File size: 9,243 Bytes
cd5fcb4 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 |
# Copyright (c) 2017-present, Facebook, Inc.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
#
'''
SICK Relatedness and Entailment
'''
from __future__ import absolute_import, division, unicode_literals
import os
import io
import logging
import numpy as np
from sklearn.metrics import mean_squared_error
from scipy.stats import pearsonr, spearmanr
from senteval.tools.relatedness import RelatednessPytorch
from senteval.tools.validation import SplitClassifier
class SICKEval(object):
def __init__(self, task_path, seed=1111):
logging.debug('***** Transfer task : SICK-Relatedness*****\n\n')
self.seed = seed
train = self.loadFile(os.path.join(task_path, 'SICK_train.txt'))
dev = self.loadFile(os.path.join(task_path, 'SICK_trial.txt'))
test = self.loadFile(os.path.join(task_path, 'SICK_test_annotated.txt'))
self.sick_data = {'train': train, 'dev': dev, 'test': test}
def do_prepare(self, params, prepare):
samples = self.sick_data['train']['X_A'] + \
self.sick_data['train']['X_B'] + \
self.sick_data['dev']['X_A'] + \
self.sick_data['dev']['X_B'] + \
self.sick_data['test']['X_A'] + self.sick_data['test']['X_B']
return prepare(params, samples)
def loadFile(self, fpath):
skipFirstLine = True
sick_data = {'X_A': [], 'X_B': [], 'y': []}
with io.open(fpath, 'r', encoding='utf-8') as f:
for line in f:
if skipFirstLine:
skipFirstLine = False
else:
text = line.strip().split('\t')
sick_data['X_A'].append(text[1].split())
sick_data['X_B'].append(text[2].split())
sick_data['y'].append(text[3])
sick_data['y'] = [float(s) for s in sick_data['y']]
return sick_data
def run(self, params, batcher):
sick_embed = {'train': {}, 'dev': {}, 'test': {}}
bsize = params.batch_size
for key in self.sick_data:
logging.info('Computing embedding for {0}'.format(key))
# Sort to reduce padding
sorted_corpus = sorted(zip(self.sick_data[key]['X_A'],
self.sick_data[key]['X_B'],
self.sick_data[key]['y']),
key=lambda z: (len(z[0]), len(z[1]), z[2]))
self.sick_data[key]['X_A'] = [x for (x, y, z) in sorted_corpus]
self.sick_data[key]['X_B'] = [y for (x, y, z) in sorted_corpus]
self.sick_data[key]['y'] = [z for (x, y, z) in sorted_corpus]
for txt_type in ['X_A', 'X_B']:
sick_embed[key][txt_type] = []
for ii in range(0, len(self.sick_data[key]['y']), bsize):
batch = self.sick_data[key][txt_type][ii:ii + bsize]
embeddings = batcher(params, batch)
sick_embed[key][txt_type].append(embeddings)
sick_embed[key][txt_type] = np.vstack(sick_embed[key][txt_type])
sick_embed[key]['y'] = np.array(self.sick_data[key]['y'])
logging.info('Computed {0} embeddings'.format(key))
# Train
trainA = sick_embed['train']['X_A']
trainB = sick_embed['train']['X_B']
trainF = np.c_[np.abs(trainA - trainB), trainA * trainB]
trainY = self.encode_labels(self.sick_data['train']['y'])
# Dev
devA = sick_embed['dev']['X_A']
devB = sick_embed['dev']['X_B']
devF = np.c_[np.abs(devA - devB), devA * devB]
devY = self.encode_labels(self.sick_data['dev']['y'])
# Test
testA = sick_embed['test']['X_A']
testB = sick_embed['test']['X_B']
testF = np.c_[np.abs(testA - testB), testA * testB]
testY = self.encode_labels(self.sick_data['test']['y'])
config = {'seed': self.seed, 'nclasses': 5}
clf = RelatednessPytorch(train={'X': trainF, 'y': trainY},
valid={'X': devF, 'y': devY},
test={'X': testF, 'y': testY},
devscores=self.sick_data['dev']['y'],
config=config)
devspr, yhat = clf.run()
pr = pearsonr(yhat, self.sick_data['test']['y'])[0]
sr = spearmanr(yhat, self.sick_data['test']['y'])[0]
pr = 0 if pr != pr else pr
sr = 0 if sr != sr else sr
se = mean_squared_error(yhat, self.sick_data['test']['y'])
logging.debug('Dev : Spearman {0}'.format(devspr))
logging.debug('Test : Pearson {0} Spearman {1} MSE {2} \
for SICK Relatedness\n'.format(pr, sr, se))
return {'devspearman': devspr, 'pearson': pr, 'spearman': sr, 'mse': se,
'yhat': yhat, 'ndev': len(devA), 'ntest': len(testA)}
def encode_labels(self, labels, nclass=5):
"""
Label encoding from Tree LSTM paper (Tai, Socher, Manning)
"""
Y = np.zeros((len(labels), nclass)).astype('float32')
for j, y in enumerate(labels):
for i in range(nclass):
if i+1 == np.floor(y) + 1:
Y[j, i] = y - np.floor(y)
if i+1 == np.floor(y):
Y[j, i] = np.floor(y) - y + 1
return Y
class SICKEntailmentEval(SICKEval):
def __init__(self, task_path, seed=1111):
logging.debug('***** Transfer task : SICK-Entailment*****\n\n')
self.seed = seed
train = self.loadFile(os.path.join(task_path, 'SICK_train.txt'))
dev = self.loadFile(os.path.join(task_path, 'SICK_trial.txt'))
test = self.loadFile(os.path.join(task_path, 'SICK_test_annotated.txt'))
self.sick_data = {'train': train, 'dev': dev, 'test': test}
def loadFile(self, fpath):
label2id = {'CONTRADICTION': 0, 'NEUTRAL': 1, 'ENTAILMENT': 2}
skipFirstLine = True
sick_data = {'X_A': [], 'X_B': [], 'y': []}
with io.open(fpath, 'r', encoding='utf-8') as f:
for line in f:
if skipFirstLine:
skipFirstLine = False
else:
text = line.strip().split('\t')
sick_data['X_A'].append(text[1].split())
sick_data['X_B'].append(text[2].split())
sick_data['y'].append(text[4])
sick_data['y'] = [label2id[s] for s in sick_data['y']]
return sick_data
def run(self, params, batcher):
sick_embed = {'train': {}, 'dev': {}, 'test': {}}
bsize = params.batch_size
for key in self.sick_data:
logging.info('Computing embedding for {0}'.format(key))
# Sort to reduce padding
sorted_corpus = sorted(zip(self.sick_data[key]['X_A'],
self.sick_data[key]['X_B'],
self.sick_data[key]['y']),
key=lambda z: (len(z[0]), len(z[1]), z[2]))
self.sick_data[key]['X_A'] = [x for (x, y, z) in sorted_corpus]
self.sick_data[key]['X_B'] = [y for (x, y, z) in sorted_corpus]
self.sick_data[key]['y'] = [z for (x, y, z) in sorted_corpus]
for txt_type in ['X_A', 'X_B']:
sick_embed[key][txt_type] = []
for ii in range(0, len(self.sick_data[key]['y']), bsize):
batch = self.sick_data[key][txt_type][ii:ii + bsize]
embeddings = batcher(params, batch)
sick_embed[key][txt_type].append(embeddings)
sick_embed[key][txt_type] = np.vstack(sick_embed[key][txt_type])
logging.info('Computed {0} embeddings'.format(key))
# Train
trainA = sick_embed['train']['X_A']
trainB = sick_embed['train']['X_B']
trainF = np.c_[np.abs(trainA - trainB), trainA * trainB]
trainY = np.array(self.sick_data['train']['y'])
# Dev
devA = sick_embed['dev']['X_A']
devB = sick_embed['dev']['X_B']
devF = np.c_[np.abs(devA - devB), devA * devB]
devY = np.array(self.sick_data['dev']['y'])
# Test
testA = sick_embed['test']['X_A']
testB = sick_embed['test']['X_B']
testF = np.c_[np.abs(testA - testB), testA * testB]
testY = np.array(self.sick_data['test']['y'])
config = {'nclasses': 3, 'seed': self.seed,
'usepytorch': params.usepytorch,
'classifier': params.classifier,
'nhid': params.nhid}
clf = SplitClassifier(X={'train': trainF, 'valid': devF, 'test': testF},
y={'train': trainY, 'valid': devY, 'test': testY},
config=config)
devacc, testacc = clf.run()
logging.debug('\nDev acc : {0} Test acc : {1} for \
SICK entailment\n'.format(devacc, testacc))
return {'devacc': devacc, 'acc': testacc,
'ndev': len(devA), 'ntest': len(testA)}
|