def get_paragraph_forward_pair(tag,
ruleterm_doc_results,
is_training,
debug=False,
ignore_non_verifiable=False):
if tag == 'dev':
d_list = common.load_jsonl(config.FEVER_DEV)
elif tag == 'train':
d_list = common.load_jsonl(config.FEVER_TRAIN)
elif tag == 'test':
d_list = common.load_jsonl(config.FEVER_TEST)
else:
raise ValueError(f"Tag:{tag} not supported.")
if debug:
d_list = d_list[:100]
ruleterm_doc_results = ruleterm_doc_results[:100]
ruleterm_doc_results_dict = list_dict_data_tool.list_to_dict(
ruleterm_doc_results, 'id')
db_cursor = fever_db.get_cursor()
fitems = build_full_wiki_document_forward_item(ruleterm_doc_results_dict,
d_list, is_training,
db_cursor,
ignore_non_verifiable)
return fitems
def get_sentences(tag, is_training, debug=False):
if tag == 'dev':
d_list = common.load_jsonl(config.FEVER_DEV)
elif tag == 'train':
d_list = common.load_jsonl(config.FEVER_TRAIN)
elif tag == 'test':
d_list = common.load_jsonl(config.FEVER_TEST)
else:
raise ValueError(f"Tag:{tag} not supported.")
if debug:
# d_list = d_list[:10]
d_list = d_list[:50]
# d_list = d_list[:200]
doc_results = common.load_jsonl(
config.RESULT_PATH /
f"doc_retri_results/fever_results/merged_doc_results/m_doc_{tag}.jsonl"
)
doc_results_dict = list_dict_data_tool.list_to_dict(doc_results, 'id')
fever_db_cursor = fever_db.get_cursor(config.FEVER_DB)
forward_items = build_full_wiki_document_forward_item(
doc_results_dict,
d_list,
is_training=is_training,
db_cursor=fever_db_cursor)
return forward_items
def get_inference_pair(tag,
is_training,
sent_result_path,
debug_num=None,
evidence_filtering_threshold=0.01):
# sent_result_path = ""
if tag == 'dev':
d_list = common.load_jsonl(config.FEVER_DEV)
elif tag == 'train':
d_list = common.load_jsonl(config.FEVER_TRAIN)
elif tag == 'test':
d_list = common.load_jsonl(config.FEVER_TEST)
else:
raise ValueError(f"Tag:{tag} not supported.")
if debug_num is not None:
# d_list = d_list[:10]
d_list = d_list[:50]
# d_list = d_list[:200]
d_dict = list_dict_data_tool.list_to_dict(d_list, 'id')
threshold_value = evidence_filtering_threshold
# sent_list = common.load_jsonl(
# config.RESULT_PATH / "doc_retri_results/fever_results/sent_results/4-14-sent_results_v0/train_sent_results.jsonl")
# sent_list = common.load_jsonl(
# config.RESULT_PATH / "doc_retri_results/fever_results/sent_results/4-14-sent_results_v0/i(5000)|e(0)|s01(0.9170917091709171)|s05(0.8842384238423843)|seed(12)_dev_sent_results.json")
# debug_num = None if not debug else 2971
# debug_num = None
if isinstance(sent_result_path, Path):
sent_list = common.load_jsonl(sent_result_path, debug_num)
elif isinstance(sent_result_path, list):
sent_list = sent_result_path
else:
raise ValueError(
f"{sent_result_path} is not of a valid argument type which should be [list, Path]."
)
list_dict_data_tool.append_subfield_from_list_to_dict(sent_list,
d_dict,
'oid',
'fid',
check=True)
filltered_sent_dict = select_top_k_and_to_results_dict(
d_dict, top_k=5, threshold=threshold_value)
list_dict_data_tool.append_item_from_dict_to_list(
d_list, filltered_sent_dict, 'id',
['predicted_evidence', 'predicted_scored_evidence'])
fever_db_cursor = fever_db.get_cursor(config.FEVER_DB)
forward_items = build_nli_forward_item(d_list,
is_training=is_training,
db_cursor=fever_db_cursor)
return forward_items, d_list
def get_nli_pair(tag,
is_training,
sent_level_results_list,
debug=None,
sent_top_k=5,
sent_filter_value=0.05):
if tag == 'dev':
d_list = common.load_jsonl(config.FEVER_DEV)
elif tag == 'train':
d_list = common.load_jsonl(config.FEVER_TRAIN)
elif tag == 'test':
d_list = common.load_jsonl(config.FEVER_TEST)
else:
raise ValueError(f"Tag:{tag} not supported.")
if debug:
d_list = d_list[:100]
# sent_dict = list_dict_data_tool.list_to_dict(sent_level_results_list):
d_dict = list_dict_data_tool.list_to_dict(d_list, 'id')
if debug:
id_set = set([item['id'] for item in d_list])
new_sent_list = []
for item in sent_level_results_list:
if item["qid"] in id_set:
new_sent_list.append(item)
sent_level_results_list = new_sent_list
list_dict_data_tool.append_subfield_from_list_to_dict(
sent_level_results_list, d_dict, 'qid', 'fid', check=True)
filltered_sent_dict = select_top_k_and_to_results_dict(
d_dict,
score_field_name='prob',
top_k=sent_top_k,
filter_value=sent_filter_value,
result_field='predicted_evidence')
list_dict_data_tool.append_item_from_dict_to_list_hotpot_style(
d_list, filltered_sent_dict, 'id',
['predicted_evidence', 'selected_scored_results'])
fever_db_cursor = fever_db.get_cursor(config.FEVER_DB)
forward_items = build_nli_forward_item(d_list,
is_training=is_training,
db_cursor=fever_db_cursor)
return forward_items, d_list
def get_sentence_forward_pair(tag,
ruleterm_doc_results,
is_training,
debug=False,
ignore_non_verifiable=False,
top_k=5,
filter_value=0.005):
if tag == 'dev':
d_list = common.load_jsonl(config.FEVER_DEV)
elif tag == 'train':
d_list = common.load_jsonl(config.FEVER_TRAIN)
elif tag == 'test':
d_list = common.load_jsonl(config.FEVER_TEST)
else:
raise ValueError(f"Tag:{tag} not supported.")
if debug:
d_list = d_list[:100]
ruleterm_doc_results = ruleterm_doc_results[:100]
# ruleterm_doc_results_dict = list_dict_data_tool.list_to_dict(ruleterm_doc_results, 'id')
d_o_dict = list_dict_data_tool.list_to_dict(d_list, 'id')
copied_d_o_dict = copy.deepcopy(d_o_dict)
# copied_d_list = copy.deepcopy(d_list)
list_dict_data_tool.append_subfield_from_list_to_dict(ruleterm_doc_results,
copied_d_o_dict,
'qid',
'fid',
check=True)
cur_results_dict_filtered = select_top_k_and_to_results_dict(
copied_d_o_dict,
score_field_name='prob',
top_k=top_k,
filter_value=filter_value)
db_cursor = fever_db.get_cursor()
fitems = build_full_wiki_sentence_forward_item(cur_results_dict_filtered,
d_list, is_training,
db_cursor,
ignore_non_verifiable)
return fitems
random.shuffle(pos_items)
random.shuffle(neg_items)
neg_sample_count = int(pos_count / ratio)
sampled_neg = neg_items[:neg_sample_count]
print(f"After Sampling, we have {pos_count}/{len(sampled_neg)} (pos/neg).")
sampled_list = sampled_neg + pos_items
random.shuffle(sampled_list)
return sampled_list
if __name__ == '__main__':
d_list = common.load_jsonl(config.FEVER_DEV)
doc_results = common.load_jsonl(
config.RESULT_PATH /
"doc_retri_results/fever_results/merged_doc_results/m_doc_dev.jsonl")
doc_results_dict = list_dict_data_tool.list_to_dict(doc_results, 'id')
fever_db_cursor = fever_db.get_cursor(config.FEVER_DB)
forward_items = build_full_wiki_document_forward_item(
doc_results_dict, d_list, is_training=False, db_cursor=fever_db_cursor)
# print(forward_items)
# for item in forward_items:
# if item['s_labels'] == 'true':
# print(item['query'], item['context'], item['sid'], item['cid'], item['fid'], item['s_labels'])
print(len(forward_items))
# down_sample_neg(forward_items, ratio=0.2)
def adv_simi_sample_with_prob_v1_1(input_file,
additional_file,
prob_dict_file,
tokenized=False):
cursor = fever_db.get_cursor()
d_list = common.load_jsonl(input_file)
if prob_dict_file is None:
prob_dict_file = dict()
if isinstance(additional_file, list):
additional_d_list = additional_file
else:
additional_d_list = common.load_jsonl(additional_file)
additional_data_dict = dict()
for add_item in additional_d_list:
additional_data_dict[add_item['id']] = add_item
sampled_data_list = []
count = 0
for item in tqdm(d_list):
# e_list = check_sentences.check_and_clean_evidence(item)
sampled_e_list, flags = sample_additional_data_for_item_v1_1(
item, additional_data_dict)
# print(flags)
for i, (sampled_evidence,
flag) in enumerate(zip(sampled_e_list, flags)):
# Do not copy, might change in the future for error analysis
# new_item = copy.deepcopy(item)
new_item = dict()
# print(new_item['claim'])
# print(e_list)
# print(sampled_evidence)
# print(flag)
evidence_text_list = evidence_list_to_text_list(
cursor,
sampled_evidence,
contain_head=True,
id_tokenized=tokenized)
evidences = sorted(sampled_evidence, key=lambda x: (x[0], x[1]))
item_id = int(item['id'])
evidence_text_list_with_prob = []
for text, (doc_id, ln) in zip(evidence_text_list, evidences):
ssid = (int(item_id), doc_id, int(ln))
if ssid not in prob_dict_file:
count += 1
# print("Some sentence pair don't have 'prob'.")
prob = 0.5
else:
prob = prob_dict_file[ssid]['prob']
assert item['claim'] == prob_dict_file[ssid]['claim']
evidence_text_list_with_prob.append((text, prob))
new_item['id'] = str(item['id']) + '#' + str(i)
if tokenized:
new_item['claim'] = item['claim']
else:
raise NotImplemented("Non tokenized is not implemented.")
# new_item['claim'] = ' '.join(easy_tokenize(item['claim']))
new_item['evid'] = evidence_text_list_with_prob
new_item['verifiable'] = item['verifiable']
new_item['label'] = item['label']
# print("C:", new_item['claim'])
# print("E:", new_item['evid'])
# print("L:", new_item['label'])
# print()
sampled_data_list.append(new_item)
cursor.close()
print(count)
return sampled_data_list
def select_sent_with_prob_for_eval(input_file,
additional_file,
prob_dict_file,
tokenized=False,
pipeline=False):
"""
This method select sentences with upstream sentence retrieval.
:param input_file: This should be the file with 5 sentences selected.
:return:
"""
cursor = fever_db.get_cursor()
if prob_dict_file is None:
prob_dict_file = dict()
if isinstance(additional_file, list):
additional_d_list = additional_file
else:
additional_d_list = common.load_jsonl(additional_file)
additional_data_dict = dict()
for add_item in additional_d_list:
additional_data_dict[add_item['id']] = add_item
d_list = common.load_jsonl(input_file)
for item in tqdm(d_list):
e_list = additional_data_dict[item['id']]['predicted_sentids']
if not pipeline:
assert additional_data_dict[item['id']]['label'] == item['label']
assert additional_data_dict[
item['id']]['verifiable'] == item['verifiable']
assert additional_data_dict[item['id']]['id'] == item['id']
pred_evidence_list = []
for i, cur_e in enumerate(e_list):
doc_id = cur_e.split(fever_scorer.SENT_LINE)[0]
ln = int(cur_e.split(
fever_scorer.SENT_LINE)[1]) # Important changes Bugs: July 21
pred_evidence_list.append((doc_id, ln))
pred_evidence = check_sentences.Evidences(pred_evidence_list)
evidence_text_list = evidence_list_to_text_list(cursor,
pred_evidence,
contain_head=True,
id_tokenized=tokenized)
evidences = sorted(pred_evidence, key=lambda x: (x[0], x[1]))
item_id = int(item['id'])
evidence_text_list_with_prob = []
for text, (doc_id, ln) in zip(evidence_text_list, evidences):
ssid = (item_id, doc_id, int(ln))
if ssid not in prob_dict_file:
# print("Some sentence pair don't have 'prob'.")
prob = 0.5
else:
prob = prob_dict_file[ssid]['prob']
assert item['claim'] == prob_dict_file[ssid]['claim']
evidence_text_list_with_prob.append((text, prob))
if tokenized:
pass
else:
raise NotImplemented("Non tokenized is not implemented.")
# item['claim'] = ' '.join(easy_tokenize(item['claim']))
item['evid'] = evidence_text_list_with_prob
item['predicted_evidence'] = convert_evidence2scoring_format(e_list)
item['predicted_sentids'] = e_list
# This change need to be saved.
# item['predicted_label'] = additional_data_dict[item['id']]['label']
return d_list
