def labels_precision_recall(system_output_path,
gold_path,
labels=["nsubj", "nsubjpass"],
fine_grained_deprels=True):
system = udtree.from_files(system_output_path)
gold = udtree.from_files(gold_path)
system_correct, system_incorrect, gold_count = 0, 0, 0
for system_tree, gold_tree in zip(system, gold):
(tree_correct,
tree_incorrect) = match_tree_attachments(system_tree, gold_tree, True,
fine_grained_deprels=fine_grained_deprels)
for _, _, system_label, _, _ in tree_correct:
if system_label in labels:
system_correct += 1
gold_count += 1
for _, _, system_label, _, gold_label in tree_incorrect:
if gold_label in labels:
gold_count += 1
if system_label in labels:
system_incorrect += 1
if system_correct + system_incorrect == 0:
precision = float("NaN")
else:
precision = system_correct / (system_correct + system_incorrect)
if gold_count == 0:
recall = float("NaN")
else:
recall = system_correct / (gold_count)
return precision, recall
def weighted_las(system_output_path, gold_path, weights):
system = udtree.from_files(system_output_path)
gold = udtree.from_files(gold_path)
correct, incorrect = 0, 0
for system_tree, gold_tree in zip(system, gold):
(tree_correct,
tree_incorrect) = match_tree_attachments(system_tree, gold_tree, True,
fine_grained_deprels=False)
for _, _, _, _, gold_label in tree_correct:
correct += weights[gold_label]
for _, _, _, _, gold_label in tree_incorrect:
incorrect += weights[gold_label]
if (correct + incorrect) == 0:
return float("NaN")
return correct / (correct + incorrect)
def attachment_score(system_output_path, gold_path, labeled=True, fine_grained_deprels=True, include_punct=False):
puncts = None
if not include_punct:
puncts = set.union(is_only_punctuation(gold_path), {'punct'})
system = udtree.from_files(system_output_path)
gold = udtree.from_files(gold_path)
correct, incorrect = 0, 0
for system_tree, gold_tree in zip(system, gold):
(tree_correct,
tree_incorrect) = match_tree_attachments(system_tree, gold_tree, labeled,
fine_grained_deprels=fine_grained_deprels,
ignore_deprels=puncts)
correct += len(tree_correct)
incorrect += len(tree_incorrect)
if (correct + incorrect) == 0:
return float("NaN")
return correct / (correct + incorrect)
def root_distance_las(system_output_path, gold_path, include_punct=False):
puncts = None
if not include_punct:
puncts = set.union(is_only_punctuation(gold_path), {'punct'})
system = udtree.from_files(system_output_path)
gold = udtree.from_files(gold_path)
correct, incorrect = 0, 0
for system_tree, gold_tree in zip(system, gold):
(tree_correct,
tree_incorrect) = match_tree_attachments(system_tree, gold_tree, True,
fine_grained_deprels=False,
ignore_deprels=puncts)
for index, _, _, _, gold_label in tree_correct:
correct += 1 / root_distance(gold_tree, index)
for index, _, _, _, gold_label in tree_incorrect:
incorrect += 1 / root_distance(gold_tree, index)
if (correct + incorrect) == 0:
return float("NaN")
return correct / (correct + incorrect)
def get_nonprojectivity_ratios():
lang_trees = lang_utils.get_ud_paths('../resources/universaldependencies1-2/universal-dependencies-1.2/', type_='train', format_='conllu', coarse=False)
np = dict()
tot = dict()
for lang, path in lang_trees.items():
trees = list(udtree.from_files(path))
tot[lang] = len(trees)
np[lang] = get_np_trees(trees)
rel = {}
for lang, n in zip(np.keys(), np.values()):
rel[lang] = len(n) / tot[lang]
return rel
def is_only_punctuation(gold_path):
is_punct = {}
puncts = set(string.punctuation)
for gold_tree in udtree.from_files(gold_path):
for token, deprel in zip(gold_tree.tokens, gold_tree.deprels):
all_puncts = True
for char in token:
if char not in puncts:
all_puncts = False
break
if all_puncts:
is_punct[deprel] = True
else:
is_punct[deprel] = False
return {deprel for deprel,val in is_punct.items() if val}
def convert(to_convert_files):
for lang, files in to_convert_files.items():
file_ending = ".conllx"
if not fine_grained_deprels:
file_ending = ".coarse_deprels.conllx"
if lang == "Czech" and len(files) > 1:
file_name = "cs-ud-train" + file_ending
else:
file_name = files[0].split("/")[-1].split(".")[0] + file_ending
outfile = join(project_base, "UD_" + lang, file_name)
trees = udtree.from_files(files)
with open (outfile, "w") as w:
for tree in trees:
for word in tree.sentence_structure:
word['deps'] = None
word['misc'] = None
if not any(tree.postags): # Copy CPOSTAG to POSTAG
tree.postags = tree.cpostags
w.write("\n".join(tree.to_conllx_format(fine_grained_deprels=fine_grained_deprels)) + "\n\n")
for i, head in zip(tree.ids, tree.heads):
if head-1 < 0:
continue
else:
for j, inner_head in zip(tree.ids[i:head-1], tree.heads[i:head-1]):
if inner_head < i or inner_head > head:
is_non_projective = True
return is_non_projective
def get_np_trees(trees):
np = []
for tree in trees:
if is_non_projective(tree):
np.append(tree)
return np
def get_non-projectivity_ratios():
lang_trees = lang_utils.get_ud_paths('../resources/universaldependencies1-2/universal-dependencies-1.2/', type_='train', format_='conllu', coarse=False)
np = dict()
tot = dict()
for lang, path in lang_trees.items():
trees = list(udtree.from_files(path))
tot[lang] = len(trees)
np[lang] = get_np_trees(trees)
rel = {}
for lang, n in zip(np.keys(), np.values()):
rel[lang] = len(n) / tot[lang]