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 transform_LR_input(claim_text: str, line_text: str, debug: bool = False):
# remove punctuation that are otherwise part of tokens
preprocessed_claim = preprocess_claim_text(claim_text)
# remove artifacts like -LRB- etc.
preprocessed_line = preprocess_doc_text(line_text)
claim_vector = transform_sentence_to_vector(preprocessed_claim, debug)
line_vector = transform_sentence_to_vector(preprocessed_line, debug)
return get_vector_difference(claim_vector, line_vector)
def transform_NN_input(claim_text: str, line_text: str):
# remove punctuation that are otherwise part of tokens
preprocessed_claim = preprocess_claim_text(claim_text)
# remove artifacts like -LRB- etc.
preprocessed_line = preprocess_doc_text(line_text)
claim_vector = transform_sentence_to_vector(preprocessed_claim, args.debug)
line_vector = transform_sentence_to_vector(preprocessed_line, args.debug)
combined_claim_line_vector = get_vector_difference(claim_vector, line_vector)
return combined_claim_line_vector
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 retrieve_documents_for_claim(claim: str, claim_id: int):
print(
colored('Retrieving documents for claim [{}]: "{}"'.format(
claim_id, claim),
attrs=['bold']))
preprocessed_claim = preprocess_claim_text(claim)
claim_terms = process_normalise_tokenise_filter(preprocessed_claim)
# only docs that appear in index for at least one claim term to be considered
doc_candidates = get_candidate_documents_for_claim(claim_terms,
mode='raw_count')
scoring_function = get_query_likelihood_score_no_smoothing
if args.smoothing == 'laplace':
scoring_function = get_query_likelihood_score_laplace_smoothing
if args.smoothing == 'laplace_lindstone':
scoring_function = get_query_likelihood_score_laplace_lindstone_smoothing
if args.smoothing == 'jelinek_mercer':
scoring_function = get_query_likelihood_score_jelinek_mercer_smoothing
if args.smoothing == 'dirichlet':
scoring_function = get_query_likelihood_score_dirichlet_smoothing
# query likelihood scores for each claim-doc combination
docs_with_query_likelihood_scores = [
scoring_function(claim_terms, doc_with_terms)
for doc_with_terms in doc_candidates.items()
]
# zero values lead to random retrievals if all documents evaluate to zero, so might rather want to show no results
if (args.remove_zero_likelihood):
docs_with_query_likelihood_scores = list(
filter(lambda x: x[1] != 0, docs_with_query_likelihood_scores))
# sort by query likelihood and limit to top results
docs_with_query_likelihood_scores.sort(key=itemgetter(1), reverse=True)
result_docs = docs_with_query_likelihood_scores[:
DOCS_TO_RETRIEVE_PER_CLAIM]
result_directory = '{}{}/'.format(RETRIEVED_PROBABILISTIC_DIRECTORY,
args.smoothing or 'no_smoothing')
display_or_store_result(claim, claim_id, result_docs, result_directory,
args.print)
def retrieve_documents_for_claim(claim: str, claim_id: int):
print(colored('Retrieving documents for claim [{}]: "{}"'.format(claim_id, claim), attrs=['bold']))
preprocessed_claim = preprocess_claim_text(claim)
claim_terms = process_normalise_tokenise_filter(preprocessed_claim)
claim_vector = get_tfidf_vector_for_claim(claim_terms)
claim_norm = get_tfidf_vector_norm(claim_terms, args.variant)
# only docs that appear in index for at least one claim term to be considered
doc_candidates = get_candidate_documents_for_claim(claim_terms)
# similarity scores for each claim-doc combination
docs_with_similarity_scores = [
scoring_function(claim_terms, claim_vector, claim_norm, doc_with_terms) for doc_with_terms in
doc_candidates.items()]
# sort by similarity and limit to top results
docs_with_similarity_scores.sort(key=itemgetter(1), reverse=True)
result_docs = docs_with_similarity_scores[:DOCS_TO_RETRIEVE_PER_CLAIM]
display_or_store_result(claim, claim_id, result_docs, RETRIEVED_TFIDF_DIRECTORY, args.print)
def transform_NN_input(
claim_text: str, line_text: str, line_contains_references: int,
line_position_absolute: int, line_position_relative: float,
num_evidence_docs_for_claim: int, num_lines_for_evidence: int,
num_coordination_terms_claim: int, num_coordination_terms_title: int):
# remove punctuation that are otherwise part of tokens
preprocessed_claim = preprocess_claim_text(claim_text)
# remove artifacts like -LRB- etc.
preprocessed_line = preprocess_doc_text(line_text)
claim_vector = transform_sentence_to_vector(preprocessed_claim, args.debug)
line_vector = transform_sentence_to_vector(preprocessed_line, args.debug)
combined_claim_line_vector = get_vector_difference(claim_vector,
line_vector)
additional_features = np.array(
(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))
return np.concatenate((combined_claim_line_vector, additional_features))
