|
Faculty Mentor: |
|
Alex Hauptmann (alex@cs.cmu.edu) |
|
Students: |
|
Zhirong Wang (Zhirong@cs) Jichuan Chang (cjc@cs) Ningning Hu (hnn@cs) |
Our research will
address the problem of story segmentation in the Informedia Digital Video
Library project, focusing on how to use closed-captioning cues to automatically
detect the story boundary and segment the word stream into coherent regions of
text.
With a collection
of text data annotated with the boundaries between regions, the problem of text
segmentation is to learn how to place breaks in unlabeled text by observing a
set of labeled examples. In our source data, the boundaries are denoted with
paragraph separators “>>>”.
Segmentation is
an integral and critical process in the Informedia digital video library. The
success of information retrieval in Informedia hinges on an important
assumption: that we can segment the whole new broadcast into individual
paragraphs or stories. For example, given a large unpartitioned collection of
expository text and a users query, return a collection of coherent segments
matching the query. Without segmentation technique, an IR application may be
able to locate positions in its database which are strong matched with the
user’s query, but be unable to determine how much of the surrounding data to
provide to the user.
One method to
segment news broadcast is using text (such as transcripts) as an aid. Usually
there is closed captioning along with the news broadcast, with hints in it
denoting story boundaries. How can we learn from this information to segment
text into stories? [1] points out that a seemingly simple problem can actually
prove quite difficult to automate, and that a tool for partitioning a stream of
text (or multimedia) into coherent regions would be of great benefit to a
number of existing applications.
Training data is
provided by new provider and stored in Informedia disks, with paragraph
separators (>>>) denoting the boundary of story. For example,
cwv199910091830.cc contains the closed captioning text for CNN WorldView on
Oct. 19, 1999. The closed captioning text contain (maybe incorrect) timing information,
and every news lasts about 18:30 minutes. Below is an example:
|
000076 ANDRIA: FROM CNN
IN ATLANTA, 000079 SEEN LIVE AROUND
THE WORLD, THIS 000081 IS
"WORLDVIEW." I'M ANDRIA 000082 HALL.
>>> THE SEARCH FOR 000084 SURVIVORS
CONTINUES IN SOUTHEASTERN 000087 MEXICO, WHERE THE
DEATH TOLL IS 000088 RISING BY THE
HOUR. IT'S NOW 000089 CLOSE TO 300.
RESCUERS ARE 000091 DIGGING THROUGH A
SEA OF MUD, 000093 SEARCHING FOR
HUNDREDS WHO ARE 000094 MISSING. THE U.S.
STATE DEPARTMENT, 000096 WARNING TRAVELERS
TO USE 000099 EXTREME CAUTION
IN THE REGION. |
We can remove all
the separators to produce testing data. Although these separators are not
always correct, the goal of our work is to build algorithm and system that can
put back all the separators in their original positions. Even ideally, the
system should improve on closed-caption boundaries.
People have tried
various approaches to do text segmentation. [1] introduces a new statistical
approach to automatically partitioning text into coherent segments. We will use
the probabilistically motivated error metric to measure our result. [2] [6]
uses the tiles for segmentation of paragraphs by topic. [3] has concentrated on
automatic segmentation of stories from news broadcasts using phrase templates.
BNN system is heavily tailored towards a specific news show format, namely CNN
Prime News. In this project, we will use machine learning and information
retrieval techniques to attack the problem.
The goal of our research is
to find the effective learning algorithm as well as features for text
segmentation. We will try to focus on topic changes detection and adjacent
sentence features extraction methods to do the segmentation job. In both of the
methods, we will also use difference algorithms: decision tree, neural network,
SVM, etc. We will compare their results together.
One simple way of testing
is compare the result of our segmentation with the original closed-captioning.
There maybe several kinds of errors: (1) incorrect location of detected boundary;
(2) boundary not detected, (3) extra boundary detected. We will use the error
metric introduced in [1] as well as traditional precision and recall metrics to
measure the output.
In this project, we will
try to answer the following questions:
(1) Which methods is the best in our implementation and
testing, explain why
(2) Whether the result of our segmentation is better than
the segmentation used in the Informedia Digital Video Library project, explain
the reason.
(3) Compare the performance of our method with the latest
text segmentation techniques, explain the difference.
We will regard our work in
this project as successful if we finally find some method which produces
satisfactory error-rate. If all the
methods we have tried are not good enough, reasons must be provided to justify
the result.
A project plan, outlining
in detail what type experiments will be performed and when (take your best
guess). Also describe other means of achieving the project goal, such as
consulting specific literature, talking to specific experts.
A description of the tasks
that will be performed by the individual team member. Our requirement: Each
team member has to engage in interesting machine learning work, so having one
student program the software interface while another runs machine learning
experiments is not a good division of labor.
[1] Beeferman, D., Berger, A., and Lafferty, J., "Text Segmentation Using Exponential Models," Proceedings of Empirical Methods in Natural Language Processing, AAAI 97, Providence, RI, 1997
[2] Hearst, M.A. and Plaunt, C.,
“Subtopicstructuring for full-length document access,” in ProcACM SIGIR-93
Int’l Conf. On Research and Development in Information Retrieval, pp. 59 – 68,
Pittsburgh, PA, 1993.
[3] Merlino, A., Morey, D., and Maybury,M., “Broadcast News Navigation using Story Segmentation,” ACM Multimedia 1997, November 1997
[4] Hauptmann, A., Witbrock, M., "Story Segmentation and Detection of Commercials in Broadcast News Video," ADL-98 Advances in Digital Libraries Conference, Santa Barbara, CA., April 22-24, 1998
[5] Lafferty, J., Berger, A., Beeferman, D., “Statistical Models for Text
Segmentation,” Special Issue on Natural Language Learning, C. Cardie and R.
Mooney, eds. 34(1-3), pp 177-210, 1999
[6] Hearst, M.A. (1994). Multi-paragraph segmentation of expository text. In Proceedings of the 32nd Annual Meeting of the Association for Computational Linguistics.