THE TOTAL POINTS THAT THE AUTHOR WHOTURES ARE APPLIED
13. Total Points: The total points that the author who
tures are applied.
gives the answer sentence receives.Step 2. Context assignment: every context sen-
tence is assigned to the most relevant question sen-
The previous literature (Shah et al., 2010) hinted
tence. We compute the semantic similarity(Simpson
that some cQA features, such as Sentence Length,
Has Link and Best Answer Star, may be more im-
and Crowe, 2005) between sentences or sub ques-
Figure 1: Four kinds of the contextual factors are considered for answer summarization in our general CRF basedmodels.tions as:
swer sentences x
i
, x
j
and their corresponding
replied questions Qr
i
, Qr
j
. If the similarity of Qr
i
sim(w
1
, w
2
)
sim(x, y) = 2 × ∑
and Qr
j
is above some upper threshold τ
uq
, this
| x | + | y | (2)
(w
1
,w
2
)∈M(x,y)
means that x
i
and x
j
are very similar and likely to
provide similar viewpoint to answer similar ques-
where M (x, y) denotes synset pairs matched in sen-
tions. In this case, we want to select either x
i
or
tences x and y; and the similarity between the two
x
j
as answer. This is done by setting the contextual
synsets w
1
and w
2
is computed to be inversely pro-
factor cf
2
such that x
i
and x
j
have opposite labels,
portional to the length of the path in Wordnet.
{
One answer sentence may related to more than
exp ν, y
i
∗ y
j
= − 1
cf
2
=
one sub questions to some extent. Thus, we de-
exp − ν, otherwise
fine the replied question Qr
i
as the sub question
with the maximal similarity to sentence x
i
: Qr
i
=
Assuming that sentence x
i
is selected as a sum-
argmax
Q
j
sim(x
i
, Q
j
). It is intuitive that different
mary sentence, and its next local neighborhood sen-
summary sentences aim at answering different sub
tence x
i+1
by the same author is dissimilar to it but
questions. Therefore, we design the following two
it is relevant to the original multi-sentence question,
contextual factors based on the similarity of replied
then it is reasonable to also pick x
i+1
as a summary
questions.
sentence because it may offer new viewpoints by
Dissimilar Replied Question Factor: Given two
the author. Meanwhile, other local and non-local
answer sentences x
i
, x
j
and their corresponding
sentences which are similar to it at above the up-
replied questions Qr
i
, Qr
j
. If the similarity
2
of Qr
i
per threshold will probably not be selected as sum-
and Qr
j
is below some threshold τ
lq
, it means that
mary sentences as they offer similar viewpoint as
x
i
and x
j
will present different viewpoints to answer
discussed above. Therefore, we propose the follow-
different sub questions. In this case, it is likely that
ing two kinds of contextual factors for selecting the
x
i
and x
j
are both summary sentences; we ensure
answer sentences in the CRF model.
this by setting the contextual factor cf
1
with a large
Local Novelty Factor: If the similarity of answer
value of exp ν , where ν is a positive real constant
sentence x
i
and x
i+1
given by the same author is
often assigned to value 1; otherwise we set cf
1
to
below a lower threshold τ
ls
, but their respective sim-
exp − ν for penalization.
ilarities to the sub questions both exceed an upper
threshold τ
us
, then we will boost the probability of
exp ν, y
i
= y
j
= 1
selecting both as summary sentences by setting:
cf
1
=
exp ν, y
i
= y
i+1
= 1
cf
3
=
Similar Replied Question Factor: Given two an-
2
We use the semantic similarity of Equation 2 for all ourRedundance Factor: If the similarity of answer
similarity measurement in this paper.sentence x
i
and x
j
is greater than the upper thresh-
where N denotes the total number of training sam-
old τ
us
, then they are likely to be redundant and
ples. we compute the log-likelihood gradient com-
hence should be given opposite labels. This is done
ponent of θ in the first term of Equation 4 as in
by setting:
usual CRFs. However, the second term of Equation
{ exp ν, y
i
∗ y
j
= − 1
θ
g
∥
2
be-
4 is non-differentiable when some special ∥ − →
comes exactly zero. To tackle this problem, an ad-
cf
4
=
ditional variable is added for each group (Schmidt ,
θ
g
∥
2