File size: 7,018 Bytes
a4e3fbc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
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
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
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
import json
import os
import tarfile
import zipfile
import gzip
import requests
from itertools import chain
from glob import glob
import gdown

from datasets import load_dataset

k = 10  # the 3rd level negative-distance ranking
m = 5  # the 3rd level negative-distance ranking
top_n = 10  # threshold of positive pairs in the 1st and 2nd relation


def wget(url, cache_dir: str = './cache', gdrive_filename: str = None):
    """ wget and uncompress data_iterator """
    os.makedirs(cache_dir, exist_ok=True)
    if url.startswith('https://drive.google.com'):
        assert gdrive_filename is not None, 'please provide fileaname for gdrive download'
        gdown.download(url, f'{cache_dir}/{gdrive_filename}', quiet=False)
        filename = gdrive_filename
    else:
        filename = os.path.basename(url)
    with open(f'{cache_dir}/{filename}', "wb") as f:
        r = requests.get(url)
        f.write(r.content)
    path = f'{cache_dir}/{filename}'

    if path.endswith('.tar.gz') or path.endswith('.tgz') or path.endswith('.tar'):
        if path.endswith('.tar'):
            tar = tarfile.open(path)
        else:
            tar = tarfile.open(path, "r:gz")
        tar.extractall(cache_dir)
        tar.close()
        os.remove(path)
    elif path.endswith('.zip'):
        with zipfile.ZipFile(path, 'r') as zip_ref:
            zip_ref.extractall(cache_dir)
        os.remove(path)
    elif path.endswith('.gz'):
        with gzip.open(path, 'rb') as f:
            with open(path.replace('.gz', ''), 'wb') as f_write:
                f_write.write(f.read())
        os.remove(path)


def get_training_data():
    """ Get RelBERT training data

    Returns
    -------
    pairs: dictionary of list (positive pairs, negative pairs)
    {'1b': [[0.6, ('office', 'desk'), ..], [[-0.1, ('aaa', 'bbb'), ...]]
    """
    cache_dir = 'cache'
    os.makedirs(cache_dir, exist_ok=True)
    remove_relation = None
    path_answer = f'{cache_dir}/Phase2Answers'
    path_scale = f'{cache_dir}/Phase2AnswersScaled'
    url = 'https://drive.google.com/u/0/uc?id=0BzcZKTSeYL8VYWtHVmxUR3FyUmc&export=download'
    filename = 'SemEval-2012-Platinum-Ratings.tar.gz'
    if not (os.path.exists(path_scale) and os.path.exists(path_answer)):
        wget(url, gdrive_filename=filename, cache_dir=cache_dir)
    files_answer = [os.path.basename(i) for i in glob(f'{path_answer}/*.txt')]
    files_scale = [os.path.basename(i) for i in glob(f'{path_scale}/*.txt')]
    assert files_answer == files_scale, f'files are not matched: {files_scale} vs {files_answer}'
    positives = {}
    negatives = {}
    positives_limit = {}
    all_relation_type = {}
    # score_range = [90.0, 88.7]  # the absolute value of max/min prototypicality rating
    for i in files_scale:
        relation_id = i.split('-')[-1].replace('.txt', '')
        if remove_relation and int(relation_id[:-1]) in remove_relation:
            continue
        with open(f'{path_answer}/{i}', 'r') as f:
            lines_answer = [_l.replace('"', '').split('\t') for _l in f.read().split('\n')
                            if not _l.startswith('#') and len(_l)]
            relation_type = list(set(list(zip(*lines_answer))[-1]))
            assert len(relation_type) == 1, relation_type
            relation_type = relation_type[0]
        with open(f'{path_scale}/{i}', 'r') as f:
            # list of tuple [score, ("a", "b")]
            scales = [[float(_l[:5]), _l[6:].replace('"', '')] for _l in f.read().split('\n')
                      if not _l.startswith('#') and len(_l)]
            scales = sorted(scales, key=lambda _x: _x[0])
            # positive pairs are in the reverse order of prototypicality score
            positive_pairs = [[s, tuple(p.split(':'))] for s, p in filter(lambda _x: _x[0] > 0, scales)]
            positive_pairs = sorted(positive_pairs, key=lambda x:  x[0], reverse=True)
            positives[relation_id] = list(list(zip(*positive_pairs))[1])
            positives_limit[relation_id] = list(list(zip(*positive_pairs[:min(top_n, len(positive_pairs))]))[1])
            negatives[relation_id] = [tuple(p.split(':')) for s, p in filter(lambda _x: _x[0] < 0, scales)]
        all_relation_type[relation_id] = relation_type
    parent = list(set([i[:-1] for i in all_relation_type.keys()]))

    # 1st level relation contrast (among parent relations)
    relation_pairs_1st = []
    relation_pairs_1st_validation = []
    for p in parent:
        child_positive = list(filter(lambda x: x.startswith(p), list(all_relation_type.keys())))
        child_negative = list(filter(lambda x: not x.startswith(p), list(all_relation_type.keys())))
        positive_pairs = []
        negative_pairs = []
        for c in child_positive:
            positive_pairs += positives_limit[c]
        for c in child_negative:
            negative_pairs += positives_limit[c]

        relation_pairs_1st += [{
            "positives": positive_pairs, "negatives": negative_pairs, "relation_type": p, "level": "parent"
        }]

    # 2nd level relation contrast (among child relations) & 3rd level relation contrast (within child relations)
    relation_pairs_2nd = []
    relation_pairs_2nd_validation = []
    for p in all_relation_type.keys():
        positive_pairs = positives_limit[p]
        negative_pairs = []
        for n in all_relation_type.keys():
            if p == n:
                continue
            negative_pairs += positives[n]

        relation_pairs_2nd += [{
            "positives": positive_pairs, "negatives": negative_pairs, "relation_type": p, "level": "child"
        }]

    relation_pairs_3rd = []
    for p in all_relation_type.keys():
        positive_pairs = positives[p]
        negative_pairs = positive_pairs + negatives[p]
        for n, anchor in enumerate(positive_pairs):
            if n > m:
                continue
            for _n, posi in enumerate(positive_pairs):
                if n < _n and len(negative_pairs) > _n + k:
                    relation_pairs_3rd += [{
                        "positives": [(anchor, posi)],
                        "negatives": [(anchor, neg) for neg in negative_pairs[_n+k:]],
                        "relation_type": p,
                        "level": "child_prototypical"
                    }]

    train = relation_pairs_1st + relation_pairs_2nd + relation_pairs_3rd

    # conceptnet as the validation set
    cn = load_dataset('relbert/conceptnet_high_confidence_v2')
    valid = list(chain(*cn.values()))
    for i in valid:
        i['level'] = 'N/A'
    return train, valid


if __name__ == '__main__':
    data_train, data_validation = get_training_data()
    print(f"- training data     : {len(data_train)}")
    print(f"- validation data   : {len(data_validation)}")
    with open('dataset/train.jsonl', 'w') as f_writer:
        f_writer.write('\n'.join([json.dumps(i) for i in data_train]))
    with open('dataset/valid.jsonl', 'w') as f_writer:
        f_writer.write('\n'.join([json.dumps(i) for i in data_validation]))