def to_relevance_examples(self,
index_path: str,
is_duo: bool = False) -> List[RelevanceExample]:
loader = MsMarcoPassageLoader(index_path)
example_map = {}
for (qid, text, rel_cands), cands in self.query_passage_tuples():
if qid not in example_map:
example_map[qid] = [convert_to_unicode(text), [], [], []]
example_map[qid][1].append([cand for cand in cands][0])
try:
passages = [loader.load_passage(cand) for cand in cands]
example_map[qid][2].append([
convert_to_unicode(passage.all_text)
for passage in passages
][0])
except ValueError:
logging.warning(f'Skipping {passages}')
continue
example_map[qid][3].append(cands[0] in rel_cands)
mean_stats = defaultdict(list)
for ex in self.examples:
int_rels = np.array(list(map(int, example_map[ex.qid][3])))
p = int_rels.sum() / (len(ex.candidates) -
1) if is_duo else int_rels.sum()
mean_stats['Random P@1'].append(np.mean(int_rels))
n = len(ex.candidates) - p
N = len(ex.candidates)
if len(ex.candidates) <= 1000:
mean_stats['Random R@1000'].append(1 if 1 in int_rels else 0)
numer = np.array(
[sp.comb(n, i) / (N - i)
for i in range(0, n + 1) if i != N]) * p
if n == N:
numer = np.append(numer, 0)
denom = np.array([sp.comb(N, i) for i in range(0, n + 1)])
rr = 1 / np.arange(1, n + 2)
rmrr = np.sum(numer * rr / denom)
mean_stats['Random MRR'].append(rmrr)
rmrr10 = np.sum(numer[:10] * rr[:10] / denom[:10])
mean_stats['Random MRR@10'].append(rmrr10)
ex_index = len(ex.candidates)
for rel_cand in ex.relevant_candidates:
if rel_cand in ex.candidates:
ex_index = min(ex.candidates.index(rel_cand), ex_index)
mean_stats['Existing MRR'].append(
1 / (ex_index + 1) if ex_index < len(ex.candidates) else 0)
mean_stats['Existing MRR@10'].append(1 /
(ex_index +
1) if ex_index < 10 else 0)
for k, v in mean_stats.items():
logging.info(f'{k}: {np.mean(v)}')
return [
RelevanceExample(
Query(text=query_text, id=qid),
list(
map(lambda s: Text(s[1], dict(docid=s[0])),
zip(cands, cands_text))), rel_cands)
for qid, (query_text, cands, cands_text,
rel_cands) in example_map.items()
]
def to_relevance_examples(self, index_path: str) -> List[RelevanceExample]:
loader = Cord19AbstractLoader(index_path)
example_map = {}
for (qid, text,
rel_cands), cands in tqdm(self.query_document_tuples()):
if qid not in example_map:
example_map[qid] = [convert_to_unicode(text), [], [], [], []]
example_map[qid][1].append([cand for cand in cands][0])
try:
passages = [loader.load_document(cand) for cand in cands]
# Sometimes this abstract is empty.
example_map[qid][2].append([
convert_to_unicode(passage.abstract)
for passage in passages
][0])
example_map[qid][4].append([
convert_to_unicode(passage.title) for passage in passages
][0])
except ValueError as e:
logging.error(e)
logging.warning('Skipping passages')
continue
example_map[qid][3].append(cands[0] in rel_cands)
mean_stats = defaultdict(list)
for ex in self.examples:
int_rels = np.array(list(map(int, example_map[ex.qid][3])))
p = int(int_rels.sum())
mean_stats['Expected P@1 for Random Ordering'].append(
np.mean(int_rels))
n = len(ex.candidates) - p
N = len(ex.candidates)
if len(ex.candidates) <= 1000:
mean_stats['Expected R@1000 for Random Ordering'].append(
1 if 1 in int_rels else 0)
numer = np.array(
[sp.comb(n, i) / (N - i)
for i in range(0, n + 1) if i != N]) * p
if n == N:
numer = np.append(numer, 0)
denom = np.array([sp.comb(N, i) for i in range(0, n + 1)])
rr = 1 / np.arange(1, n + 2)
rmrr = np.sum(numer * rr / denom)
mean_stats['Expected MRR for Random Ordering'].append(rmrr)
rmrr10 = np.sum(numer[:10] * rr[:10] / denom[:10])
mean_stats['Expected MRR@10 for Random Ordering'].append(rmrr10)
ex_index = len(ex.candidates)
for rel_cand in ex.relevant_candidates:
if rel_cand in ex.candidates:
ex_index = min(ex.candidates.index(rel_cand), ex_index)
mean_stats['Existing MRR'].append(
1 / (ex_index + 1) if ex_index < len(ex.candidates) else 0)
mean_stats['Existing MRR@10'].append(1 /
(ex_index +
1) if ex_index < 10 else 0)
for k, v in mean_stats.items():
logging.info(f'{k}: {np.mean(v)}')
rel = [
RelevanceExample(
Query(text=query_text, id=qid),
list(
map(lambda s: Text(s[1], dict(docid=s[0]), title=s[2]),
zip(cands, cands_text, title))), rel_cands)
for qid, (query_text, cands, cands_text, rel_cands,
title) in example_map.items()
]
return rel
