Method of topic summarization and content anatomy
A topic is summarized through content anatomy. The topic is summarized temporally. In an evolution graph generated through the present invention for the topic, all important events are picked up and cause-result relationships between the events are clearly shown. Hence, difficulty in understanding evolution of the topic is effectively reduced.
1. A method of topic summarization and content anatomy, comprising steps of:
(a) decomposing a plurality of documents related to a topic into a non-overlapping sequence of a plurality of blocks and describing a theme-identifying problem with said blocks through a constraint optimization method to find and express a plurality of themes to obtain a matrix comprising a plurality of eigenvectors;
(b) analyzing changes in said eigenvectors through an R-S endpoint detection algorithm to detect a plurality of temporally different events of each of said themes and obtain a plurality of summaries of said temporally different events; and
(c) calculating a context similarity of each of said temporally different events to obtain a temporal closeness per two of said temporally different events, and so forth, to form an evolution graph of said topic by associating all of said temporally different events according to said temporal closeness of each of said temporally different events;
wherein, in step (c), said context similarity of said event ⊖x is calculated to obtain a keyword vector through the following formula:
e
x
·
cv
=
1
1
+
e
x
·
eb
-
e
x
·
bb
∑
i
=
e
x
·
bb
e
x
·
eb
v
i
,
e
x
·
ev
⋆
b
_
i
,
;
wherein said temporal closeness of two of said events is calculated to obtain a temporal weigh through the following formula:
TW
(
e
i
,
e
j
)
=
{
1
e
j
·
bb
-
e
i
·
eb
n
,
if
e
j
·
bb
>
e
i
·
eb
,
1
2
⋆
(
min
(
e
i
·
eb
,
e
j
·
eb
)
-
e
j
·
bb
)
(
e
i
·
eb
-
e
i
·
bb
)
+
(
e
j
·
eb
-
e
j
·
bb
)
,
if
e
j
·
bb
≤
e
i
·
eb
.
2. The method according to claim 1 , wherein said block comprises at least one content selected from a group consisting of a sentence, a paragraph and a document.
3. The method according to claim 1 , wherein an eigenvalue of said eigenvector determines a relationship weight of said block to a respective theme.
4. The method according to claim 1 , wherein, in step (a), an inter-block association matrix between each two of said blocks is obtained through a matrix multiplication with said blocks of said topic.
5. The method according to claim 4 , wherein said inter-block association matrix is a symmetric matrix.
6. The method according to claim 1 , wherein, in step (b), temporal blocks in said eigenvectors are obtained as events in said theme through said R-S endpoint detection algorithm; and said block having a highest weigh is obtained as a summary of said event.
7. The method according to claim 1 , wherein, in step (c), said context similarity of two of said events is obtained from a cosine value of said keyword vectors of said events.
8. The method according to claim 1 , wherein, in step (c), a weight is obtained through a temporal weight function according to a relationship of two of said events; and two events having a values of temporal closeness larger than a threshold are regarded as related to further obtain an evolution graph of said topic.
9. The method according to claim 8 , wherein said temporal closeness is selected from a group consisting of non-overlapping, partial overlapping and complete overlapping.