CS 224N -- Ling 237: Natural Language Processing Spring 2003 FAQs

FAQs

Homework #1:

• What level of detail is desired in drawing the parse trees in Part A1 ?
  
  We are just looking for something that is generally correct - you are not expected to be accurate on each and every word. It might be useful to look at Pennbank trees to answer your doubts about how specific kinds of constructions should be handled. All the same, you need not follow the precise Penn Treebank notation. You might also find it useful to look at Chapter 3 in the textbook for ideas.
  

Homework #2:

• In problem 3, why do we have conditional probability for absolute discounting and joint probability for linear discounting ?
  
  There is no particular reason for this. For uniformity, you can assume linear discounting is also defined for the conditional probability P(wn|w1,..,wn-1) and give your proof for this case. Doing the proof for the joint probability is also fine - in this case N0 will be the number of unseen word n-tuples and N will be the number of times any word n-tuple was seen.
  


• In problem 6b, some files have almost all of their text between subsample tags. Should we also delete this text?
  
  You may choose to keep or delete the text between these tags. Let us know what you did.
  

Programming Project #2:

• How can I get a morphological analyzer if I want one ?
  
  Some possibilities are listed below:
  • On the Leland systems, /afs/ir/class/cs224n/src/morph-1.5 contains the Penn XTAG project morphological analyzer. It does a fairly good job of determining and returning the base form (lemma) and inflection of English words. It is written in C.
  • WordNet has a sort of morphological analyzer: you have to give a part of speech, but it then returns a base form. Not as comprehensive as the above, but there is a Java interface. These were installed for cs276a. You can find details here.
    The version in the cs276a directory seems like it'll still run fine for us.
  • You could use a stemmer. These generally do more in that they also strip words of their inflections, also removing derivational affixes, but are more heuristic, rather than lexicon based. One is the porter stemmer here. The page includes versions in C, Java and Perl (among others).
  
  
• How can I get a part of speech tagger ?
  
  
  • On the Leland systems, /afs/ir/class/cs224n/src/ contains the Brill and icopost taggers that you can use.
  • For an easier but approximate solution, you can use the list given in: /afs/ir/class/cs224n/src/icopost-0.9.1/data/english-wsj-train-0-18.lex This file contains a list of all tokens in a particular corpus, and the list of part of speech tags that were observed each. Counts are also kept for number of total occurences and the number of occurences as each tag.
  
  

Homework #3:

• In problem 1, 'an' does not seem to have a rule in the grammar?
  
  This was a small bug in the version handed out in lecture. Replace the rule   Det->the   by the rule  Det->an . The correction has been made on the homework version available on the handouts page.
  


• In problem 4, the third tree can not be generated by the grammar. Is there a bug?
  
  Yes, there is a bug in the rule set. Replace S -> AB with S -> BA in the rule set.
  


• In problem 4b, how do you define training set error rate?
  
  We look at the PCFG as a classifier that, for an input sentence , returns its most likely parse tree. If the returned tree is different than the tree that appears for the sentence in the training set, then this is a training set error. The error rate is the fraction of errors and is a number between 0 an 1. Error rate = num_wrong/num_total. For the example in problem 4b, we will have num_total=11.
  


• In problem 4c, how do you define conditional log-likelihood?
  
  The conditional probability for a tree t for a sequence of words W according to a PCFG grammar is P(t|W)=P(t,W)/P(W)=P(t)/P(W), since the sentence W is part of the tree. Then the conditional log-likelihood of the training corpus will be the sum of the logs of the conditional probabilities of the trees that occur in the corpus.
  


• In problem 5a, do we need to specify lexical entries and grammar rules that will handle the sentence or is it enough to provide the final semantics of the sentence?
  
  It is enough to write an expression for the semantics of the whole sentence.
  


• In problem 5b, I have difficulty with handling the plural noun phrase without a determiner.
  
  You can assume that the sentence actually is "Which customers purchased the red leather jackets?".