def resolve_evidences(evidences, t2l2s, actual=True):
"""
actual: [[[id, id, "title", linum], [id, id, "title", linum]], [[id, id, "title", linum]], ... ]
not actual ( predicted ):
[(title, linum), (title, linum), ...]
"""
if actual:
evidence_sentences = list()
for evidence_set in evidences:
evidence_linum = [(title, linum)
for _, _, title, linum in evidence_set]
for title, _ in evidence_linum:
if title is not None and title not in t2l2s:
pass
# print(title, len(evidences), evidences)
evidence_linum = [(title, linum)
for _, _, title, linum in evidence_set
if title in t2l2s]
# continue if evidence_linum is empty
if not evidence_linum:
evidence_sentences.append(["**Not Found**"])
continue
sentences = get_evidence_sentence_list(evidence_linum, t2l2s)
evidence_sentences.append(sentences)
else:
evidence_sentences = get_evidence_sentence_list(evidences, t2l2s)
return evidence_sentences
def read_ir_result(path, prependlinum=False, concatev=False):
"""
Returns
instances: list of dictionary
update instance['predicted_sentences'] with list of evidences (list of str)
"""
instances = read_jsonl(path)
t2jnum = titles_to_jsonl_num(
wikipedia_dir=abs_path("data/wiki-pages/wiki-pages/"),
doctitles=abs_path("data/doctitles"))
titles = list()
# make list of titles
for instance in instances:
titles.extend([title for title, _ in instance["predicted_sentences"]])
# load title2line2sentences
t2l2s = load_doclines(titles, t2jnum)
for instance in instances:
if concatev:
instance["evidence"] = [" ".join(get_evidence_sentence_list(
instance["predicted_sentences"], t2l2s, prependlinum=prependlinum))]
else:
instance["evidence"] = get_evidence_sentence_list(
instance["predicted_sentences"], t2l2s, prependlinum=prependlinum)
return instances
def _convert_instance(instance, t2l2s, prependlinum, prependtitle, use_ir_prediction):
"""convert single instance to either one or multiple instances
Args
instance: instance of FEVER dataset.
t2l2s: output of titles_to_jsonl_num
Returns
list of converted instances
"""
def _evidence_format(evidences):
"""return evidence sentence from (possibly) multiple evidence sentences"""
return " ".join(evidences)
converted_instances = list()
# assert instance["evidence"] == [[[hoge, hoge, title, linum], [hoge, hoge, title, linum]], [[..],[..],..], ...]
if use_ir_prediction:
evidence_linum = [(title, linum) for title, linum in instance["predicted_sentences"]
if title in t2l2s]
for eidx, (title, linum) in enumerate(evidence_linum):
converted_instances.append(
snli_format(
id="{}-{}".format(instance["id"], str(eidx)),
pair_id="{}-{}".format(instance["id"], str(eidx)),
label=convert_label(instance["label"]),
evidence=_evidence_format(
get_evidence_sentence_list(
[(title, linum)], t2l2s, prependlinum=prependlinum, prependtitle=prependtitle)),
claim=instance["claim"]))
else:
for eidx, evidence_set in enumerate(instance["evidence"]):
evidence_linum = [(title, linum)
for _, _, title, linum in evidence_set
if title in t2l2s]
# continue if evidence_linum is empty
if not evidence_linum:
continue
converted_instances.append(
snli_format(
id="{}-{}".format(instance["id"], str(eidx)),
pair_id="{}-{}".format(instance["id"], str(eidx)),
label=convert_label(instance["label"]),
evidence=_evidence_format(
get_evidence_sentence_list(
evidence_linum, t2l2s, prependlinum=prependlinum, prependtitle=prependtitle)),
claim=instance["claim"]))
return converted_instances
def read_ir_result(path,
n_sentences=5,
prependlinum=False,
prependtitle=False,
concatev=False):
"""
Returns
instances: list of dictionary
update instance['predicted_sentences'] with list of evidences (list of str)
"""
short_evidences_counter = 0
instances = read_jsonl(path)
# only read n_sentences
for instance in instances:
if len(instance["predicted_sentences"]) < n_sentences:
short_evidences_counter += 1
instance["predicted_sentences"] = instance[
"predicted_sentences"][:n_sentences]
print("short_evidences: {} / {}".format(short_evidences_counter,
len(instances)))
t2jnum = titles_to_jsonl_num(
wikipedia_dir=abs_path("data/wiki-pages/wiki-pages/"),
doctitles=abs_path("data/doctitles"))
titles = list()
# make list of titles
for instance in instances:
titles.extend([title for title, _ in instance["predicted_sentences"]])
# load title2line2sentences
t2l2s = load_doclines(titles, t2jnum)
for instance in instances:
if concatev:
instance["evidence"] = [
" ".join(
get_evidence_sentence_list(instance["predicted_sentences"],
t2l2s,
prependlinum=prependlinum,
prependtitle=prependtitle))
]
else:
instance["evidence"] = get_evidence_sentence_list(
instance["predicted_sentences"],
t2l2s,
prependlinum=prependlinum,
prependtitle=prependtitle)
return instances
def read_ir_result(path, n_sentences=5):
'''
读取句子检索的结果
'''
short_evidences_counter = 0
instances = read_jsonl(path)
for instance in instances:
if len(instance['predicted_sentences']) < n_sentences:
short_evidences_counter += 1
instance['predicted_sentences'] = instance[
'predicted_sentences'][:n_sentences] # 只保留前 n 个句子
print('short_evidences: {} / {}'.format(short_evidences_counter,
len(instances)))
t2jnum = titles_to_jsonl_num(
wikipedia_dir=abs_path('data/wiki-pages/'),
doctitles=abs_path('data/preprocessed_data/doctitles'))
titles = list()
# 获取所有标题的列表
for instance in instances:
titles.extend([title for title, _ in instance['predicted_sentences']])
t2l2s = load_doclines(titles, t2jnum)
# 证据语句
for instance in instances:
instance['evidence'] = get_evidence_sentence_list(
instance['predicted_sentences'], t2l2s)
return instances