Newer
Older
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
import torch
from collections import Counter
import argparse
import os
import json
from tqdm import tqdm
from torch.nn.utils.rnn import pad_sequence
import pandas as pd
from model.Transformer import BaseTransformer
from itertools import chain
class fullDataset(torch.utils.data.Dataset):
def __init__(self, dir, tokenizer, split='train', model='crossencoder', kg='none', kg_token="[KG]", n_labels=2):
self.path = dir
if kg !='none':
df = pd.read_csv(self.path, index_col=0)[['Arg1', 'Arg2', 'rel', kg, 'split']]
else:
df = pd.read_csv(self.path, index_col=0)[['Arg1', 'Arg2', 'rel', 'split']]
df = df[df['split'] == split]
self.model = model
comet_rels = {'isBefore':'is before', 'isAfter': 'is after', 'Causes': 'causes', 'CausedBy':'is caused by', 'HinderedBy': 'is hindered by', 'Hinders': 'hinders', 'xIntent': 'because they intend to', 'xNeed': 'because they need to'}
if n_labels == 2:
df = df[df['rel']<2]
if kg != 'none':
self.knowledge = df[kg].values.tolist()
if kg == 'comet' or kg == 'comet_link':
self.knowledge = [eval(x) for x in self.knowledge]
self.knowledge = [chain.from_iterable(x) for x in self.knowledge]
self.knowledge = [[comet_rels[x] if x in comet_rels.keys() else x for x in k] for k in self.knowledge]
self.knowledge = [' '.join(x) for x in self.knowledge]
if kg == 'conceptnet':
self.knowledge = [". ".join(eval(x)) for x in self.knowledge]
if self.model == 'hybrid' or self.model == 'attn':
self.args = tokenizer(df['Arg1'].values.tolist(), df['Arg2'].values.tolist(), padding=True, truncation=True)
self.knowledge = tokenizer(self.knowledge, padding=True, truncation=True)
else:
zipped_args = zip(self.knowledge, df['Arg2'].values.tolist())
arg2 = [k + ' ' + kg_token + ' ' + a for k, a in zipped_args]
self.args = tokenizer(df['Arg1'].astype(str).values.tolist(),
arg2,
padding=True, truncation=True)
else:
self.args = tokenizer(df['Arg1'].values.tolist(), df['Arg2'].values.tolist(), padding=True, truncation=True)
self.labels = df['rel'].astype(int).values.tolist()
def __len__(self):
return len(self.labels)
def __getitem__(self, i):
label = self.labels[i]
item = {key: torch.tensor(val[i]) for key, val in self.args.items()}
item['labels'] = torch.tensor(label)
if self.model == 'hybrid' or self.model == 'attn':
knowledge = {key: torch.tensor(val[i]) for key, val in self.knowledge.items()}
return item, knowledge
else:
return item
def injection_collator(batch):
txt = {}
kg = {}
txt_keys = batch[0][0].keys()
kg_keys = batch[0][1].keys()
for k in txt_keys:
txt[k] = torch.stack([t[0][k] for t in batch])
for k in kg_keys:
kg[k] = torch.stack([t[1][k] for t in batch])
return txt, kg
def default_collator(batch):
txt = {}
txt_keys = batch[0].keys()
for k in txt_keys:
txt[k] = torch.stack([t[k] for t in batch])
return txt
class LinkPredDataset(Dataset):
def __init__(self, split='train'):
print(f'loading {split} data')
df = pd.read_csv(f'./data/atomic2020/{split}.tsv', delimiter='\t', names=['head', 'rel', 'tail'])
n = int(df['rel'].value_counts().mean())
print(df['rel'].value_counts())
print(n)
no_rel = None
diff = n
i = 1
while diff > 0:
if diff == n:
heads = df[(df['head'].str.contains('none')) | (df['head'] != 'NaN')]['head'].dropna()
tails = df[(df['tail'] != 'none') | (df['tail'] != 'NaN')]['tail'].dropna()
heads = heads.sample(n=diff, random_state=i, ignore_index=True, replace=True)
tails = tails.sample(n=diff, random_state=i, ignore_index=True, replace=True)
temp = pd.concat([heads, tails], axis=1).rename(columns={0:'head', 1:'tail'})
temp = pd.merge(temp, df[['head', 'tail']], on=['head', 'tail'],indicator='i', how='left')
no_rel = temp.query('i == "left_only"').drop('i', 1).reset_index(drop=True)
no_rel['rel'] = 'None'
diff = max(0, n - no_rel.shape[0])
else:
heads = df[(df['head'].str.contains('none')) | (df['head'] != 'NaN')]['head'].dropna()
tails = df[(df['tail'] != 'none') | (df['tail'] != 'NaN')]['tail'].dropna()
heads = heads.sample(n=diff, random_state=i, ignore_index=True, replace=True)
tails = tails.sample(n=diff, random_state=i, ignore_index=True, replace=True)
temp = pd.concat([heads, tails], axis=1).rename(columns={0:'head', 1:'tail'})
temp = pd.merge(temp, df[['head', 'tail']], on=['head', 'tail'], indicator='i', how='left')
temp = temp.query('i == "left_only"').drop('i', 1).reset_index(drop=True)
no_rel = pd.concat([no_rel, temp], ignore_index=True)
no_rel['rel'] = 'None'
no_rel = no_rel.drop_duplicates()
diff = max(0, n - no_rel.shape[0])
i += 1
assert (no_rel.shape[0] == n)
print(f'Adding {n} none relations to the existing {df.shape[0]} samples')
df = pd.concat([df, no_rel], ignore_index=True)
df = df[df['rel'].isin(ix2rel)]
self.tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')
self.args = list(zip(df['head'].astype(str).tolist(), df['tail'].astype(str).tolist()))
self.labels = [rel2ix[rel] for rel in df['rel']]
def __len__(self):
return len(self.labels)
def __getitem__(self, idx):
item = self.tokenizer(*self.args[idx], truncation=True, padding='max_length')
item = {key: torch.tensor(val) for key, val in item.items()}
item['labels'] = torch.tensor(self.labels[idx])
return item