conll05.py 8.7 KB
Newer Older
D
dangqingqing 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
D
dangqingqing 已提交
14
"""
Q
qijun 已提交
15
Conll05 dataset.
Q
qijun 已提交
16 17 18 19 20
Paddle semantic role labeling Book and demo use this dataset as an example.
Because Conll05 is not free in public, the default downloaded URL is test set
of Conll05 (which is public). Users can change URL and MD5 to their Conll
dataset. And a pre-trained word vector model based on Wikipedia corpus is used
to initialize SRL model.
D
dangqingqing 已提交
21 22
"""

Q
qijun 已提交
23 24 25
import tarfile
import gzip
import itertools
R
root 已提交
26
import paddle.v2.dataset.common
Q
qijun 已提交
27

Y
Your Name 已提交
28
__all__ = ['test, get_dict', 'get_embedding', 'convert']
Y
Yu Yang 已提交
29

D
dangqingqing 已提交
30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53
DATA_URL = 'http://www.cs.upc.edu/~srlconll/conll05st-tests.tar.gz'
DATA_MD5 = '387719152ae52d60422c016e92a742fc'
WORDDICT_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/wordDict.txt'
WORDDICT_MD5 = 'ea7fb7d4c75cc6254716f0177a506baa'
VERBDICT_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/verbDict.txt'
VERBDICT_MD5 = '0d2977293bbb6cbefab5b0f97db1e77c'
TRGDICT_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/targetDict.txt'
TRGDICT_MD5 = 'd8c7f03ceb5fc2e5a0fa7503a4353751'
EMB_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/emb'
EMB_MD5 = 'bf436eb0faa1f6f9103017f8be57cdb7'

UNK_IDX = 0


def load_dict(filename):
    d = dict()
    with open(filename, 'r') as f:
        for i, line in enumerate(f):
            d[line.strip()] = i
    return d


def corpus_reader(data_path, words_name, props_name):
    """
54
    Read one corpus. It returns an iterator. Each element of
D
dangqingqing 已提交
55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106
    this iterator is a tuple including sentence and labels. The sentence is
    consist of a list of word IDs. The labels include a list of label IDs.
    :return: a iterator of data.
    :rtype: iterator
    """

    def reader():
        tf = tarfile.open(data_path)
        wf = tf.extractfile(words_name)
        pf = tf.extractfile(props_name)
        with gzip.GzipFile(fileobj=wf) as words_file, gzip.GzipFile(
                fileobj=pf) as props_file:
            sentences = []
            labels = []
            one_seg = []
            for word, label in itertools.izip(words_file, props_file):
                word = word.strip()
                label = label.strip().split()

                if len(label) == 0:  # end of sentence
                    for i in xrange(len(one_seg[0])):
                        a_kind_lable = [x[i] for x in one_seg]
                        labels.append(a_kind_lable)

                    if len(labels) >= 1:
                        verb_list = []
                        for x in labels[0]:
                            if x != '-':
                                verb_list.append(x)

                        for i, lbl in enumerate(labels[1:]):
                            cur_tag = 'O'
                            is_in_bracket = False
                            lbl_seq = []
                            verb_word = ''
                            for l in lbl:
                                if l == '*' and is_in_bracket == False:
                                    lbl_seq.append('O')
                                elif l == '*' and is_in_bracket == True:
                                    lbl_seq.append('I-' + cur_tag)
                                elif l == '*)':
                                    lbl_seq.append('I-' + cur_tag)
                                    is_in_bracket = False
                                elif l.find('(') != -1 and l.find(')') != -1:
                                    cur_tag = l[1:l.find('*')]
                                    lbl_seq.append('B-' + cur_tag)
                                    is_in_bracket = False
                                elif l.find('(') != -1 and l.find(')') == -1:
                                    cur_tag = l[1:l.find('*')]
                                    lbl_seq.append('B-' + cur_tag)
                                    is_in_bracket = True
                                else:
107 108
                                    raise RuntimeError('Unexpected label: %s' %
                                                       l)
D
dangqingqing 已提交
109 110 111 112 113 114 115 116 117 118

                            yield sentences, verb_list[i], lbl_seq

                    sentences = []
                    labels = []
                    one_seg = []
                else:
                    sentences.append(word)
                    one_seg.append(label)

119 120 121 122
        pf.close()
        wf.close()
        tf.close()

D
dangqingqing 已提交
123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171
    return reader


def reader_creator(corpus_reader,
                   word_dict=None,
                   predicate_dict=None,
                   label_dict=None):
    def reader():
        for sentence, predicate, labels in corpus_reader():

            sen_len = len(sentence)

            verb_index = labels.index('B-V')
            mark = [0] * len(labels)
            if verb_index > 0:
                mark[verb_index - 1] = 1
                ctx_n1 = sentence[verb_index - 1]
            else:
                ctx_n1 = 'bos'

            if verb_index > 1:
                mark[verb_index - 2] = 1
                ctx_n2 = sentence[verb_index - 2]
            else:
                ctx_n2 = 'bos'

            mark[verb_index] = 1
            ctx_0 = sentence[verb_index]

            if verb_index < len(labels) - 1:
                mark[verb_index + 1] = 1
                ctx_p1 = sentence[verb_index + 1]
            else:
                ctx_p1 = 'eos'

            if verb_index < len(labels) - 2:
                mark[verb_index + 2] = 1
                ctx_p2 = sentence[verb_index + 2]
            else:
                ctx_p2 = 'eos'

            word_idx = [word_dict.get(w, UNK_IDX) for w in sentence]

            ctx_n2_idx = [word_dict.get(ctx_n2, UNK_IDX)] * sen_len
            ctx_n1_idx = [word_dict.get(ctx_n1, UNK_IDX)] * sen_len
            ctx_0_idx = [word_dict.get(ctx_0, UNK_IDX)] * sen_len
            ctx_p1_idx = [word_dict.get(ctx_p1, UNK_IDX)] * sen_len
            ctx_p2_idx = [word_dict.get(ctx_p2, UNK_IDX)] * sen_len

D
dangqingqing 已提交
172
            pred_idx = [predicate_dict.get(predicate)] * sen_len
D
dangqingqing 已提交
173 174
            label_idx = [label_dict.get(w) for w in labels]

D
dangqingqing 已提交
175 176
            yield word_idx, ctx_n2_idx, ctx_n1_idx, \
              ctx_0_idx, ctx_p1_idx, ctx_p2_idx, pred_idx, mark, label_idx
D
dangqingqing 已提交
177

D
update  
dangqingqing 已提交
178
    return reader
D
dangqingqing 已提交
179 180 181


def get_dict():
Q
qijun 已提交
182 183 184
    """
    Get the word, verb and label dictionary of Wikipedia corpus.
    """
R
root 已提交
185 186 187 188 189 190 191 192 193
    word_dict = load_dict(
        paddle.v2.dataset.common.download(WORDDICT_URL, 'conll05st',
                                          WORDDICT_MD5))
    verb_dict = load_dict(
        paddle.v2.dataset.common.download(VERBDICT_URL, 'conll05st',
                                          VERBDICT_MD5))
    label_dict = load_dict(
        paddle.v2.dataset.common.download(TRGDICT_URL, 'conll05st',
                                          TRGDICT_MD5))
D
dangqingqing 已提交
194 195 196 197
    return word_dict, verb_dict, label_dict


def get_embedding():
Q
qijun 已提交
198 199 200
    """
    Get the trained word vector based on Wikipedia corpus.
    """
R
root 已提交
201
    return paddle.v2.dataset.common.download(EMB_URL, 'conll05st', EMB_MD5)
D
dangqingqing 已提交
202 203 204


def test():
Q
qijun 已提交
205 206 207
    """
    Conll05 test set creator.

Q
qijun 已提交
208
    Because the training dataset is not free, the test dataset is used for
Q
qijun 已提交
209 210 211
    training. It returns a reader creator, each sample in the reader is nine
    features, including sentence sequence, predicate, predicate context,
    predicate context flag and tagged sequence.
Q
qijun 已提交
212

Q
qijun 已提交
213
    :return: Training reader creator
Q
qijun 已提交
214 215
    :rtype: callable
    """
D
dangqingqing 已提交
216 217
    word_dict, verb_dict, label_dict = get_dict()
    reader = corpus_reader(
R
root 已提交
218
        paddle.v2.dataset.common.download(DATA_URL, 'conll05st', DATA_MD5),
D
dangqingqing 已提交
219 220 221
        words_name='conll05st-release/test.wsj/words/test.wsj.words.gz',
        props_name='conll05st-release/test.wsj/props/test.wsj.props.gz')
    return reader_creator(reader, word_dict, verb_dict, label_dict)
Y
Yancey1989 已提交
222 223


224
def fetch():
R
root 已提交
225 226 227 228 229 230 231
    paddle.v2.dataset.common.download(WORDDICT_URL, 'conll05st', WORDDICT_MD5)
    paddle.v2.dataset.common.download(VERBDICT_URL, 'conll05st', VERBDICT_MD5)
    paddle.v2.dataset.common.download(TRGDICT_URL, 'conll05st', TRGDICT_MD5)
    paddle.v2.dataset.common.download(EMB_URL, 'conll05st', EMB_MD5)
    paddle.v2.dataset.common.download(DATA_URL, 'conll05st', DATA_MD5)


Y
Your Name 已提交
232
def convert(path):
R
root 已提交
233 234 235
    """
    Converts dataset to recordio format
    """
236 237
    paddle.v2.dataset.common.convert(path, test(), 1000, "conl105_train")
    paddle.v2.dataset.common.convert(path, test(), 1000, "conl105_test")