![](bg4.png)
Table 4: Dataset for our method.
1 2 3 4 5 6
Ce Ve * Tr * 16 * Kl * 18 * Go*
Ve Ce * Tr * Go* Kl 16 18
Kl Tr * Ce * 18 * Bu * 16 Ve
Tr Kl * Ce * Ve * Go Bu 16
Go Gh* Ti Ce * Tr Ve * Pi
18 16 * Kl * Ce * Ve Tr Ti
16 18 * Ce * Kl Ve Tr Ti
Gh Gk* Bu Tr Ti Pi Po
Bu Kl * Tr Pi Go Gh Ti
Ti Go Pi Ce Bu Gh Tr
Pi Ti Go Bu Tr Gh Ve
Po Go Gh Tr Ve Pi Ti
Note: *: correct
4.3 Grouping
Five subjects evaluated whether the generated
groups were appropriate by five values (5: very
appropriate; 4: appropriate; 3: intermediate; 2: not
very appropriate; 1: inappropriate).
The average value was 4.3 (See Table 2), and the
result shows the grouping algorithm is satisfactory.
4.4 Character Connections
Five subjects evaluated the created character
connections by five values as grouping. Finally, we
conducted interviews about the above experiments.
The average evaluation value for the character
connection was 4.4. All subjects believed that Cell
was the most important character, although Goku is
the hero of the Dragon Ball. They felt that the
positions of Cell (center) and Goku (upper left) were
appropriate.
These results suggest that the created character
connections expressed the contents of the volume
well and the frequency of the characters indicates
the importance.
5 RELATED WORK
No previous research creates character connections
from manga or comics.
Some research create character connections from
other media. Goto et al. (2008) create character
charts from EPG texts that introduce movies. Both
studies use natural language understanding
techniques to identify relations between characters.
They do not deal with manga or comics.
Spysee, whose algorithm is based on Matsuo et
al. (2006) extracts person information from the Web
and displays social networks. Some famous
character names were input as person names. For
example, Cell is connected not only with such other
characters as Krillin but also a voice actor of Cell in
animation based on manga. This method is not
adequate to express the character connections of
designated units such as chapters or volumes.
Ogasawara et al. (2008) extracted persons from
broadcast videos to construct a human correlation
graph and examined both text and image processing;
they didn’t create graphs.
6 SUMMARY
We presented a method to create character
connections from manga using the frequencies of
characters and their co-occurrences by referring to
frames. Preliminary experiments using Dragon Ball
vol. 32 suggest the usefulness of our approach. Since
this is merely the first step of our research, we need
to improve our algorithms and conduct further
experiments using different manga. We believe our
approach is applicable to other types of comics and
should be investigated in the future.
REFERENCES
Reid, C., 2009. Graphic Novel Sales Up 5%; Manga Off
17% , In Publishers Weekly.
Goto, J., Yagi, Y., Aizawa, A., Sekine, S., 2008.
Generation of Correlation Charts from TV programs
based on Anaphora Resolution. In The 22nd Annual
Conference of the Japanese Society for Artificial
Intelligence. (CD-ROM) (in Japanese).
Matsuo, Y., Mori, J., Hamasaki, M., Ishida, K., Nishimura,
T., Takeda, H., Hashida, K., Ishiduka, M., 2006.
Polyphonet: An advanced social network extraction
system, In WWW2006, 397-406.
Ogasawara, T., Takahashi, T., Ide, I, Muase, H., 2005.
Construction of a Human Correlation Graph from
Broadcasted Video, In The 22nd Annual Conference of
the Japanese Society for Artificial Intelligence. (CD-
ROM) (in Japanese).
ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence
680