off from what is expected, you
may lose points.
2. I’m getting probabilities greater than 1.0!
The logprob function of nltk’s language model returns the negative
log probability.
3. I can’t get my translation probabilities to match.
• One little thing that may cause a slight difference in computing
the translation probability: I add 0.001 to each of the inputs (to
guard against zeros) while summing up across the foreign string.
• You will need to add a None to the English sentence in the
translate score function. The assert is there to make sure it
hasn’t happened outside of the function.
• Make sure you compute the lm probabilities conditioning the first
word on the empty string (’’).
4