segmentation - Andrew T. Duchowski

advertisement
Evaluating Eye Movement Differences when
Processing Subtitles
Abstract
Feifan Zheng
COMPUTER SCIENCE
BARD COLLEGE
fz6131@bard.edu
Andrew T. Duchowski
COMPUTER SCIENCE
CLEMSON UNIVERSITY
duchowski@acm.org
Elisa Perego
DEPARTMENT OF PSYCHOLOGY
UNIVERSITY OF TRIESTE
Discussion
Subtitle types appear to have similar influences on the eye movements
during subtitled film viewing, failing to support our previous hypotheses.
Several reasons that might cause the contradiction are concluded:
Our experimental study enables us to obtain a richer picture of the
visual processing of subtitled films. By analyzing eye movement data,
we aim to determine whether different subtitle types influence gaze
switching during subtitled film viewing. Our hypothesis is that different
subtitle formats can cause different tradeoffs between image processing
and text processing. We specifically hypothesize that real-time subtitles
can have a disruptive effect on information processing and recognition
performance. Eye tracking data, however, shows that there is no
significant difference in gaze patterns among the four types of subtitles
tested.
• Quality of calibrations
The visualization of the eye movement recordings indicate there were
errors in the calibration process. These errors are non-negligible and
resulted in unreliable eye movement measurements (see Figure 5).
Introduction & Background
Previous research has shown that different types of text groupings in
respoken subtitles elicit different viewing behaviors (Rajendran, 2011).
This study examines segmented film (not live) subtitles.
The particular type of segmentation that was tested in this study is
between the noun phrase (NP) structure. We considered the following
NP splits in our study:
1. Noun + Adjective;
2. Noun + Prepositional Phrase;
3. Adjective + Noun;
4. Determiner + Noun.
Figure 2. Screenshots with heatmaps and scanpaths of 1 of the 28 modified
places from the testing video. Subtitle types: BENE (top left), MALE (top right),
PYRAMID (bottom left), and WFW (bottom right).
Results
• The results were analyzed along four metrics between subjects
(comparing data among four types of subtitles). The metrics were
number of gazepoint crossovers (see Figure 3), number of fixation
crossovers, percentage of gazepoint duration on subtitles,
percentage of fixation duration on subtitles.
Methodology
• The experiment was a between-subjects factorial design with 4
levels.
• Participants watched the same
10-minute video excerpt from a
Hungarian drama (see Figure 1).
The video was subtitled with one
of four types of subtitles, modified
in 28 places. They were
instructed to watch the clip for
content in both the subtitles and
video.
Figure 3. Example of a gazepoint/saccadic crossover. A crossover occurs when
the gazepoint/fixation moves from video section to subtitle section, or vise
p-values
from ANOVA of the four metrics were 0.233, 0.5503,
versa.
• The
0.7329 and 0.7613, respectively, showing no significant differences
in subtitle types among the four metrics (see Figure 4).
Figure 5. The overall scanpaths of eye movements from four different subjects. The red
rectangles indicate the subtitle sections. The first, third and last were badly calibrated.
• Duration of the testing video
The duration of the test video in the previous study was less than one
minute. In our study, the test video we used in this study lasted 10
minutes. A reasonable hypothesis is that the effect of subtitle difference
has been concealed within the long duration of the current video.
• Lack of participants
Since this study is designed between subjects, a large number of
participants is required to complete this study. The lack of statistical
significance maybe due to the small number of participants per group.
Despite the fact that the result contradicts our initial hypotheses, we find
that our data coincides with a previous study, which showed that
different types of text grouping in film subtitles do not elicit diverse
viewing behaviors and generate different tradeoffs between image
processing and text processing (Perego 2010). A key factor may be the
nature of live (respoken) subtitling.
Conclusion
Our study has shown that the type of text chunking has little
significance on the visual behavior of our participants. This result
contradicts our initial hypothesis, in favor of an earlier observation.
However, it should be noted that the duration of the video, the quality
of the calibrations, and the lack of participants produced considerable
error within our data. In absence of these sources of error, replication
of this study may produce stronger results.
• Participants took recognition tests
and filled out questionnaires a
Figure 1. A participant is watching the
week
after
they
watched
the
testing video in front of the eye
video.
tracking equipment.
• The four types of subtitles tested were:
1. BENE (well-segmented subtitles, NP structures are kept in the same
line);
2. MALE (ill-segmented subtitles, NP structures are broken into two
lines);
3. PYRAMID (well-segmented subtitles, the upper line was always
shorter than the lower line);
4. WFW (real-time subtitles, words showed up one by one).
(see Figure 2)
References
Rajendran, D. J., Duchowski, A. T., Orero, P., Martinez, J., and Romero-Fresco,
P., “Effects of Text Chunking on Subtitling: A Quantitative and Qualitative
Examination'', Perspectives: Studies in Translatology, Special Issue: When
Modalities Merge, Arumi, M., Matamala, A. and Orero, P., Eds., 2011 (to appear).
Perego, E., Del Missier, F., Porta M., and Mosconi M., The Cognitive
Effectiveness of Subtitle Processing. Media Psychology, 13:243–272, 2010.
Figure 4. Graphs showing means and standard errors for between-subject
gazepoint crossover count (top left), percentage of gazepoint duration on
subtitles (top right), fixation crossover count (bottom left), and percentage of
fixation duration on subtitles (bottom right) .
Acknowledgement
This research was supported, in part, by NSF Research Experience for
Undergraduates (REU) Site Grant CNS-0850695.
Download