Information Technology Reference
In-Depth Information
in which
|
H
|
denotes the length of the hypothesis H, that is, the number of
predicates.
Note that pairs of target concepts are provided by a domain experts so as to
guide the search process.
•
Structure
(
How good is the structure of the rhetorical roles?
): measures how
much of the rules' structure is exhibited in the current hypothesis.
Since we have previous pre-processed information for bi-grams of roles, the struc-
ture can be computed by following a Markov chain [23] as follows:
Structure
(
H
)=
Prob
(
r
1
)
∗
|H|
i
=2
Prob
(
r
i
| r
i−
1
)
where
r
i
represents the
i−th
role of the hypothesis H,
Prob
(
r
i
| r
i−
1
) denotes the
conditional probability that role
r
i−
1
immediately precedes
r
i
.
Prob
(
r
i
) denotes
the probability that no role precedes
r
i
, that is, it is at the beginning of the
structure (i.e.,
Prob
(
r
i
|< start >
)).
<START>
0.28
0.09
0.49
0.12
1.0
0.08
conclusion
0.41
0.53
0.03
goal
object
0.06
0.05
0.23
0.56
method
0.54
0.16
0.35
Fig. 9.3.
Markov Model for Roles Structure Learned from sampled technical doc-
uments
For example, part of a Markov chain of rhetorical roles learned by the model from
a specific technical domain can be seen in figure 9.3. Here it can be observed that
some structure tags are more frequent than others (i.e., the sequence of rhetorical
roles goal-method (0.54) is more likely than the sequence goal-conclusion (0.08)).
•
Cohesion
(
How likely is a predicate action to be associated with some specific
rhetorical role?
): measures the degree of “connection” between rhetorical infor-
mation (i.e., roles) and predicate actions. The issue here is how likely (according
to the rules) some predicate relation
P
in the current hypothesis is to be associ-
ated with role
r
. Formally,
cohesion
for hypothesis
H
is expressed as:
cohesion(H) =
r
i
,P
i
∈H
Prob
(
P
i
|r
i
)
|H|
where
Prob
(
P
i
| r
i
) states the conditional probability of the predicate
P
i
given
the rhetorical role
r
i
.
Search WWH ::
Custom Search