Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: daixieit

955G5

MSc EXAMINATION

August 2021 (A3)

Applied Natural Language processing

(b） You  collect  100,000  posts  which  have  been  labelled  as  positive  or negative by humans and divide them into a training and testing set with 50,000 posts in each.

i. You run myfunction（） on your training set.  ln order to visualise the word type frequency distribution, what do you need to do with the output of myfunction（）?                                                   [4 marks]

ii.  sketch the shape of the word type frequency distribution you would  expect to see.                                                                    [4 marks] 

iii.  Explain why this shape of distribution is a problem for NLP systems. [4 marks]

iv.  Explain  4  modiications  or  pre-processing  techniques  you  could introduce to myfunction（）  which would  alleviate this  problem to some extent.   lnclude  examples to illustrate your answer.   which do you think would be the most effective and why?           [10 marks]

(c） You decide to build a Naive Bayes classiier.

i. what probabilities need to be estimated based on the training data?  why?            [4 marks] 

ii.  Explain how to calculate these from the training data.         [4 marks]

iii.  Explain how and why you would smooth the probabilities.  [5 marks] 

(d） You test your classiier on the testing set and its accuracy is 90%. Give 2 possible different conclusions you could  reach from this evidence.

Explain how an evaluation based on precison and recall could help you come to the correct conclusion.                                               [5 marks]

2.   (a） Describe 2 applications in NLP where it could be useful to be able to compute the similarity between pairs of documents.                  [5 marks]

(b） consider  Table  1， which  gives  the  frequencies  of  5  potential word features in a very small corpus of 5 documents. Assume that there are no other words and no other documents in the corpus.

Table 1: Frequencies of 5 word Features in 5 Documents

i.  For documents A and B compute the tf-idf score associated with each of the 5 word features given.       [10 marks]

ii.  For documents A and B compute the positive pointwise mutual  information  (PPMI)  between the document and each of the 5  word features given.                                                        [10 marks] 

iii. why is a representation based on tf-idf or PPMI better than one based  on  raw  frequency when  considering  document  similarity? [5 marks]

(c） consider the following snippet of code

def   afunc ( docA ，docB ）:

the  sum=0

for   ( key ，value ） in   docA . items ( ）:

the sum+= value *docB . get ( key ，0）

return  the   sum

Assuming that the input to afunc（）consists of two dictionaries，storing the frequency-based representations of Doc A and Doc B given in Table 1:

i. what would be the output?                                                   [5 marks]

ii.  Explain how the output is produced by the code.                 [5 marks]

iii.  Explain how this is related to the similarity of 2 documents in the  vector space.           [5 marks] 

(d） Give  an  example of  how  variation  and  ambiguity  each  might  cause unexpected  results when calculating document similarity  in this way. [5 marks]

3. consider the following document snippet, taken from wikipedia:

on  28  June   1999,   Jack   Ma   and   17   friends   and  students  founded  Alibaba.com,  a  china-based  B2B  marketplace  site,  in  his  Hangzhou apartment.   ln  october  1999,  Alibaba  received  a  Us$25  million  investment  from  Goldman  sachs  and  softBank.  Alibaba.com was expected to improve the domestic e-commerce  market and perfect an e-commerce platform for chinese enterprises, especially small and  medium-sized  enterprises  (sMEs）,  to  help  export chinese products to the global market as well as address  world Trade organization (wTo）challenges. ln 2002, Alibaba.com  became proitable three years after launch.  Ma wanted to improve  the global  e-commerce  system,  so  from  2003  onward,  Alibaba  launched Taobao Marketplace, Alipay, Alimama.com, and Lynx.

(a） ldentify 4 different types of named entity mentioned in the text. Give at  least one example of each.                                                      [8 marks] 

(b） what is the difference between a system which carries out coreference resolution  and  a  system  which  carries  out  named  entity  linking? [5 marks]

(c） Give 2 examples of variation and 2 examples of ambiguity for named entities.  Explain why variation and ambiguity are problems for named entity linking.                                                                            [8 marks] 

(d） outline a plan for building a named entity linking system.  How would you account for variation and ambiguity?             [12 marks]

(e） For each of the following applications, would Named Entity Recognition and Named Entity Linking components be useful? Justify your answers. [12 marks] 

i.  lnformation-Retrieval based Factoid Question Answering system.   

ii.  knowledge-based Factoid Question Answering system.

iii. Automatic speech Transcription.

iv.  Machine Translation.

(f） Describe 2 additional challenges for Named Entity Recognition in the context of 1 or more of the applications from question 3e.         [5 marks]