def preprocess_claim_with_doc(claim_with_docs: tuple) -> list:
claim_id = claim_with_docs[0]
# remove any NOT_VERIFIABLE claims that were processed earlier
if not claim_is_verifiable(claim_id, dataset=args.dataset):
return []
claim = get_claim(claim_id, dataset=args.dataset)
evidence_map = get_evidence_page_line_map(claim_id, args.dataset)
print('Preprocessing docs for claim [{}]: {}'.format(
claim_id, evidence_map.keys()))
preprocessed_pairs = []
for page_id, relevant_line_ids in evidence_map.items():
wiki_page = retrieve_wiki_page(page_id)
for line_id in relevant_line_ids:
# add the relevant claim/sentence pair...
positive_line = wiki_page.lines[line_id]
positive_input = transform_LR_input(claim, positive_line.text)
preprocessed_pairs.append(
(claim_id, page_id, line_id, positive_input, 1))
# ...and, to keep it balanced, one irrelevant sample
negative_line = get_irrelevant_line(wiki_page, relevant_line_ids)
negative_input = transform_LR_input(claim, negative_line.text)
preprocessed_pairs.append(
(claim_id, page_id, negative_line.id, negative_input, 0))
return preprocessed_pairs
def preprocess_claim(claim_row: pd.Series) -> list:
claim_id, verifiable, label, claim, evidence = claim_row[1].values
if not verifiable == 'VERIFIABLE':
return []
print('Preprocessing docs for claim [{}]'.format(claim_id))
# output will be the same for all evidence items belonging to this claim
#label = 'entails' if label == 'SUPPORTS' else 'neutral'
label = 1 if label == 'SUPPORTS' else 0
preprocessed_pairs = []
evidence_map = get_evidence_page_line_map(claim_id, args.dataset)
evidence_sentences = []
for page_id, relevant_line_ids in evidence_map.items():
wiki_page = retrieve_wiki_page(page_id)
evidence_sentences.extend(
[wiki_page.lines[id].text for id in relevant_line_ids])
premise = ' '.join(evidence_sentences)
premise = recreate_punctuation_in_doc_text(premise)
hypothesis = recreate_punctuation_in_doc_text(claim)
pair = {
'label': label,
'sentence1': premise,
'sentence2': hypothesis,
'claim_id': claim_id
}
preprocessed_pairs.append(pair)
return preprocessed_pairs
def preprocess_claim(claim_row: pd.Series) -> list:
claim_id, verifiable, label, claim, evidence = claim_row[1].values
if not verifiable == 'VERIFIABLE':
return []
print('Preprocessing docs for claim [{}]'.format(claim_id))
# output will be the same for all evidence items belonging to this claim
output = 1 if label == 'SUPPORTS' else 0
preprocessed_pairs = []
evidence_map = get_evidence_page_line_map(claim_id, args.dataset)
for page_id, relevant_line_ids in evidence_map.items():
wiki_page = retrieve_wiki_page(page_id)
for line_id in relevant_line_ids:
line = wiki_page.lines[line_id]
line_text = line.text
line_contains_references = 1 if line.anchors else 0
line_position_absolute = line.id
line_position_relative = line.id / len(wiki_page.lines)
num_evidence_docs_for_claim = len(evidence_map.keys())
num_lines_for_evidence = len(relevant_line_ids)
num_coordination_terms_claim = get_num_coordination_terms(
line_text,
preprocess_claim_text(claim).split())
num_coordination_terms_title = get_num_coordination_terms(
line_text, preprocess_doc_title(page_id))
input = transform_NN_input(
claim, line_text, line_contains_references,
line_position_absolute, line_position_relative,
num_evidence_docs_for_claim, num_lines_for_evidence,
num_coordination_terms_claim, num_coordination_terms_title)
preprocessed_pairs.append(
(claim_id, page_id, line_id, input, output))
return preprocessed_pairs
def preprocess_claim(claim_row: pd.Series) -> list:
claim_id, verifiable, label, claim, evidence = claim_row[1].values
if not verifiable == 'VERIFIABLE':
return []
print('Preprocessing docs for claim [{}]'.format(claim_id))
# output will be the same for all evidence items belonging to this claim
output = 1 if label == 'SUPPORTS' else 0
preprocessed_pairs = []
evidence_map = get_evidence_page_line_map(claim_id, args.dataset)
evidence_sentences = []
num_evidence_docs_for_claim = len(evidence_map.keys())
num_references = 0
num_evidence_items = 0
num_evidence_words = 0
num_coordination_terms_evidence_claim = 0
num_coordination_terms_titles_claim = 0
evidence_sentence_positions = []
# concat evidence (can be from multiple wiki_pages and/or lines)
for page_id, relevant_line_ids in evidence_map.items():
wiki_page = retrieve_wiki_page(page_id)
evidence_sentences.extend(
[wiki_page.lines[id].text for id in relevant_line_ids])
# count metrics and subtract features
for line_id in relevant_line_ids:
line = wiki_page.lines[line_id]
line_text = line.text
num_evidence_words += len(line_text.split())
num_references += len(line.anchors)
num_evidence_items += 1
evidence_sentence_positions.append(line.id)
num_coordination_terms_evidence_claim += get_num_coordination_terms(
line_text,
preprocess_claim_text(claim).split())
num_coordination_terms_titles_claim += get_num_coordination_terms(
line_text, preprocess_doc_title(page_id))
combined_evidence = ' '.join(evidence_sentences)
avg_sentence_position = np.mean(evidence_sentence_positions)
input = transform_NN_input(claim, combined_evidence,
num_evidence_docs_for_claim, num_references,
num_evidence_items,
num_coordination_terms_evidence_claim,
num_coordination_terms_titles_claim,
avg_sentence_position, num_evidence_words)
preprocessed_pairs.append((claim_id, input, output))
return preprocessed_pairs
def display_or_store_result(claim: str,
claim_id: int,
result_docs: list,
dir_path: str,
display_only: bool = False):
if display_only:
print(colored('Results for claim "{}":'.format(claim), 'yellow'))
for doc in result_docs:
page_id = doc[0]
wiki_page = retrieve_wiki_page(page_id)
print(wiki_page)
else:
#result_path = '{}{}.jsonl'.format(path, claim_id)
#write_list_to_jsonl(result_path, result_docs)
print(
colored(
'Storing results for claim "{}"\n{}:'.format(
claim, result_docs), 'yellow'))
write_list_to_oneline_csv(dir_path, claim_id, result_docs)
def preprocess_claim(claim_row: pd.Series) -> list:
claim_id, verifiable, label, claim, evidence = claim_row[1].values
if not verifiable == 'VERIFIABLE':
return []
print('Preprocessing docs for claim [{}]'.format(claim_id))
# output will be the same for all evidence items belonging to this claim
output = 1 if label == 'SUPPORTS' else 0
preprocessed_pairs = []
evidence_map = get_evidence_page_line_map(claim_id, args.dataset)
for page_id, relevant_line_ids in evidence_map.items():
wiki_page = retrieve_wiki_page(page_id)
for line_id in relevant_line_ids:
line_text = wiki_page.lines[line_id].text
input = transform_NN_input(claim, line_text)
preprocessed_pairs.append((claim_id, page_id, line_id, input, output))
return preprocessed_pairs
def preprocess_claim_with_doc(claim_with_docs: tuple) -> list:
partial_result = []
claim_id = claim_with_docs[0]
# remove any NOT_VERIFIABLE claims that were processed earlier
if not claim_is_verifiable(claim_id, dataset=args.dataset):
return []
claim = get_claim(claim_id, dataset=args.dataset)
evidence_map = get_evidence_page_line_map(claim_id, args.dataset)
retrieved_doc_ids = set(claim_with_docs[1])
evidence_doc_ids = evidence_map.keys()
docs_for_training = retrieved_doc_ids.union(evidence_doc_ids)
print('Preprocessing docs for claim [{}]: {}'.format(
claim_id, docs_for_training))
docs = [retrieve_wiki_page(id) for id in docs_for_training]
for doc in docs:
partial_result.extend(
preprocess_doc(claim_id, claim, doc, evidence_map))
return partial_result
# where the actual retrieval takes place
import argparse
import time
from dataaccess.access_wiki_page import retrieve_wiki_page
parser = argparse.ArgumentParser()
parser.add_argument("--id",
help="ID of a document to retrieve for test purposes",
required=True)
parser.add_argument(
"--complete",
help="print complete doc text instead of ID + text preview",
action="store_true")
args = parser.parse_args()
if __name__ == '__main__':
if (args.id):
start_time = time.time()
wiki_document = retrieve_wiki_page(args.id)
print('Retrieved document "{}" after {:.5f} seconds'.format(
args.id,
time.time() - start_time))
if (args.complete):
print(wiki_document.text)
else:
print(wiki_document)
else:
print('Please add ID to retrieve')