chunker_training.txt

====================
Treebank Chunk Trees
====================

>>> from nltk.corpus import treebank_chunk
>>> treebank_chunk.chunked_sents()[0]
Tree('S', [Tree('NP', [('Pierre', 'NNP'), ('Vinken', 'NNP')]), (',', ','), Tree('NP', [('61', 'CD'), ('years', 'NNS')]), ('old', 'JJ'), (',', ','), ('will', 'MD'), ('join', 'VB'), Tree('NP', [('the', 'DT'), ('board', 'NN')]), ('as', 'IN'), Tree('NP', [('a', 'DT'), ('nonexecutive', 'JJ'), ('director', 'NN'), ('Nov.', 'NNP'), ('29', 'CD')]), ('.', '.')])

>>> from nltk import chunk
>>> import nltk.data
>>> chunker = nltk.data.load(chunk._MULTICLASS_NE_CHUNKER)
>>> score = chunker.evaluate(treebank_chunk.chunked_sents())
>>> print score
ChunkParse score:
    IOB Accuracy:  45.4%
    Precision:      0.0%
    Recall:         0.0%
    F-Measure:      0.0%

# treebank_chunk doesn't have named entities, so score is not really accurate

>>> score.accuracy()
0.4536624203821656
>>> score.precision()
0.0
>>> score.recall()
0.0
>>> len(score.correct())
24667
>>> len(score.incorrect())
5659
>>> len(score.missed())
24667

======================
Simplified Chunk Trees
======================

>>> import chunk_parser
>>> simple_chunks = [chunk_parser.simplify_chunk(c) for c in treebank_chunk.chunked_sents()]
>>> simple_chunks[0]
Tree('S', [Tree('NP', [('Pierre', 'NP'), ('Vinken', 'NP')]), (',', ','), Tree('NP', [('61', 'NUM'), ('years', 'N')]), ('old', 'ADJ'), (',', ','), ('will', 'MOD'), ('join', 'V'), Tree('NP', [('the', 'DET'), ('board', 'N')]), ('as', 'P'), Tree('NP', [('a', 'DET'), ('nonexecutive', 'ADJ'), ('director', 'N'), ('Nov.', 'NP'), ('29', 'NUM')]), ('.', '.')])

>>> chunker = chunk_parser.ChunkParser(simple_chunks)
>>> score = chunker.evaluate(simple_chunks)
>>> print score
ChunkParse score:
    IOB Accuracy:  97.6%
    Precision:     93.0%
    Recall:        94.4%
    F-Measure:     93.7%

# score is self-accurate and doesn't reflect real-world usage

>>> import os, os.path, pickle
>>> path = os.path.expanduser('~/nltk_data/chunkers/')
>>> os.makedirs(path)
>>> f = open(os.path.join(path, 'chunk_parser.pickle'), 'wb')
>>> pickle.dump(chunker, f)
>>> f.close()

>>> iob_tagged = chunker._parse_to_tagged(simple_chunks[0])
>>> iob_tagged
[(('Pierre', 'NP'), 'B-NP'), (('Vinken', 'NP'), 'I-NP'), ((',', ','), 'O'), (('61', 'NUM'), 'B-NP'), (('years', 'N'), 'I-NP'), (('old', 'ADJ'), 'O'), ((',', ','), 'O'), (('will', 'MOD'), 'O'), (('join', 'V'), 'O'), (('the', 'DET'), 'B-NP'), (('board', 'N'), 'I-NP'), (('as', 'P'), 'O'), (('a', 'DET'), 'B-NP'), (('nonexecutive', 'ADJ'), 'I-NP'), (('director', 'N'), 'I-NP'), (('Nov.', 'NP'), 'I-NP'), (('29', 'NUM'), 'I-NP'), (('.', '.'), 'O')]
>>> chunker._tagged_to_parse(iob_tagged) == simple_chunks[0]
True
>>> iob_tagged[0]
(('Pierre', 'NP'), 'B-NP')
>>> untagged, tags = zip(*iob_tagged)
>>> chunker._tagger._feature_detector(untagged, 0, [])
{'nextpos': 'np', 'pos+prevtag': 'NP+None', 'nextword': 'vinken', 'word': 'Pierre', 'prefix3': 'pie', 'wordlen': 6, 'prevpos': None, 'pos': 'NP', 'prevtag': None, 'prevword': None, 'shape': 'upcase', 'bias': True, 'en-wordlist': False, 'shape+prevtag': 'None+None', 'suffix3': 'rre', 'word+nextpos': 'pierre+np'}