named_entity.py

# Natural Language Toolkit: Chunk parsing API
#
# Copyright (C) 2001-2019 NLTK Project
# Author: Edward Loper <edloper@gmail.com>
# URL: <http://nltk.org/>
# For license information, see LICENSE.TXT

"""
Named entity chunker
"""
from __future__ import print_function
from __future__ import unicode_literals

import os, re, pickle
from xml.etree import ElementTree as ET

from nltk.tag import ClassifierBasedTagger, pos_tag

try:
    from nltk.classify import MaxentClassifier
except ImportError:
    pass

from nltk.tree import Tree
from nltk.tokenize import word_tokenize
from nltk.data import find

from nltk.chunk.api import ChunkParserI
from nltk.chunk.util import ChunkScore


class NEChunkParserTagger(ClassifierBasedTagger):
    """
    The IOB tagger used by the chunk parser.
    """

    def __init__(self, train):
        ClassifierBasedTagger.__init__(
            self, train=train, classifier_builder=self._classifier_builder
        )

    def _classifier_builder(self, train):
        return MaxentClassifier.train(
            train, algorithm='megam', gaussian_prior_sigma=1, trace=2
        )

    def _english_wordlist(self):
        try:
            wl = self._en_wordlist
        except AttributeError:
            from nltk.corpus import words

            self._en_wordlist = set(words.words('en-basic'))
            wl = self._en_wordlist
        return wl

    def _feature_detector(self, tokens, index, history):
        word = tokens[index][0]
        pos = simplify_pos(tokens[index][1])
        if index == 0:
            prevword = prevprevword = None
            prevpos = prevprevpos = None
            prevshape = prevtag = prevprevtag = None
        elif index == 1:
            prevword = tokens[index - 1][0].lower()
            prevprevword = None
            prevpos = simplify_pos(tokens[index - 1][1])
            prevprevpos = None
            prevtag = history[index - 1][0]
            prevshape = prevprevtag = None
        else:
            prevword = tokens[index - 1][0].lower()
            prevprevword = tokens[index - 2][0].lower()
            prevpos = simplify_pos(tokens[index - 1][1])
            prevprevpos = simplify_pos(tokens[index - 2][1])
            prevtag = history[index - 1]
            prevprevtag = history[index - 2]
            prevshape = shape(prevword)
        if index == len(tokens) - 1:
            nextword = nextnextword = None
            nextpos = nextnextpos = None
        elif index == len(tokens) - 2:
            nextword = tokens[index + 1][0].lower()
            nextpos = tokens[index + 1][1].lower()
            nextnextword = None
            nextnextpos = None
        else:
            nextword = tokens[index + 1][0].lower()
            nextpos = tokens[index + 1][1].lower()
            nextnextword = tokens[index + 2][0].lower()
            nextnextpos = tokens[index + 2][1].lower()

        # 89.6
        features = {
            'bias': True,
            'shape': shape(word),
            'wordlen': len(word),
            'prefix3': word[:3].lower(),
            'suffix3': word[-3:].lower(),
            'pos': pos,
            'word': word,
            'en-wordlist': (word in self._english_wordlist()),
            'prevtag': prevtag,
            'prevpos': prevpos,
            'nextpos': nextpos,
            'prevword': prevword,
            'nextword': nextword,
            'word+nextpos': '{0}+{1}'.format(word.lower(), nextpos),
            'pos+prevtag': '{0}+{1}'.format(pos, prevtag),
            'shape+prevtag': '{0}+{1}'.format(prevshape, prevtag),
        }

        return features


class NEChunkParser(ChunkParserI):
    """
    Expected input: list of pos-tagged words
    """

    def __init__(self, train):
        self._train(train)

    def parse(self, tokens):
        """
        Each token should be a pos-tagged word
        """
        tagged = self._tagger.tag(tokens)
        tree = self._tagged_to_parse(tagged)
        return tree

    def _train(self, corpus):
        # Convert to tagged sequence
        corpus = [self._parse_to_tagged(s) for s in corpus]

        self._tagger = NEChunkParserTagger(train=corpus)

    def _tagged_to_parse(self, tagged_tokens):
        """
        Convert a list of tagged tokens to a chunk-parse tree.
        """
        sent = Tree('S', [])

        for (tok, tag) in tagged_tokens:
            if tag == 'O':
                sent.append(tok)
            elif tag.startswith('B-'):
                sent.append(Tree(tag[2:], [tok]))
            elif tag.startswith('I-'):
                if sent and isinstance(sent[-1], Tree) and sent[-1].label() == tag[2:]:
                    sent[-1].append(tok)
                else:
                    sent.append(Tree(tag[2:], [tok]))
        return sent

    @staticmethod
    def _parse_to_tagged(sent):
        """
        Convert a chunk-parse tree to a list of tagged tokens.
        """
        toks = []
        for child in sent:
            if isinstance(child, Tree):
                if len(child) == 0:
                    print("Warning -- empty chunk in sentence")
                    continue
                toks.append((child[0], 'B-{0}'.format(child.label())))
                for tok in child[1:]:
                    toks.append((tok, 'I-{0}'.format(child.label())))
            else:
                toks.append((child, 'O'))
        return toks


def shape(word):
    if re.match('[0-9]+(\.[0-9]*)?|[0-9]*\.[0-9]+$', word, re.UNICODE):
        return 'number'
    elif re.match('\W+$', word, re.UNICODE):
        return 'punct'
    elif re.match('\w+$', word, re.UNICODE):
        if word.istitle():
            return 'upcase'
        elif word.islower():
            return 'downcase'
        else:
            return 'mixedcase'
    else:
        return 'other'


def simplify_pos(s):
    if s.startswith('V'):
        return "V"
    else:
        return s.split('-')[0]


def postag_tree(tree):
    # Part-of-speech tagging.
    words = tree.leaves()
    tag_iter = (pos for (word, pos) in pos_tag(words))
    newtree = Tree('S', [])
    for child in tree:
        if isinstance(child, Tree):
            newtree.append(Tree(child.label(), []))
            for subchild in child:
                newtree[-1].append((subchild, next(tag_iter)))
        else:
            newtree.append((child, next(tag_iter)))
    return newtree


def load_ace_data(roots, fmt='binary', skip_bnews=True):
    for root in roots:
        for root, dirs, files in os.walk(root):
            if root.endswith('bnews') and skip_bnews:
                continue
            for f in files:
                if f.endswith('.sgm'):
                    for sent in load_ace_file(os.path.join(root, f), fmt):
                        yield sent


def load_ace_file(textfile, fmt):
    print('  - {0}'.format(os.path.split(textfile)[1]))
    annfile = textfile + '.tmx.rdc.xml'

    # Read the xml file, and get a list of entities
    entities = []
    with open(annfile, 'r') as infile:
        xml = ET.parse(infile).getroot()
    for entity in xml.findall('document/entity'):
        typ = entity.find('entity_type').text
        for mention in entity.findall('entity_mention'):
            if mention.get('TYPE') != 'NAME':
                continue  # only NEs
            s = int(mention.find('head/charseq/start').text)
            e = int(mention.find('head/charseq/end').text) + 1
            entities.append((s, e, typ))

    # Read the text file, and mark the entities.
    with open(textfile, 'r') as infile:
        text = infile.read()

    # Strip XML tags, since they don't count towards the indices
    text = re.sub('<(?!/?TEXT)[^>]+>', '', text)

    # Blank out anything before/after <TEXT>
    def subfunc(m):
        return ' ' * (m.end() - m.start() - 6)

    text = re.sub('[\s\S]*<TEXT>', subfunc, text)
    text = re.sub('</TEXT>[\s\S]*', '', text)

    # Simplify quotes
    text = re.sub("``", ' "', text)
    text = re.sub("''", '" ', text)

    entity_types = set(typ for (s, e, typ) in entities)

    # Binary distinction (NE or not NE)
    if fmt == 'binary':
        i = 0
        toks = Tree('S', [])
        for (s, e, typ) in sorted(entities):
            if s < i:
                s = i  # Overlapping!  Deal with this better?
            if e <= s:
                continue
            toks.extend(word_tokenize(text[i:s]))
            toks.append(Tree('NE', text[s:e].split()))
            i = e
        toks.extend(word_tokenize(text[i:]))
        yield toks

    # Multiclass distinction (NE type)
    elif fmt == 'multiclass':
        i = 0
        toks = Tree('S', [])
        for (s, e, typ) in sorted(entities):
            if s < i:
                s = i  # Overlapping!  Deal with this better?
            if e <= s:
                continue
            toks.extend(word_tokenize(text[i:s]))
            toks.append(Tree(typ, text[s:e].split()))
            i = e
        toks.extend(word_tokenize(text[i:]))
        yield toks

    else:
        raise ValueError('bad fmt value')


# This probably belongs in a more general-purpose location (as does
# the parse_to_tagged function).
def cmp_chunks(correct, guessed):
    correct = NEChunkParser._parse_to_tagged(correct)
    guessed = NEChunkParser._parse_to_tagged(guessed)
    ellipsis = False
    for (w, ct), (w, gt) in zip(correct, guessed):
        if ct == gt == 'O':
            if not ellipsis:
                print("  {:15} {:15} {2}".format(ct, gt, w))
                print('  {:15} {:15} {2}'.format('...', '...', '...'))
                ellipsis = True
        else:
            ellipsis = False
            print("  {:15} {:15} {2}".format(ct, gt, w))


def build_model(fmt='binary'):
    print('Loading training data...')
    train_paths = [
        find('corpora/ace_data/ace.dev'),
        find('corpora/ace_data/ace.heldout'),
        find('corpora/ace_data/bbn.dev'),
        find('corpora/ace_data/muc.dev'),
    ]
    train_trees = load_ace_data(train_paths, fmt)
    train_data = [postag_tree(t) for t in train_trees]
    print('Training...')
    cp = NEChunkParser(train_data)
    del train_data

    print('Loading eval data...')
    eval_paths = [find('corpora/ace_data/ace.eval')]
    eval_trees = load_ace_data(eval_paths, fmt)
    eval_data = [postag_tree(t) for t in eval_trees]

    print('Evaluating...')
    chunkscore = ChunkScore()
    for i, correct in enumerate(eval_data):
        guess = cp.parse(correct.leaves())
        chunkscore.score(correct, guess)
        if i < 3:
            cmp_chunks(correct, guess)
    print(chunkscore)

    outfilename = '/tmp/ne_chunker_{0}.pickle'.format(fmt)
    print('Saving chunker to {0}...'.format(outfilename))

    with open(outfilename, 'wb') as outfile:
        pickle.dump(cp, outfile, -1)

    return cp


if __name__ == '__main__':
    # Make sure that the pickled object has the right class name:
    from nltk.chunk.named_entity import build_model

    build_model('binary')
    build_model('multiclass')