1 CONDITIONAL RANDOM FIELDSSUB-QUESTIONS ALONG WITH THEIR CONTEXTS,...
3.1 Conditional Random Fields
sub-questions along with their contexts, then se-
We utilize the probabilistic graphical model to solve
quentially retrieve the sub questions one by one,
the answer summarization task, Figure 1 gives some
and return similar questions and their best answers
illustrations, in which the sites correspond to the
(Wang et al., 2010). This strategy works well in gen-
sentences and the edges are utilized to model the
eral, however, as the automatic question segmenta-
interactions between sentences. Specifically, let x
tion is imperfect and the matched similar questions
be the sentence sequence to all answers within a
are likely to be generated in different contextual sit-
question thread, and y be the corresponding label se-
uations, this strategy often could not combine multi-
quence. Every component y
i
of y has a binary value,
ple independent best answers of sub questions seam-
with +1 for the summary sentence and -1 otherwise.
lessly and may introduce redundancy in final answer.
Then under CRF (Lafferty et al., 2001), the condi-
On general problem of cQA answer summariza-
tional probability of y given x obeys the following
tion, Liu et al.(2008) manually classified both ques-
distribution:
tions and answers into different taxonomies and ap-
plied clustering algorithms for answer summariza-
p(y | x) = 1
µ
l
g
l
(v, y |
v
, x)
Z(x) exp( ∑
tion.They utilized textual features for open and opin-
v∈V,l
(1)
ion type questions. Through exploiting metadata,
+ ∑
λ
k
f
k
(e, y |
e
, x)),
Tomasoni and Huang(2010) introduced four char-
e∈E,k
acteristics (constraints) of summarized answer and
combined them in an additional model as well as
where Z(x) is the normalization constant called
a multiplicative model. In order to leverage con-
partition function, g
l
denotes the cQA feature func-
text, Yang et al.(2011) employed a dual wing fac-
tion of site l , f
k
denotes the function of edge k ( mod-
tor graph to mutually enhance the performance of
eling the interactions between sentences), µ and λ
social document summarization with user generated
are respectively the weights of function of sites and
content like tweets. Wang et al. (2011) learned on-
edges, and y |
t