def eval_trees(silver_tree,
gold_tree,
rst_span_perf,
rst_nuc_perf,
rst_rel_perf,
use_parseval=False):
# RST-Parseval
met = Metrics(levels=['span', 'nuclearity', 'relation'],
use_parseval=use_parseval)
met.eval(gold_tree, silver_tree)
assert met.span_perf.hit_num / met.span_num == rst_span_perf
assert met.nuc_perf.hit_num / met.span_num == rst_nuc_perf
assert met.rela_perf.hit_num / met.span_num == rst_rel_perf
def eval_trees(silver_tree, gold_tree, rst_span_perf, rst_nuc_perf, span_perf, nuc_perf):
# RST-Parseval
met = Metrics(use_parseval=False, levels=['span', 'nuclearity'])
met.eval(silver_tree, gold_tree)
assert met.span_perf.hit_num / met.span_num == rst_span_perf
assert met.nuc_perf.hit_num / met.span_num == rst_nuc_perf
# Parseval
met = Metrics(use_parseval=True, levels=['span', 'nuclearity'])
met.eval(silver_tree, gold_tree)
assert met.span_perf.hit_num / met.span_num == span_perf
assert met.nuc_perf.hit_num / met.span_num == nuc_perf
def eval_parser(self,
dev_data=None,
path='./examples',
use_parseval=False):
""" Test the parsing performance"""
# Evaluation
met = Metrics(use_parseval, levels=['span', 'nuclearity'])
# ----------------------------------------
# Read all files from the given path
if dev_data is None:
eval_list = [
os.path.join(path, fname) for fname in os.listdir(path)
if fname.endswith('.merge')
]
else:
eval_list = dev_data
pred_forms = []
gold_forms = []
total_cost = 0
for eval_instance in eval_list:
# ----------------------------------------
# Read *.merge file
doc = Doc()
if dev_data is not None:
gold_rst = eval_instance
doc = eval_instance.doc
else:
doc.read_from_fmerge(eval_instance)
eval_instance = eval_instance.replace('.dis', '.merge')
fdis = eval_instance.replace('.merge', '.dis')
gold_rst = RstTree(fdis, eval_instance)
gold_rst.build()
_, gold_action_seq = gold_rst.generate_action_samples()
gold_action_seq = list(
map(lambda x: self.data_helper.action_map[x], gold_action_seq))
pred_rst, cost = self.parser.sr_parse(
[doc], [torch.cuda.LongTensor(gold_action_seq)],
None,
is_train=False)
total_cost += cost
pred_rst = pred_rst[0]
# ----------------------------------------
# Evaluate with gold RST tree
met.eval(gold_rst, pred_rst)
print("Total cost: ", total_cost)
met.report()
def eval_parser(self,
dev_data=None,
path='./examples',
save_preds=True,
use_parseval=False):
""" Test the parsing performance"""
# Evaluation
met = Metrics(levels=['span', 'nuclearity', 'relation'],
use_parseval=use_parseval)
# ----------------------------------------
# Read all files from the given path
if dev_data is None:
dev_data = [
os.path.join(path, fname) for fname in os.listdir(path)
if fname.endswith('.dis')
]
total_cost = 0
for eval_instance in dev_data:
# ----------------------------------------
fmerge = eval_instance.replace('.dis', '.merge')
doc = Doc()
doc.read_from_fmerge(fmerge)
gold_rst = RstTree(eval_instance, fmerge)
gold_rst.build()
# tok_edus = [nltk.word_tokenize(edu) for edu in doc.doc_edus]
tok_edus = [edu.split(" ") for edu in doc.doc_edus]
tokens = flatten(tok_edus)
coref_document = Document(raw_text=None,
tokens=tokens,
sents=tok_edus,
corefs=[],
speakers=["0"] * len(tokens),
genre="nw",
filename=None)
coref_document.token_dict = doc.token_dict
coref_document.edu_dict = doc.edu_dict
doc = coref_document
gold_action_seq, gold_rel_seq = gold_rst.decode_rst_tree()
gold_action_seq = [action_map[x] for x in gold_action_seq]
gold_relation_seq = [
relation_map[x.lower()] for x in gold_rel_seq if x is not None
]
pred_rst, cost = self.parser.sr_parse(
doc, torch.cuda.LongTensor(gold_action_seq),
torch.cuda.LongTensor(gold_relation_seq))
total_cost += cost
if save_preds:
if not os.path.isdir('../data/predicted_trees'):
os.mkdir('../data/predicted_trees')
filename = eval_instance.split(os.sep)[-1]
filepath = f'../data/predicted_trees/{self.config[MODEL_NAME]}_{filename}'
pred_brackets = pred_rst.bracketing()
# Write brackets into file
Evaluator.writebrackets(filepath, pred_brackets)
# ----------------------------------------
# Evaluate with gold RST tree
met.eval(gold_rst, pred_rst)
print("Total cost: ", total_cost)
if use_parseval:
print("Reporting original Parseval metric.")
else:
print("Reporting RST Parseval metric.")
met.report()
def eval_parser(self,
path='./examples',
report=False,
bcvocab=None,
draw=True):
""" Test the parsing performance"""
# Evaluation
met = Metrics(levels=['span', 'nuclearity', 'relation'])
# ----------------------------------------
# Read all files from the given path
doclist = [
os.path.join(path, fname) for fname in os.listdir(path)
if fname.endswith('.merge')
]
pred_forms = []
gold_forms = []
depth_per_relation = {}
for fmerge in doclist:
# ----------------------------------------
# Read *.merge file
doc = Doc()
doc.read_from_fmerge(fmerge)
# ----------------------------------------
# Parsing
pred_rst = self.parser.sr_parse(doc, bcvocab)
if draw:
pred_rst.draw_rst(fmerge.replace(".merge", ".ps"))
# Get brackets from parsing results
pred_brackets = pred_rst.bracketing()
fbrackets = fmerge.replace('.merge', '.brackets')
# Write brackets into file
Evaluator.writebrackets(fbrackets, pred_brackets)
# ----------------------------------------
# Evaluate with gold RST tree
if report:
fdis = fmerge.replace('.merge', '.dis')
gold_rst = RstTree(fdis, fmerge)
gold_rst.build()
met.eval(gold_rst, pred_rst)
for node in pred_rst.postorder_DFT(pred_rst.tree, []):
pred_forms.append(node.form)
for node in gold_rst.postorder_DFT(gold_rst.tree, []):
gold_forms.append(node.form)
nodes = gold_rst.postorder_DFT(gold_rst.tree, [])
inner_nodes = [
node for node in nodes
if node.lnode is not None and node.rnode is not None
]
for idx, node in enumerate(inner_nodes):
relation = node.rnode.relation if node.form == 'NS' else node.lnode.relation
rela_class = RstTree.extract_relation(relation)
if rela_class in depth_per_relation:
depth_per_relation[rela_class].append(node.depth)
else:
depth_per_relation[rela_class] = [node.depth]
lnode_text = ' '.join([
gold_rst.doc.token_dict[tid].word
for tid in node.lnode.text
])
lnode_lemmas = ' '.join([
gold_rst.doc.token_dict[tid].lemma
for tid in node.lnode.text
])
rnode_text = ' '.join([
gold_rst.doc.token_dict[tid].word
for tid in node.rnode.text
])
rnode_lemmas = ' '.join([
gold_rst.doc.token_dict[tid].lemma
for tid in node.rnode.text
])
# if rela_class == 'Topic-Change':
# print(fmerge)
# print(relation)
# print(lnode_text)
# print(rnode_text)
# print()
if report:
met.report()
def eval_parser(self, data_dir, output_dir='./examples', report=False, bcvocab=None, draw=True, isFlat=False):
""" Test the parsing performance"""
# Evaluation
met = Metrics(levels=['span', 'nuclearity', 'relation'])
# ----------------------------------------
# Read all files from the given path
with open(os.path.join(output_dir, "Treebank/TEST", "processed_data.p"), 'rb') as file:
doclist = pickle.load(file)
fnames = [fn for fn in os.listdir(data_dir) if fn.endswith(".out")]
pred_forms = []
gold_forms = []
depth_per_relation = {}
for lines, fname in zip(doclist, fnames):
# ----------------------------------------
# Read *.merge file
doc = Doc()
doc.read_from_fmerge(lines)
fout = os.path.join(data_dir, fname)
print(fout)
# ----------------------------------------
# Parsing
print("************************ predict rst ************************")
pred_rst = self.parser.sr_parse(doc, self.isFlat,bcvocab)
if draw:
pred_rst.draw_rst(fout+'.ps')
# Get brackets from parsing results
pred_brackets = pred_rst.bracketing(self.isFlat)
fbrackets = fout+'.brackets'
# Write brackets into file
Evaluator.writebrackets(fbrackets, pred_brackets)
# ----------------------------------------
# Evaluate with gold RST tree
if report:
print("************************ gold rst ************************")
fdis = fout+'.dis'
gold_rst = RstTree(fdis, lines, isFlat)
gold_rst.build()
met.eval(gold_rst, pred_rst, self.isFlat)
if isFlat:
for node in pred_rst.postorder_flat_DFT(pred_rst.tree, []):
pred_forms.append(node.form)
for node in gold_rst.postorder_flat_DFT(gold_rst.tree, []):
gold_forms.append(node.form)
nodes = gold_rst.postorder_flat_DFT(gold_rst.tree, [])
else:
for node in pred_rst.postorder_DFT(pred_rst.tree, []):
pred_forms.append(node.form)
for node in gold_rst.postorder_DFT(gold_rst.tree, []):
gold_forms.append(node.form)
nodes = gold_rst.postorder_DFT(gold_rst.tree, [])
inner_nodes = [node for node in nodes if node.lnode is not None and node.rnode is not None]
for idx, node in enumerate(inner_nodes):
relation = node.rnode.relation if node.form == 'NS' else node.lnode.relation
rela_class = RstTree.extract_relation(relation)
if rela_class in depth_per_relation:
depth_per_relation[rela_class].append(node.depth)
else:
depth_per_relation[rela_class] = [node.depth]
if report:
met.report()