You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

181 lines
6.1 KiB
Python

# Natural Language Toolkit: RTE Classifier
#
# Copyright (C) 2001-2019 NLTK Project
# Author: Ewan Klein <ewan@inf.ed.ac.uk>
# URL: <http://nltk.org/>
# For license information, see LICENSE.TXT
"""
Simple classifier for RTE corpus.
It calculates the overlap in words and named entities between text and
hypothesis, and also whether there are words / named entities in the
hypothesis which fail to occur in the text, since this is an indicator that
the hypothesis is more informative than (i.e not entailed by) the text.
TO DO: better Named Entity classification
TO DO: add lemmatization
"""
from __future__ import print_function
from nltk.tokenize import RegexpTokenizer
from nltk.classify.util import accuracy, check_megam_config
from nltk.classify.maxent import MaxentClassifier
class RTEFeatureExtractor(object):
"""
This builds a bag of words for both the text and the hypothesis after
throwing away some stopwords, then calculates overlap and difference.
"""
def __init__(self, rtepair, stop=True, use_lemmatize=False):
"""
:param rtepair: a ``RTEPair`` from which features should be extracted
:param stop: if ``True``, stopwords are thrown away.
:type stop: bool
"""
self.stop = stop
self.stopwords = set(
[
'a',
'the',
'it',
'they',
'of',
'in',
'to',
'is',
'have',
'are',
'were',
'and',
'very',
'.',
',',
]
)
self.negwords = set(['no', 'not', 'never', 'failed', 'rejected', 'denied'])
# Try to tokenize so that abbreviations, monetary amounts, email
# addresses, URLs are single tokens.
tokenizer = RegexpTokenizer('[\w.@:/]+|\w+|\$[\d.]+')
# Get the set of word types for text and hypothesis
self.text_tokens = tokenizer.tokenize(rtepair.text)
self.hyp_tokens = tokenizer.tokenize(rtepair.hyp)
self.text_words = set(self.text_tokens)
self.hyp_words = set(self.hyp_tokens)
if use_lemmatize:
self.text_words = set(self._lemmatize(token) for token in self.text_tokens)
self.hyp_words = set(self._lemmatize(token) for token in self.hyp_tokens)
if self.stop:
self.text_words = self.text_words - self.stopwords
self.hyp_words = self.hyp_words - self.stopwords
self._overlap = self.hyp_words & self.text_words
self._hyp_extra = self.hyp_words - self.text_words
self._txt_extra = self.text_words - self.hyp_words
def overlap(self, toktype, debug=False):
"""
Compute the overlap between text and hypothesis.
:param toktype: distinguish Named Entities from ordinary words
:type toktype: 'ne' or 'word'
"""
ne_overlap = set(token for token in self._overlap if self._ne(token))
if toktype == 'ne':
if debug:
print("ne overlap", ne_overlap)
return ne_overlap
elif toktype == 'word':
if debug:
print("word overlap", self._overlap - ne_overlap)
return self._overlap - ne_overlap
else:
raise ValueError("Type not recognized:'%s'" % toktype)
def hyp_extra(self, toktype, debug=True):
"""
Compute the extraneous material in the hypothesis.
:param toktype: distinguish Named Entities from ordinary words
:type toktype: 'ne' or 'word'
"""
ne_extra = set(token for token in self._hyp_extra if self._ne(token))
if toktype == 'ne':
return ne_extra
elif toktype == 'word':
return self._hyp_extra - ne_extra
else:
raise ValueError("Type not recognized: '%s'" % toktype)
@staticmethod
def _ne(token):
"""
This just assumes that words in all caps or titles are
named entities.
:type token: str
"""
if token.istitle() or token.isupper():
return True
return False
@staticmethod
def _lemmatize(word):
"""
Use morphy from WordNet to find the base form of verbs.
"""
lemma = nltk.corpus.wordnet.morphy(word, pos=nltk.corpus.wordnet.VERB)
if lemma is not None:
return lemma
return word
def rte_features(rtepair):
extractor = RTEFeatureExtractor(rtepair)
features = {}
features['alwayson'] = True
features['word_overlap'] = len(extractor.overlap('word'))
features['word_hyp_extra'] = len(extractor.hyp_extra('word'))
features['ne_overlap'] = len(extractor.overlap('ne'))
features['ne_hyp_extra'] = len(extractor.hyp_extra('ne'))
features['neg_txt'] = len(extractor.negwords & extractor.text_words)
features['neg_hyp'] = len(extractor.negwords & extractor.hyp_words)
return features
def rte_featurize(rte_pairs):
return [(rte_features(pair), pair.value) for pair in rte_pairs]
def rte_classifier(algorithm):
from nltk.corpus import rte as rte_corpus
train_set = rte_corpus.pairs(['rte1_dev.xml', 'rte2_dev.xml', 'rte3_dev.xml'])
test_set = rte_corpus.pairs(['rte1_test.xml', 'rte2_test.xml', 'rte3_test.xml'])
featurized_train_set = rte_featurize(train_set)
featurized_test_set = rte_featurize(test_set)
# Train the classifier
print('Training classifier...')
if algorithm in ['megam', 'BFGS']: # MEGAM based algorithms.
# Ensure that MEGAM is configured first.
check_megam_config()
clf = lambda x: MaxentClassifier.train(featurized_train_set, algorithm)
elif algorithm in ['GIS', 'IIS']: # Use default GIS/IIS MaxEnt algorithm
clf = MaxentClassifier.train(featurized_train_set, algorithm)
else:
err_msg = str(
"RTEClassifier only supports these algorithms:\n "
"'megam', 'BFGS', 'GIS', 'IIS'.\n"
)
raise Exception(err_msg)
print('Testing classifier...')
acc = accuracy(clf, featurized_test_set)
print('Accuracy: %6.4f' % acc)
return clf