def pairwise_bootstrap(file1, file2, gold_attr, pred_attr, cmp_func, iters):
d1, d2 = DataSet(), DataSet()
log_info('Loading File1: %s' % file1)
d1.load_from_arff(file1)
log_info('Loading File2: %s' % file2)
d2.load_from_arff(file2)
gold = d1.attrib_as_vect(gold_attr)
p1 = d1.attrib_as_vect(pred_attr)
p2 = d2.attrib_as_vect(pred_attr)
p1_better, p2_better, ties = 0, 0, 0
for i in xrange(iters):
sample = rnd.randint(0, len(gold), len(gold))
s_p1_good = sum(1 if cmp_func(gold[i], p1[i]) else 0 for i in sample)
s_p2_good = sum(1 if cmp_func(gold[i], p2[i]) else 0 for i in sample)
log_info('Round %d: File1 - %2.2f vs. File2 - %2.2f' %
(i, float(s_p1_good) / len(gold) * 100,
float(s_p2_good) / len(gold) * 100))
if s_p1_good > s_p2_good:
p1_better += 1
elif s_p2_good > s_p1_good:
p2_better += 1
else:
ties += 1
print ('File1 better: %d (%2.2f) | File2 better: %d (%2.2f) |' +
' ties: %d (%2.2f)') % (p1_better, float(p1_better) / iters * 100,
p2_better, float(p2_better) / iters * 100,
ties, float(ties) / iters * 100,)
def evaluate_oov(data, source_attr, target_attr, forms_attr,
oov_test_file, oov_part):
"""\
Out-of-vocabulary evaluation
"""
log_info('Loading known lemmas and forms from: ' + oov_test_file)
train = DataSet()
train.load_from_arff(oov_test_file)
if oov_part < 1:
log_info('Using only %f-part of the file.' % oov_part)
train = train.subset(0, int(round(oov_part * len(train))), copy=False)
known_forms = {i[target_attr].lower() for i in train}
known_lemmas = {i[source_attr].lower() for i in train}
oov_forms = [1 if i[target_attr].lower() not in known_forms else 0
for i in data]
oov_lemmas = [1 if i[source_attr] not in known_lemmas else 0
for i in data]
data.add_attrib(Attribute('OOV_FORM', 'numeric'), oov_forms)
data.add_attrib(Attribute('OOV_LEMMA', 'numeric'), oov_lemmas)
oov_forms_count = sum(oov_forms)
oov_forms_good = count_correct(data, target_attr, forms_attr,
lambda i: i['OOV_FORM'])
print_score(oov_forms_good, oov_forms_count, 'OOV forms')
oov_lemmas_count = sum(oov_lemmas)
oov_lemmas_good = count_correct(data, target_attr, forms_attr,
lambda i: i['OOV_LEMMA'])
print_score(oov_lemmas_good, oov_lemmas_count, 'OOV lemmas')
def evaluate_oov(data, source_attr, target_attr, forms_attr, oov_test_file,
oov_part):
"""\
Out-of-vocabulary evaluation
"""
log_info('Loading known lemmas and forms from: ' + oov_test_file)
train = DataSet()
train.load_from_arff(oov_test_file)
if oov_part < 1:
log_info('Using only %f-part of the file.' % oov_part)
train = train.subset(0, int(round(oov_part * len(train))), copy=False)
known_forms = {i[target_attr].lower() for i in train}
known_lemmas = {i[source_attr].lower() for i in train}
oov_forms = [
1 if i[target_attr].lower() not in known_forms else 0 for i in data
]
oov_lemmas = [1 if i[source_attr] not in known_lemmas else 0 for i in data]
data.add_attrib(Attribute('OOV_FORM', 'numeric'), oov_forms)
data.add_attrib(Attribute('OOV_LEMMA', 'numeric'), oov_lemmas)
oov_forms_count = sum(oov_forms)
oov_forms_good = count_correct(data, target_attr, forms_attr,
lambda i: i['OOV_FORM'])
print_score(oov_forms_good, oov_forms_count, 'OOV forms')
oov_lemmas_count = sum(oov_lemmas)
oov_lemmas_good = count_correct(data, target_attr, forms_attr,
lambda i: i['OOV_LEMMA'])
print_score(oov_lemmas_good, oov_lemmas_count, 'OOV lemmas')
def main():
"""\
Main application entry: parse command line and run the test.
"""
opts, filenames = getopt.getopt(sys.argv[1:], 'ca:s:n:')
show_help = False
combine_cng = False
subsets = []
neighbors = []
substrs = []
for opt, arg in opts:
if opt == '-c':
combine_cng = True
elif opt == '-s':
sub_len, attr = arg.split(':', 1)
substrs.append((int(sub_len), attr))
elif opt == '-a':
size, attrs = arg.split(':', 1)
subsets.append((int(size), re.split(r'[, ]+', attrs)))
elif opt == '-n':
shift, attrs = arg.split(':', 1)
neighbors.append((int(shift), re.split(r'[, ]+', attrs)))
# display help and exit
if len(filenames) != 2 or not (combine_cng or substrs or
subsets or neighbors) or show_help:
display_usage()
sys.exit(1)
# run the training
filename_in, filename_out = filenames
data = DataSet()
log_info('Loading data: ' + filename_in)
data.load_from_arff(filename_in)
if substrs:
for (sub_len, attr) in substrs:
log_info(('Adding substrings from the %s of %s ' +
'up to %d characters long ...') %
(('beginning' if sub_len > 0 else 'end'),
attr, abs(sub_len)))
add_substr_attributes(data, sub_len, attr)
if combine_cng:
log_info('Combining case, number, gender ...')
combine_tag_num_gen_cas(data)
if subsets:
for (set_size, set_attrs) in subsets:
log_info('Combining up to %d attributes from [%s] ...' %
(set_size, ','.join(set_attrs)))
combine_subsets(data, set_attrs, set_size)
if neighbors:
for (shift, attrs) in neighbors:
log_info('Adding neighbor %d\'s attributes [%s] ...' %
(shift, ','.join(attrs)))
add_neighbor_attributes(data, shift, attrs)
log_info('Saving data: ' + filename_out)
data.save_to_arff(filename_out)
def convert(in_file, out_file, feat_no, use_feat_names, cpos_chars):
"""\
This does the conversion to ARFF.
"""
fh_in = file_stream(in_file)
buf = []
sent_id = 1
word_id = 1
for line in fh_in:
line = line.rstrip('\r\n')
if not line:
sent_id += 1
word_id = 1
continue
# split the CoNLL format, removing unwanted stuff
_, form, lemma, _, pos, _, feat, _ = line.split('\t', 7)
# copy attributes
inst = {
'Form': form, 'Lemma': lemma, 'Tag_POS': pos,
'word_id': word_id, 'sent_id': in_file + '-' + str(sent_id)
}
# computing form-lemma diff (edit script)
escr_front, escr_midback = edit_script(lemma, form)
inst['LemmaFormDiff_Front'] = escr_front
inst['LemmaFormDiff_Back'] = escr_midback
# lemma suffixes
for i in xrange(1, 9):
inst['LemmaSuff_' + str(i)] = lemma[-i:]
# coarse POS
inst['Tag_CPOS'] = pos[:cpos_chars]
# POS features
feats = feat.split('|', feat_no - 1)
feats += [''] * (feat_no - len(feats))
for feat_ord, feat in enumerate(feats, start=1):
if use_feat_names:
feat_name, feat_val = feat.split('=', 1)
inst['Tag_' + feat_name] = feat_val
else:
inst['Tag_FEAT' + str(feat_ord)] = feat
# increase word number
word_id += 1
# save the instance to the buffer
buf.append(inst)
# write this all out as ARFF
data = DataSet()
attr_order = ['sent_id', 'word_id', 'Lemma', 'Form',
'LemmaFormDiff_Front', 'LemmaFormDiff_Back']
for i in xrange(1, 9):
attr_order.append('LemmaSuff_' + str(i))
attr_order.extend(['Tag_POS', 'Tag_CPOS'])
data.load_from_dict(buf, {'word_id': 'numeric'}, attr_order)
data.save_to_arff(out_file)
def test_models(file_in, file_out, model_files, source_attr, target_attr,
oov_test_file, oov_part, pos_attr, test_indiv):
"""\
Test all the given models on the selected file and save the target.
If oov_test_file is set, performs also OOV evaluation.
If test_pos is True, prints detailed results for various POSs.
"""
# load testing data
log_info('Loading data: ' + file_in)
data = DataSet()
data.load_from_arff(file_in)
forms = data[source_attr]
# apply all models
for model_num, model_file in enumerate(model_files, start=1):
model = Model.load_from_file(model_file)
log_info('Applying model: ' + model_file)
rules = model.classify(data)
output_attr = 'OUTPUT_M' + str(model_num)
data.add_attrib(Attribute(output_attr, 'string'), rules)
if test_indiv:
good = count_correct(data, model.class_attr, output_attr)
print_score(good, len(data), 'Model accuracy')
forms = [inflect(form, rule) for form, rule in zip(forms, rules)]
forms_attr = 'FORMS_M' + str(model_num)
data.add_attrib(Attribute(forms_attr, 'string'), forms)
# test the final performance
log_info('Evaluating...')
good = count_correct(data, target_attr, forms_attr)
print_score(good, len(data), 'ALL')
# evaluate without punctuation
evaluate_nopunct(data, source_attr, target_attr, forms_attr)
# evaluate forms different from lemma
evaluate_nolemma(data, source_attr, target_attr, forms_attr)
# load training data for OOV tests, evaluate on OOV
if oov_test_file:
evaluate_oov(data, source_attr, target_attr, forms_attr, oov_test_file,
oov_part)
# test on different POSes
if pos_attr:
evaluate_poses(data, target_attr, forms_attr, pos_attr)
# save the classification results
log_info('Saving data: ' + file_out)
data.save_to_arff(file_out)
def test_models(file_in, file_out, model_files, source_attr, target_attr,
oov_test_file, oov_part, pos_attr, test_indiv):
"""\
Test all the given models on the selected file and save the target.
If oov_test_file is set, performs also OOV evaluation.
If test_pos is True, prints detailed results for various POSs.
"""
# load testing data
log_info('Loading data: ' + file_in)
data = DataSet()
data.load_from_arff(file_in)
forms = data[source_attr]
# apply all models
for model_num, model_file in enumerate(model_files, start=1):
model = Model.load_from_file(model_file)
log_info('Applying model: ' + model_file)
rules = model.classify(data)
output_attr = 'OUTPUT_M' + str(model_num)
data.add_attrib(Attribute(output_attr, 'string'), rules)
if test_indiv:
good = count_correct(data, model.class_attr, output_attr)
print_score(good, len(data), 'Model accuracy')
forms = [inflect(form, rule) for form, rule in zip(forms, rules)]
forms_attr = 'FORMS_M' + str(model_num)
data.add_attrib(Attribute(forms_attr, 'string'), forms)
# test the final performance
log_info('Evaluating...')
good = count_correct(data, target_attr, forms_attr)
print_score(good, len(data), 'ALL')
# evaluate without punctuation
evaluate_nopunct(data, source_attr, target_attr, forms_attr)
# evaluate forms different from lemma
evaluate_nolemma(data, source_attr, target_attr, forms_attr)
# load training data for OOV tests, evaluate on OOV
if oov_test_file:
evaluate_oov(data, source_attr, target_attr, forms_attr,
oov_test_file, oov_part)
# test on different POSes
if pos_attr:
evaluate_poses(data, target_attr, forms_attr, pos_attr)
# save the classification results
log_info('Saving data: ' + file_out)
data.save_to_arff(file_out)
def convert(in_file, out_file, feat_no, use_feat_names, cpos_chars):
"""\
This does the conversion to ARFF.
"""
fh_in = file_stream(in_file)
buf = []
sent_id = 1
word_id = 1
for line in fh_in:
line = line.rstrip('\r\n')
if not line:
sent_id += 1
word_id = 1
continue
# split the CoNLL format, removing unwanted stuff
_, form, lemma, _, pos, _, feat, _ = line.split('\t', 7)
# copy attributes
inst = {
'Form': form,
'Lemma': lemma,
'Tag_POS': pos,
'word_id': word_id,
'sent_id': in_file + '-' + str(sent_id)
}
# computing form-lemma diff (edit script)
escr_front, escr_midback = edit_script(lemma, form)
inst['LemmaFormDiff_Front'] = escr_front
inst['LemmaFormDiff_Back'] = escr_midback
# lemma suffixes
for i in xrange(1, 9):
inst['LemmaSuff_' + str(i)] = lemma[-i:]
# coarse POS
inst['Tag_CPOS'] = pos[:cpos_chars]
# POS features
feats = feat.split('|', feat_no - 1)
feats += [''] * (feat_no - len(feats))
for feat_ord, feat in enumerate(feats, start=1):
if use_feat_names:
feat_name, feat_val = feat.split('=', 1)
inst['Tag_' + feat_name] = feat_val
else:
inst['Tag_FEAT' + str(feat_ord)] = feat
# increase word number
word_id += 1
# save the instance to the buffer
buf.append(inst)
# write this all out as ARFF
data = DataSet()
attr_order = [
'sent_id', 'word_id', 'Lemma', 'Form', 'LemmaFormDiff_Front',
'LemmaFormDiff_Back'
]
for i in xrange(1, 9):
attr_order.append('LemmaSuff_' + str(i))
attr_order.extend(['Tag_POS', 'Tag_CPOS'])
data.load_from_dict(buf, {'word_id': 'numeric'}, attr_order)
data.save_to_arff(out_file)
def get_stats(data_file, train_file, source_attr, target_attr):
"""\
"""
data = DataSet()
log_info('Loading data from %s...' % data_file)
data.load_from_arff(data_file)
print_feat(data, lambda a, b: True, 'total')
print_feat(data, lambda _, i: not regex.match(r'^\p{P}', i[source_attr]),
'excluding punctuation')
print_feat(data, lambda _, i: i[source_attr].lower() != \
i[target_attr].lower(), 'inflected forms')
if train_file is not None:
log_info('Loading known data from %s...' % train_file)
train = DataSet()
train.load_from_arff(train_file)
known = {i[target_attr].lower() for i in train}
print_feat(data, lambda _, i: not i[target_attr].lower() in known,
'unknown')
def main():
"""\
Main application entry: parse command line and run the test.
"""
opts, filenames = getopt.getopt(sys.argv[1:], 'ca:s:n:')
show_help = False
combine_cng = False
subsets = []
neighbors = []
substrs = []
for opt, arg in opts:
if opt == '-c':
combine_cng = True
elif opt == '-s':
sub_len, attr = arg.split(':', 1)
substrs.append((int(sub_len), attr))
elif opt == '-a':
size, attrs = arg.split(':', 1)
subsets.append((int(size), re.split(r'[, ]+', attrs)))
elif opt == '-n':
shift, attrs = arg.split(':', 1)
neighbors.append((int(shift), re.split(r'[, ]+', attrs)))
# display help and exit
if len(filenames) != 2 or not (combine_cng or substrs or
subsets or neighbors) or show_help:
display_usage()
sys.exit(1)
# run the training
filename_in, filename_out = filenames
data = DataSet()
log_info('Loading data: ' + filename_in)
data.load_from_arff(filename_in)
if substrs:
for (sub_len, attr) in substrs:
log_info(('Adding substrings from the %s of %s ' +
'up to %d characters long ...') %
(('beginning' if sub_len > 0 else 'end'),
attr, abs(sub_len)))
add_substr_attributes(data, sub_len, attr)
if combine_cng:
log_info('Combining case, number, gender ...')
combine_tag_num_gen_cas(data)
if subsets:
for (set_size, set_attrs) in subsets:
log_info('Combining up to %d attributes from [%s] ...' %
(set_size, ','.join(set_attrs)))
combine_subsets(data, set_attrs, set_size)
if neighbors:
for (shift, attrs) in neighbors:
log_info('Adding neighbor %d\'s attributes [%s] ...' %
(shift, ','.join(attrs)))
add_neighbor_attributes(data, shift, attrs)
log_info('Saving data: ' + filename_out)
data.save_to_arff(filename_out)
def main():
"""\
Main application entry: parse command line and run the test.
"""
opts, filenames = getopt.getopt(sys.argv[1:], 'g:p:ai')
show_help = False
annot_errors = False
gold = None
predicted = 'PREDICTED'
ignore_case = False
for opt, arg in opts:
if opt == '-g':
gold = arg
elif opt == '-p':
predicted = arg
elif opt == '-a':
annot_errors = True
elif opt == '-i':
ignore_case = True
# display help and exit
if len(filenames) != 2 or not gold or show_help:
display_usage()
sys.exit(1)
# run the training
filename_in, filename_out = filenames
data = DataSet()
log_info('Loading data: ' + filename_in)
data.load_from_arff(filename_in)
if ignore_case:
cmp_func = lambda a, b: a.lower() != b.lower()
else:
cmp_func = lambda a, b: a != b
if annot_errors:
log_info('Annotating errors...')
err_ind = [
'ERR' if cmp_func(i[gold], i[predicted]) else '' for i in data
]
data.add_attrib(Attribute('ERROR_IND', 'string'), err_ind)
else:
log_info('Selecting errors...')
data = data[lambda _, i: cmp_func(i[gold], i[predicted])]
log_info('Saving data: ' + filename_out)
data.save_to_arff(filename_out)
def main():
"""\
Main application entry: parse command line and run the test.
"""
opts, filenames = getopt.getopt(sys.argv[1:], "g:p:ai")
show_help = False
annot_errors = False
gold = None
predicted = "PREDICTED"
ignore_case = False
for opt, arg in opts:
if opt == "-g":
gold = arg
elif opt == "-p":
predicted = arg
elif opt == "-a":
annot_errors = True
elif opt == "-i":
ignore_case = True
# display help and exit
if len(filenames) != 2 or not gold or show_help:
display_usage()
sys.exit(1)
# run the training
filename_in, filename_out = filenames
data = DataSet()
log_info("Loading data: " + filename_in)
data.load_from_arff(filename_in)
if ignore_case:
cmp_func = lambda a, b: a.lower() != b.lower()
else:
cmp_func = lambda a, b: a != b
if annot_errors:
log_info("Annotating errors...")
err_ind = ["ERR" if cmp_func(i[gold], i[predicted]) else "" for i in data]
data.add_attrib(Attribute("ERROR_IND", "string"), err_ind)
else:
log_info("Selecting errors...")
data = data[lambda _, i: cmp_func(i[gold], i[predicted])]
log_info("Saving data: " + filename_out)
data.save_to_arff(filename_out)
