def folds_static(in_corpus, out_corpus, params, condition_name, n_folds):
maF1s = defaultdict(list)
miF1s = defaultdict(list)
folds = list(get_static_folds())
predictions = defaultdict(dict)
decisions = defaultdict(dict)
for train_tids, test_tids, i in folds[:n_folds]:
print(f"Iteration: {i}")
ensemble_basename = condition_name.split('|')[0]
ensemble_name = "{}__{}__{}".format(
ensemble_basename, hash_of_featureset(params['feature_set']), i)
clf = EvidenceGraphClassifier(feature_function_segments,
feature_function_segmentpairs, **params)
train_txt = [g for t, g in in_corpus.items() if t in train_tids]
train_arg = [g for t, g in out_corpus.items() if t in train_tids]
try:
# load ensemble of pretrained base classifiers
clf.ensemble = load(modelpath + ensemble_name)
if params['optimize_weighting']:
# and train metaclassifier (if desired)
clf.train_metaclassifier(train_txt, train_arg)
except RuntimeError:
# train ensemble
clf.train(train_txt, train_arg)
save(clf.ensemble, modelpath + ensemble_name)
# test
test_txt = [g for t, g in in_corpus.items() if t in test_tids]
test_arg = [g for t, g in out_corpus.items() if t in test_tids]
score_msg = ''
for level, base_classifier in clf.ensemble.items():
maF1, miF1 = base_classifier.test(test_txt, test_arg)
maF1s[level].append(maF1)
miF1s[level].append(miF1)
score_msg += "{}: {:.3f}\t".format(level, maF1)
decoded_scores = []
for t in test_tids:
mst = clf.predict(in_corpus[t])
decoded_scores.append(clf.score(mst, out_corpus[t]))
predictions[i][t] = mst.get_triples()
decisions[i][t] = clf.predict_decisions(in_corpus[t])
score_msg += "decoded: {:.3f}\t".format(mean(decoded_scores))
print(score_msg)
print("Average macro and micro F1:")
for level in maF1s:
avg_maF1 = mean(maF1s[level])
avg_miF1 = mean(miF1s[level])
print(level, avg_maF1, avg_miF1)
return predictions, decisions
def init_lang_classifier(model_path):
init_language('de')
features = [
'default', 'bow', 'bow_2gram', 'first_three',
'tags', 'deps_lemma', 'deps_tag',
'punct', 'verb_main', 'verb_all', 'discourse_marker',
'context', 'clusters', 'clusters_2gram', 'discourse_relation',
'vector_left_right', 'vector_source_target',
'verb_segment', 'same_sentence', 'matrix_clause'
]
PARAMS = {'feature_set': features,
'relation_set': FULL_RELATION_SET_ADU,
'optimize': False,
'optimize_weighting': False}
clf = EvidenceGraphClassifier(feature_function_segments, feature_function_segmentpairs, **PARAMS)
clf.ensemble = eg_utils.load(model_path)
return clf
'same_sentence', 'matrix_clause'
]
params = {
'feature_set': features,
'relation_set': FULL_RELATION_SET_ADU,
'optimize': False,
'optimize_weighting': False
}
logger.info("Loading pre-trained model")
clf = EvidenceGraphClassifier(feature_function_segments,
feature_function_segmentpairs, **params)
clf.ensemble = eg_utils.load(model_path)
logger.info("Predicting on test case")
mst = clf.predict(segments)
roles = mst.get_ro_vector()
triples = mst.get_triples()
role_map = {1: 'opp', 0: 'pro'}
seg_map = {
"support": "sup",
"rebut": "reb",
"undercut": "und",