NLTK relation extraction returns nothing

Question

I am recently working on using nltk to extract relation from text. so i build a sample text:" Tom is the cofounder of Microsoft." and using following program to test and return nothing. I cannot figure out why.

I'm using NLTK version: 3.2.1, python version: 3.5.2.

Here is my code:

import re
import nltk
from nltk.sem.relextract import extract_rels, rtuple
from nltk.tokenize import sent_tokenize, word_tokenize


def test():
    with open('sample.txt', 'r') as f:
        sample = f.read()   # "Tom is the cofounder of Microsoft"

    sentences = sent_tokenize(sample)
    tokenized_sentences = [word_tokenize(sentence) for sentence in sentences]
    tagged_sentences = [nltk.tag.pos_tag(sentence) for sentence in tokenized_sentences]

    OF = re.compile(r'.*\bof\b.*')

    for i, sent in enumerate(tagged_sentences):
        sent = nltk.chunk.ne_chunk(sent) # ne_chunk method expects one tagged sentence
        rels = extract_rels('PER', 'GPE', sent, corpus='ace', pattern=OF, window=10) 
        for rel in rels:
            print('{0:<5}{1}'.format(i, rtuple(rel)))

if __name__ == '__main__':
    test()

---

1.　After some debug, if found that when i changed the input as

"Gates was born in Seattle, Washington on October 28, 1955. "

the nltk.chunk.ne_chunk() output is:

(S (PERSON Gates/NNS) was/VBD born/VBN in/IN (GPE Seattle/NNP) ,/, (GPE Washington/NNP) on/IN October/NNP 28/CD ,/, 1955/CD ./.)

The test() returns:

[PER: 'Gates/NNS'] 'was/VBD born/VBN in/IN' [GPE: 'Seattle/NNP']

2. After i changed the input as:

"Gates was born in Seattle on October 28, 1955. "

The test() retuns nothing.

3. I digged into nltk/sem/relextract.py and find this strange

output is caused by function: semi_rel2reldict(pairs, window=5, trace=False), which returns result only when len(pairs) > 2, and that's why when one sentence with less than three NEs will return None.

Is this a bug or i used NLTK in wrong way?

Answer 1

Firstly, to chunk NEs with ne_chunk, the idiom would look something like this

>>> from nltk import ne_chunk, pos_tag, word_tokenize
>>> text = "Tom is the cofounder of Microsoft"
>>> chunked = ne_chunk(pos_tag(word_tokenize(text)))
>>> chunked
Tree('S', [Tree('PERSON', [('Tom', 'NNP')]), ('is', 'VBZ'), ('the', 'DT'), ('cofounder', 'NN'), ('of', 'IN'), Tree('ORGANIZATION', [('Microsoft', 'NNP')])])

(see also https://stackoverflow.com/a/31838373/610569)

Next let's look at the extract_rels function.

def extract_rels(subjclass, objclass, doc, corpus='ace', pattern=None, window=10):
    """
    Filter the output of ``semi_rel2reldict`` according to specified NE classes and a filler pattern.
    The parameters ``subjclass`` and ``objclass`` can be used to restrict the
    Named Entities to particular types (any of 'LOCATION', 'ORGANIZATION',
    'PERSON', 'DURATION', 'DATE', 'CARDINAL', 'PERCENT', 'MONEY', 'MEASURE').
    """

When you evoke this function:

extract_rels('PER', 'GPE', sent, corpus='ace', pattern=OF, window=10)

It performs 4 processes sequentially.

1. It checks whether your subjclass and objclassare valid

i.e. https://github.com/nltk/nltk/blob/develop/nltk/sem/relextract.py#L202 :

if subjclass and subjclass not in NE_CLASSES[corpus]:
    if _expand(subjclass) in NE_CLASSES[corpus]:
        subjclass = _expand(subjclass)
    else:
        raise ValueError("your value for the subject type has not been recognized: %s" % subjclass)
if objclass and objclass not in NE_CLASSES[corpus]:
    if _expand(objclass) in NE_CLASSES[corpus]:
        objclass = _expand(objclass)
    else:
        raise ValueError("your value for the object type has not been recognized: %s" % objclass)

2. It extracts "pairs" from your NE tagged inputs:

if corpus == 'ace' or corpus == 'conll2002':
    pairs = tree2semi_rel(doc)
elif corpus == 'ieer':
    pairs = tree2semi_rel(doc.text) + tree2semi_rel(doc.headline)
else:
    raise ValueError("corpus type not recognized")

Now let's see given your input sentence Tom is the cofounder of Microsoft, what does tree2semi_rel() returns:

>>> from nltk.sem.relextract import tree2semi_rel, semi_rel2reldict
>>> from nltk import word_tokenize, pos_tag, ne_chunk
>>> text = "Tom is the cofounder of Microsoft"
>>> chunked = ne_chunk(pos_tag(word_tokenize(text)))
>>> tree2semi_rel(chunked)
[[[], Tree('PERSON', [('Tom', 'NNP')])], [[('is', 'VBZ'), ('the', 'DT'), ('cofounder', 'NN'), ('of', 'IN')], Tree('ORGANIZATION', [('Microsoft', 'NNP')])]]

So it returns a list of 2 lists, the first inner list consist of a blank list and the Tree that contains the "PERSON" tag.

[[], Tree('PERSON', [('Tom', 'NNP')])] 

The second list consist of the phrase is the cofounder of and the Tree that contains "ORGANIZATION".

Let's move on.

3. extract_rel then tries to change the pairs to some sort of relation dictionary

reldicts = semi_rel2reldict(pairs)

If we look what the semi_rel2reldict function returns with your example sentence, we see that this is where the empty list gets returns:

>>> tree2semi_rel(chunked)
[[[], Tree('PERSON', [('Tom', 'NNP')])], [[('is', 'VBZ'), ('the', 'DT'), ('cofounder', 'NN'), ('of', 'IN')], Tree('ORGANIZATION', [('Microsoft', 'NNP')])]]
>>> semi_rel2reldict(tree2semi_rel(chunked))
[]

So let's look into the code of semi_rel2reldict https://github.com/nltk/nltk/blob/develop/nltk/sem/relextract.py#L144:

def semi_rel2reldict(pairs, window=5, trace=False):
    """
    Converts the pairs generated by ``tree2semi_rel`` into a 'reldict': a dictionary which
    stores information about the subject and object NEs plus the filler between them.
    Additionally, a left and right context of length =< window are captured (within
    a given input sentence).
    :param pairs: a pair of list(str) and ``Tree``, as generated by
    :param window: a threshold for the number of items to include in the left and right context
    :type window: int
    :return: 'relation' dictionaries whose keys are 'lcon', 'subjclass', 'subjtext', 'subjsym', 'filler', objclass', objtext', 'objsym' and 'rcon'
    :rtype: list(defaultdict)
    """
    result = []
    while len(pairs) > 2:
        reldict = defaultdict(str)
        reldict['lcon'] = _join(pairs[0][0][-window:])
        reldict['subjclass'] = pairs[0][1].label()
        reldict['subjtext'] = _join(pairs[0][1].leaves())
        reldict['subjsym'] = list2sym(pairs[0][1].leaves())
        reldict['filler'] = _join(pairs[1][0])
        reldict['untagged_filler'] = _join(pairs[1][0], untag=True)
        reldict['objclass'] = pairs[1][1].label()
        reldict['objtext'] = _join(pairs[1][1].leaves())
        reldict['objsym'] = list2sym(pairs[1][1].leaves())
        reldict['rcon'] = _join(pairs[2][0][:window])
        if trace:
            print("(%s(%s, %s)" % (reldict['untagged_filler'], reldict['subjclass'], reldict['objclass']))
        result.append(reldict)
        pairs = pairs[1:]
    return result

The first thing that semi_rel2reldict() does is to check where there are more than 2 elements the output from tree2semi_rel(), which your example sentence doesn't:

>>> tree2semi_rel(chunked)
[[[], Tree('PERSON', [('Tom', 'NNP')])], [[('is', 'VBZ'), ('the', 'DT'), ('cofounder', 'NN'), ('of', 'IN')], Tree('ORGANIZATION', [('Microsoft', 'NNP')])]]
>>> len(tree2semi_rel(chunked))
2
>>> len(tree2semi_rel(chunked)) > 2
False

Ah ha, that's why the extract_rel is returning nothing.

Now comes the question of how to make extract_rel() return something even with 2 elements from tree2semi_rel()? Is that even possible?

Let's try a different sentence:

>>> text = "Tom is the cofounder of Microsoft and now he is the founder of Marcohard"
>>> chunked = ne_chunk(pos_tag(word_tokenize(text)))
>>> chunked
Tree('S', [Tree('PERSON', [('Tom', 'NNP')]), ('is', 'VBZ'), ('the', 'DT'), ('cofounder', 'NN'), ('of', 'IN'), Tree('ORGANIZATION', [('Microsoft', 'NNP')]), ('and', 'CC'), ('now', 'RB'), ('he', 'PRP'), ('is', 'VBZ'), ('the', 'DT'), ('founder', 'NN'), ('of', 'IN'), Tree('PERSON', [('Marcohard', 'NNP')])])
>>> tree2semi_rel(chunked)
[[[], Tree('PERSON', [('Tom', 'NNP')])], [[('is', 'VBZ'), ('the', 'DT'), ('cofounder', 'NN'), ('of', 'IN')], Tree('ORGANIZATION', [('Microsoft', 'NNP')])], [[('and', 'CC'), ('now', 'RB'), ('he', 'PRP'), ('is', 'VBZ'), ('the', 'DT'), ('founder', 'NN'), ('of', 'IN')], Tree('PERSON', [('Marcohard', 'NNP')])]]
>>> len(tree2semi_rel(chunked)) > 2
True
>>> semi_rel2reldict(tree2semi_rel(chunked))
[defaultdict(<type 'str'>, {'lcon': '', 'untagged_filler': 'is the cofounder of', 'filler': 'is/VBZ the/DT cofounder/NN of/IN', 'objsym': 'microsoft', 'objclass': 'ORGANIZATION', 'objtext': 'Microsoft/NNP', 'subjsym': 'tom', 'subjclass': 'PERSON', 'rcon': 'and/CC now/RB he/PRP is/VBZ the/DT', 'subjtext': 'Tom/NNP'})]

But that only confirms that extract_rel can't extract when tree2semi_rel returns pairs of < 2. What happens if we remove that condition of while len(pairs) > 2?

Why can't we do while len(pairs) > 1?

If we look closer into the code, we see the last line of populating the reldict, https://github.com/nltk/nltk/blob/develop/nltk/sem/relextract.py#L169:

reldict['rcon'] = _join(pairs[2][0][:window])

It tries to access a 3rd element of the pairs and if the length of the pairs is 2, you'll get an IndexError.

So what happens if we remove that rcon key and simply change it to while len(pairs) >= 2?

To do that we have to override the semi_rel2redict() function:

>>> from nltk.sem.relextract import _join, list2sym
>>> from collections import defaultdict
>>> def semi_rel2reldict(pairs, window=5, trace=False):
...     """
...     Converts the pairs generated by ``tree2semi_rel`` into a 'reldict': a dictionary which
...     stores information about the subject and object NEs plus the filler between them.
...     Additionally, a left and right context of length =< window are captured (within
...     a given input sentence).
...     :param pairs: a pair of list(str) and ``Tree``, as generated by
...     :param window: a threshold for the number of items to include in the left and right context
...     :type window: int
...     :return: 'relation' dictionaries whose keys are 'lcon', 'subjclass', 'subjtext', 'subjsym', 'filler', objclass', objtext', 'objsym' and 'rcon'
...     :rtype: list(defaultdict)
...     """
...     result = []
...     while len(pairs) >= 2:
...         reldict = defaultdict(str)
...         reldict['lcon'] = _join(pairs[0][0][-window:])
...         reldict['subjclass'] = pairs[0][1].label()
...         reldict['subjtext'] = _join(pairs[0][1].leaves())
...         reldict['subjsym'] = list2sym(pairs[0][1].leaves())
...         reldict['filler'] = _join(pairs[1][0])
...         reldict['untagged_filler'] = _join(pairs[1][0], untag=True)
...         reldict['objclass'] = pairs[1][1].label()
...         reldict['objtext'] = _join(pairs[1][1].leaves())
...         reldict['objsym'] = list2sym(pairs[1][1].leaves())
...         reldict['rcon'] = []
...         if trace:
...             print("(%s(%s, %s)" % (reldict['untagged_filler'], reldict['subjclass'], reldict['objclass']))
...         result.append(reldict)
...         pairs = pairs[1:]
...     return result
... 
>>> text = "Tom is the cofounder of Microsoft"
>>> chunked = ne_chunk(pos_tag(word_tokenize(text)))
>>> tree2semi_rel(chunked)
[[[], Tree('PERSON', [('Tom', 'NNP')])], [[('is', 'VBZ'), ('the', 'DT'), ('cofounder', 'NN'), ('of', 'IN')], Tree('ORGANIZATION', [('Microsoft', 'NNP')])]]
>>> semi_rel2reldict(tree2semi_rel(chunked))
[defaultdict(<type 'str'>, {'lcon': '', 'untagged_filler': 'is the cofounder of', 'filler': 'is/VBZ the/DT cofounder/NN of/IN', 'objsym': 'microsoft', 'objclass': 'ORGANIZATION', 'objtext': 'Microsoft/NNP', 'subjsym': 'tom', 'subjclass': 'PERSON', 'rcon': [], 'subjtext': 'Tom/NNP'})]

Ah! It works but there's still a 4th step in extract_rels().

4. It performs a filter of the reldict given the regex you have provided to the pattern parameter, https://github.com/nltk/nltk/blob/develop/nltk/sem/relextract.py#L222:

relfilter = lambda x: (x['subjclass'] == subjclass and
                       len(x['filler'].split()) <= window and
                       pattern.match(x['filler']) and
                       x['objclass'] == objclass)

Now let's try it with the hacked version of semi_rel2reldict:

>>> text = "Tom is the cofounder of Microsoft"
>>> chunked = ne_chunk(pos_tag(word_tokenize(text)))
>>> tree2semi_rel(chunked)
[[[], Tree('PERSON', [('Tom', 'NNP')])], [[('is', 'VBZ'), ('the', 'DT'), ('cofounder', 'NN'), ('of', 'IN')], Tree('ORGANIZATION', [('Microsoft', 'NNP')])]]
>>> semi_rel2reldict(tree2semi_rel(chunked))
[defaultdict(<type 'str'>, {'lcon': '', 'untagged_filler': 'is the cofounder of', 'filler': 'is/VBZ the/DT cofounder/NN of/IN', 'objsym': 'microsoft', 'objclass': 'ORGANIZATION', 'objtext': 'Microsoft/NNP', 'subjsym': 'tom', 'subjclass': 'PERSON', 'rcon': [], 'subjtext': 'Tom/NNP'})]
>>> 
>>> pattern = re.compile(r'.*\bof\b.*')
>>> reldicts = semi_rel2reldict(tree2semi_rel(chunked))
>>> relfilter = lambda x: (x['subjclass'] == subjclass and
...                            len(x['filler'].split()) <= window and
...                            pattern.match(x['filler']) and
...                            x['objclass'] == objclass)
>>> relfilter
<function <lambda> at 0x112e591b8>
>>> subjclass = 'PERSON'
>>> objclass = 'ORGANIZATION'
>>> window = 5
>>> list(filter(relfilter, reldicts))
[defaultdict(<type 'str'>, {'lcon': '', 'untagged_filler': 'is the cofounder of', 'filler': 'is/VBZ the/DT cofounder/NN of/IN', 'objsym': 'microsoft', 'objclass': 'ORGANIZATION', 'objtext': 'Microsoft/NNP', 'subjsym': 'tom', 'subjclass': 'PERSON', 'rcon': [], 'subjtext': 'Tom/NNP'})]

It works! Now let's see it in tuple form:

>>> from nltk.sem.relextract import rtuple
>>> rels = list(filter(relfilter, reldicts))
>>> for rel in rels:
...     print rtuple(rel)
... 
[PER: 'Tom/NNP'] 'is/VBZ the/DT cofounder/NN of/IN' [ORG: 'Microsoft/NNP']