Scribd
Upload
66 views2 pages

NLP Endsem 2016

1. The document provides instructions for a 3 hour exam with 70 total marks covering natural language processing topics. It includes two parts: Part A contains 13 multiple choice questions worth 40 marks, and Part B contains 4 long answer questions worth 30 marks where students must answer 3 of the 4. 2. The questions cover a range of NLP tasks and algorithms including building bilingual dictionaries with machine learning, the Viterbi algorithm, word embeddings, latent semantic analysis, parsing with probabilistic context-free grammars, Hidden Markov Models, smoothing techniques, and more. Detailed explanations are required for full marks. 3. Students must show their work for problems involving algorithms, math, or code

Uploaded by

Puneet Sangal
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
66 views2 pages

NLP Endsem 2016

1. The document provides instructions for a 3 hour exam with 70 total marks covering natural language processing topics. It includes two parts: Part A contains 13 multiple choice questions worth 40 marks, and Part B contains 4 long answer questions worth 30 marks where students must answer 3 of the 4. 2. The questions cover a range of NLP tasks and algorithms including building bilingual dictionaries with machine learning, the Viterbi algorithm, word embeddings, latent semantic analysis, parsing with probabilistic context-free grammars, Hidden Markov Models, smoothing techniques, and more. Detailed explanations are required for full marks. 3. Students must show their work for problems involving algorithms, math, or code

Uploaded by

Puneet Sangal
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

NLP Endsem 2016

Time: 3 hrs Total Marks: 70 Be Precise


======================================================
Important: Answer all questions from PART A, and any 3 out of 4
questions from Part B.
======================================================
Part A (40 marks)
(Answer all questions from this part)
1. Constructing a bilingual dictionary (say English to Hindi) is a non-trivial task.
How can machine learning help? Be specific in identifying the algorithm used and
explaining how parameters are estimated. [3]
2. Identify the specific problem that the Viterbi algorithm solves in the context of
HMMs. Explain the central intuition behind the algorithm using an example. [3]
3. What is CBOW in the context of Word2Vec? Why is it useful? [2]
4. Which of polysemy or synonymy is LSA better at handling, and why? Given a
rectangular matrix of size 2 x 3, how would you compute SVD of this matrix by
hand? [2+2]
5. You are given a set of sentences from an unknown language, which has never
been studied till date. How can you use Expectation Maximization to arrive at the
correct parse of these sentences? [3]
6. Are there situations where the parameters learnt from corpus in PCFG are
successful in correctly parsing some sentences, but unsuccessful over others? If
yes, explain with an example, and suggest a fix. If no, justify (your justification
must be accompanied by a proof sketch). [3]
7. We can use distributional models of similarity (KL divergence or its symmetrised
version) to estimate document and term relatedness from corpus. What
advantage does a method like Latent Semantic Analysis have over this approach?
[2]
8. Bottom up filtering is used to improve the efficiency of top down parsers. Is this
true? If yes, how? If no, correct the sentence and justify. [2]
9. Apart from decision trees, identify a rule induction technique that addresses a
classical problem in NLP. Discuss briefly the central idea behind the approach.
[3]
10. How is the success of HMM parameter learning related to an important property
of KL divergence? [2]
11. What limitation of Laplace smoothing does Good Turing smoothing overcome?
Where can Good Turing smoothing fail? Suggest a repair to overcome this
shortcoming. [2+1.5+1.5]
12. Can you think of any NLP task where knowledge of recall can reduce uncertainty
about precision (or vice versa)? Explain. [2]
13. Briefly explain the connection between branching factor and perplexity with an
example. [3]
14. What are Hearst patterns and what are they used for? Explain briefly with two
examples. Identify a limitation of Hearst patterns. [2+2]
15. There are only two words A and B in a corpus, and two (hidden) topics that can
generate these words. We have ten documents, each having 20 tokens of type A
and B. Suggest an approach to estimate (a) the probabilities with which each
topic generates A and B (b) estimate the belongingness of each document to the
two topics. Identify clearly all assumptions you may have made. Are there any
criteria that the document collection should ideally satisfy? [6]
16. What are rhetorical relations and in which context are they useful? Explain with
two examples. [3]

(Please Turn Over)


Part B ( 30 marks)
(Answer any three questions from this part)

1. (a) Use dynamic programming to compute the edit distance between words
WRONG and WINGS, assuming that the costs of insertion and deletion are 2 and
the cost of substitution is 1. [6]

(b) Assume that there are only two senses of the word bank (one pertaining to the
financial sense, and the other to the “river bank” sense) in WordNet. In a given
piece of raw text, the distributional neighbours of bank are {account, deposit,
river}. Given that each of these neighbours can have multiple senses as well, how
would you go about assigning dominance based ranks to the two senses of bank?
Show the steps in detail. [4]

2. Consider a Machine Translation parallel corpus having three sentence pairs. The
first sentence pair is “go there fast”/”jaldi udhar jaao”. The second sentence pair is
“go there”/”udhar jaao”. The third sentence pair is “go”/”jaao”. (a) Show how the
first few iterations of EM are useful in learning word alignments from this corpus.
Make clear any simplifying assumptions on top of IBM Model 3. (b) How is extra
knowledge “getting generated” in successive iterations of EM? [8+2]

3. What limitations of the basic parsing techniques does the CYK parser address? Is
there an assumption on the grammar rules that CYK can deal with? If yes, what
are these? Given the grammar below and the input sentence “w =(()(()))”, show the
steps in chart parsing using CYK. Alongside your charts showing each step,
mention clearly the rule(s) that is(are) used (if any) to advance to this step from
the previous one. [1.5 + 1.5 + 7]
S → SS
S →(S1
S1 → S)
S → ()

4. A PCFG is based on the following rules:


a. S → A B
b. B → D A
c. B → D A C
d. A → A C
e. A → a
f. A → b c
g. A → b d e
h. C → f g h
i. D → i
The corpus has the following two sentences, the first occurring 15 times and the
second 30 times:
1. a i b c f g h
2. b c i b d e
(a) Are the sentences accepted by the grammar? In case both of them are, which
of these two sentences is/are ambiguous? Show all possible parse trees of the
sentence(s).
(b) Make an APPROPRIATE initial choice of the rule probabilities. Show the
first three steps of the EM algorithm for estimating the parameters of this
PCFG. [3+7]

== The End ==

You might also like