Finding head of a noun phrase in NLTK and stanford

Question

generally A head of a nounphrase is a noun which is rightmost of the NP as shown below tree is the head of the parent NP. So

            ROOT                             
             |                                
             S                               
          ___|________________________        
         NP                           |      
      ___|_____________               |       
     |                 PP             VP     
     |             ____|____      ____|___    
     NP           |         NP   |       PRT 
  ___|_______     |         |    |        |   
 DT  JJ  NN  NN   IN       NNP  VBD       RP 
 |   |   |   |    |         |    |        |   
The old oak tree from     India fell     down

Out[40]: Tree('S', [Tree('NP', [Tree('NP', [Tree('DT', ['The']), Tree('JJ', ['old']), Tree('NN', ['oak']), Tree('NN', ['tree'])]), Tree('PP', [Tree('IN', ['from']), Tree('NP', [Tree('NNP', ['India'])])])]), Tree('VP', [Tree('VBD', ['fell']), Tree('PRT', [Tree('RP', ['down'])])])])

The following code based on a java implementation uses a simplistic rule to find the head of the NP , but i need to be based on the rules:

parsestr='(ROOT (S (NP (NP (DT The) (JJ old) (NN oak) (NN tree)) (PP (IN from) (NP (NNP India)))) (VP (VBD fell) (PRT (RP down)))))'
def traverse(t):
    try:
        t.label()
    except AttributeError:
          return
    else:
        if t.label()=='NP':
            print 'NP:'+str(t.leaves())
            print 'NPhead:'+str(t.leaves()[-1])
            for child in t:
                 traverse(child)

        else:
            for child in t:
                traverse(child)


tree=Tree.fromstring(parsestr)
traverse(tree)

The above code gives output:

NP:['The', 'old', 'oak', 'tree', 'from', 'India'] NPhead:India NP:['The', 'old', 'oak', 'tree'] NPhead:tree NP:['India'] NPhead:India

Although now its giving correct output for the sentence given but I need to incorporate a condition that only right most noun is extracted as head , currently it does not check if it were a noun (NN)

print 'NPhead:'+str(t.leaves()[-1])

So something like following in the np head condition in above code:

t.leaves().getrightmostnoun() 

Michael Collins dissertation (Appendix A) includes head-finding rules for the Penn Treebank, and hence it is not necessary that only the rightmost noun is the head. Hence the above conditions should incorporate such scenario.

For the following example as given in one of the answers:

(NP (NP the person) that gave (NP the talk)) went home

The head noun of the subject is person but the last leave node of the NP the person that gave the talk is talk.

Answer 1

I was looking for a python script using NLTK that does this task and stumbled across this post. Here's the solution I came up with. It's a little bit noisy and arbitrary, and definitely doesn't always pick the right answer (e.g. for compound nouns). But I wanted to post it in case it was helpful for others to have a solution that mostly works.

#!/usr/bin/env python

from nltk.tree import Tree

examples = [
    '(ROOT (S (NP (NP (DT The) (JJ old) (NN oak) (NN tree)) (PP (IN from) (NP (NNP India)))) (VP (VBD fell) (PRT (RP down)))))',
    "(ROOT\n  (S\n    (NP\n      (NP (DT the) (NN person))\n      (SBAR\n        (WHNP (WDT that))\n        (S\n          (VP (VBD gave)\n            (NP (DT the) (NN talk))))))\n    (VP (VBD went)\n      (NP (NN home)))))",
    '(ROOT (S (NP (NN Carnac) (DT the) (NN Magnificent)) (VP (VBD gave) (NP ((DT a) (NN talk))))))'
]

def find_noun_phrases(tree):
    return [subtree for subtree in tree.subtrees(lambda t: t.label()=='NP')]

def find_head_of_np(np):
    noun_tags = ['NN', 'NNS', 'NNP', 'NNPS']
    top_level_trees = [np[i] for i in range(len(np)) if type(np[i]) is Tree]
    ## search for a top-level noun
    top_level_nouns = [t for t in top_level_trees if t.label() in noun_tags]
    if len(top_level_nouns) > 0:
        ## if you find some, pick the rightmost one, just 'cause
        return top_level_nouns[-1][0]
    else:
        ## search for a top-level np
        top_level_nps = [t for t in top_level_trees if t.label()=='NP']
        if len(top_level_nps) > 0:
            ## if you find some, pick the head of the rightmost one, just 'cause
            return find_head_of_np(top_level_nps[-1])
        else:
            ## search for any noun
            nouns = [p[0] for p in np.pos() if p[1] in noun_tags]
            if len(nouns) > 0:
                ## if you find some, pick the rightmost one, just 'cause
                return nouns[-1]
            else:
                ## return the rightmost word, just 'cause
                return np.leaves()[-1]

for example in examples:
    tree = Tree.fromstring(example)
    for np in find_noun_phrases(tree):
        print "noun phrase:",
        print " ".join(np.leaves())
        head = find_head_of_np(np)
        print "head:",
        print head

For the examples discussed in the question and in the other answers, this is the output:

noun phrase: The old oak tree from India
head: tree
noun phrase: The old oak tree
head: tree
noun phrase: India
head: India
noun phrase: the person that gave the talk
head: person
noun phrase: the person
head: person
noun phrase: the talk
head: talk
noun phrase: home
head: home
noun phrase: Carnac the Magnificent
head: Magnificent
noun phrase: a talk
head: talk

Answer 2

There are built-in string to Tree object in NLTK (http://www.nltk.org/_modules/nltk/tree.html), see https://github.com/nltk/nltk/blob/develop/nltk/tree.py#L541.

>>> from nltk.tree import Tree
>>> parsestr='(ROOT (S (NP (NP (DT The) (JJ old) (NN oak) (NN tree)) (PP (IN from) (NP (NNP India)))) (VP (VBD fell) (PRT (RP down)))))'
>>> for i in Tree.fromstring(parsestr).subtrees():
...     if i.label() == 'NP':
...             print i
... 
(NP
  (NP (DT The) (JJ old) (NN oak) (NN tree))
  (PP (IN from) (NP (NNP India))))
(NP (DT The) (JJ old) (NN oak) (NN tree))
(NP (NNP India))


>>> for i in Tree.fromstring(parsestr).subtrees():
...     if i.label() == 'NP':
...             print i.leaves()
... 
['The', 'old', 'oak', 'tree', 'from', 'India']
['The', 'old', 'oak', 'tree']
['India']

Note that it's not always the case that right most noun is the head noun of an NP, e.g.

>>> s = '(ROOT (S (NP (NN Carnac) (DT the) (NN Magnificent)) (VP (VBD gave) (NP ((DT a) (NN talk))))))'
>>> Tree.fromstring(s)
Tree('ROOT', [Tree('S', [Tree('NP', [Tree('NN', ['Carnac']), Tree('DT', ['the']), Tree('NN', ['Magnificent'])]), Tree('VP', [Tree('VBD', ['gave']), Tree('NP', [Tree('', [Tree('DT', ['a']), Tree('NN', ['talk'])])])])])])
>>> for i in Tree.fromstring(s).subtrees():
...     if i.label() == 'NP':
...             print i.leaves()[-1]
... 
Magnificent
talk

Arguably, Magnificent can still be the head noun. Another example is when the NP includes a relative clause:

(NP (NP the person) that gave (NP the talk)) went home

The head noun of the subject is person but the last leave node of the NP the person that gave the talk is talk.