def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
qterms = set(query.tf.keys())
sd1 = ctx.f.closest_grouping_size_and_count(qterms, d1.termseq)
sd2 = ctx.f.closest_grouping_size_and_count(qterms, d2.termseq)
if sd1[0] == sd2[0]:
return prefs.strictlygreater(sd1[1], sd2[1])
else:
return prefs.strictlygreater(sd2[0], sd1[0])
def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
qts = set(query.tf.keys())
o1 = ctx.f.vocab_overlap(qts, d1.tf.keys())
o2 = ctx.f.vocab_overlap(qts, d2.tf.keys())
return prefs.strictlygreater(o2, o1)
def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
terms_to_test = set(query.termseq).intersection(
d1.termseq).intersection(d2.termseq)
avg1 = ctx.f.average_between_qterms(terms_to_test, d1.termseq)
avg2 = ctx.f.average_between_qterms(terms_to_test, d2.termseq)
return prefs.strictlygreater(avg2, avg1)
def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
tf1 = 0
tf2 = 0
for qt in query.termseq:
tf1 += d1.tf[qt]
tf2 += d2.tf[qt]
if not prefs.approximatelyEqual(tf1, tf2):
# at least 10% difference
return prefs.strictlygreater(tf1, tf2)
return 0
def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
fpsum1 = 0
fpsum2 = 0
ts = set(d1.termseq).intersection(d2.termseq)
for t in set(query.termseq):
if t in ts:
fpsum1 += d1.termseq.index(t)
fpsum2 += d2.termseq.index(t)
return prefs.strictlygreater(fpsum2, fpsum1)
def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
qterms = list(query.tf.keys())
ssum = np.zeros(len(qterms))
for i1, i2 in itertools.combinations(np.arange(len(qterms)), 2):
sim = ctx.f.synset_similarity(qterms[i1], qterms[i2])
ssum[i1] += sim
ssum[i2] += sim
tmax = qterms[np.argmax(ssum)]
return prefs.strictlygreater(d1.tf[tmax], d2.tf[tmax])
def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
terms = query.termseq
tl = len(terms)
def find_idx(seq: typing.Sequence[str]):
for i in (i for i, e in enumerate(seq) if e == terms[0]):
if i + tl <= len(seq) and seq[i:i + tl] == terms:
return i
return float('inf')
return prefs.strictlygreater(find_idx(d2.termseq),
find_idx(d1.termseq))
def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
sd1 = 0
sd2 = 0
for qt1, qt2 in itertools.combinations(query.tf.keys(), 2):
if prefs.approximatelyEqual(ctx.f.td(qt1), ctx.f.td(qt2)):
d1q1 = d1.tf[qt1]
d2q1 = d2.tf[qt1]
d1q2 = d1.tf[qt2]
d2q2 = d2.tf[qt2]
sd1 += (d2q1 == d1q1 +
d1q2) and (d2q2 == 0) and (d1q1 != 0) and (d1q2 != 0)
sd2 += (d1q1 == d2q1 +
d2q2) and (d1q2 == 0) and (d2q1 != 0) and (d2q2 != 0)
return prefs.strictlygreater(sd1, sd2)
def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
sd1 = 0
sd2 = 0
def check(t, dx, dy):
return dx.tf[t] > dy.tf[t] and len(dx) == (len(dy) + dx.tf[t] -
dy.tf[t])
for t in query.tf.keys():
if check(t, d1, d2):
sd1 += 1
elif check(t, d2, d1):
sd2 += 1
return prefs.strictlygreater(sd1, sd2)
def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
score = 0
for qt1, qt2 in itertools.combinations(query.tf.keys(), 2):
# qt1 is rarer
if not ctx.f.idf(qt1) >= ctx.f.idf(qt2):
qt1, qt2 = qt2, qt1
# term pair is valid
if not ((d1.tf[qt1] == d2.tf[qt2] and d1.tf[qt2] == d2.tf[qt1]) or
(query.tf[qt1] > query.tf[qt2])):
continue
# document with more occurrences of qt1 gets a point
diff = d1.tf[qt1] - d2.tf[qt1]
score += diff > 0 and 1 or diff < 0 and -1 or 0
return prefs.strictlygreater(score, 0)
def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
sim1 = 0
sim2 = 0
d1_terms = set(d1.tf)
d2_terms = set(d2.tf)
d_terms_both = d1_terms.intersection(d2_terms)
def sum_sim(ts):
return np.sum(self._similarity(ctx, ts, query.tf))
sim1 += sum_sim(d_terms_both)
sim2 += sum_sim(d_terms_both)
sim1 += sum_sim(d1_terms.difference(d2_terms))
sim2 += sum_sim(d2_terms.difference(d1_terms))
sim1 /= ctx.f.wordcount(d1)
sim2 /= ctx.f.wordcount(d2)
return prefs.strictlygreater(sim1, sim2)
def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
qterms = set(query.tf.keys())
sd1 = ctx.f.average_smallest_span(qterms, d1.termseq)
sd2 = ctx.f.average_smallest_span(qterms, d2.termseq)
return prefs.strictlygreater(sd2, sd1)
def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
ts = set(query.tf.keys())
s1 = ts.intersection(d1.tf.keys())
s2 = ts.intersection(d2.tf.keys())
return prefs.strictlygreater(len(s1), len(s2))
def preference(self, ctx: RerankingContext, query: Query, d1: Document,
d2: Document):
# prefer the shorter document
return prefs.strictlygreater(len(d2), len(d1))