A Human-Centered Computing Framework to Enable Personalized News Video Recommendation

hurriedtinkleΤεχνίτη Νοημοσύνη και Ρομποτική

15 Νοε 2013 (πριν από 4 χρόνια και 1 μήνα)

63 εμφανίσεις

A Human
-
Centered Computing Framework

to Enable Personalized News Video

Recommendation

오준혁
(Oh Jun
-
hyuk
)

Questions


How to detect news topic from video?


How to measure inter
-
topic association?


How
to
measure interestingness of news topic?

Contents


Introduction


Related Work


User
-
Adaptive News Topic Recommendation


News Topic Detection


Topic Association Extraction


Interestingness Scores of News Topics


Hyperbolic Topic Network Visualization


Personalized Topic Network Generation


Personalized News Video Recommendation


Algorithm Evaluation


Conclusions

Contents


Introduction


Related Work


User
-
Adaptive News Topic Recommendation


News Topic Detection


Topic Association Extraction


Interestingness Scores of News Topics


Hyperbolic Topic Network Visualization


Personalized Topic Network Generation


Personalized News Video Recommendation


Algorithm Evaluation


Conclusions

Introduction


News video recommendation in CNN


Not related to the current news topic and user profile


Users need to follow(subscribe) news topics manually

Introduction


Video recommendation in YouTube


Related to the current video topic and user profile but not visualized

Introduction


Topic
Network


Visualize news videos and represent inter
-
topic association.


Hyperbolic
visualization


Enables interactive topic network
navigation(browsing).


Recommend the news topics of interest according to the
personal
preferences.

Contents


Introduction


Related Work


User
-
Adaptive News Topic Recommendation


News Topic Detection


Topic Association Extraction


Interestingness Scores of News Topics


Hyperbolic Topic Network Visualization


Personalized Topic Network Generation


Personalized News Video Recommendation


Algorithm Evaluation


Conclusions

Related Work

1. Automatic news topic detection


Identification of individual topics within a broadcast news video
by
detecting the boundaries

where the topic of discussion changes.


Special program structures and styles can be used to detect
boundaries.

Intro

News

Intro

News

Intro

News





KBS news broadcast

Related Work

2
. News visualization


Existing visualization systems disclose all the available news topics
to news seekers without considering interestingness of topics
.



It will be better to provide a
small

number of
interesting

news!

Contents


Introduction


Related Work


User
-
Adaptive News Topic Recommendation


News Topic Detection


Topic Association Extraction


Interestingness Scores of News Topics


Hyperbolic Topic Network Visualization


Personalized Topic Network Generation


Personalized News Video Recommendation


Algorithm Evaluation


Conclusions

User
-
Adaptive News Topic Recommendation


Goal : recommend the
news
topics of interest

by
incorporating topic network
and hyperbolic visualization.



To
-
do List


News Topic Detection


Topic Association Extraction


Interestingness Scores of
News Topics


Hyperbolic Topic Network
Visualization


Personalized Topic Network
Generation



Contents


Introduction


Related Work


User
-
Adaptive News Topic Recommendation


News Topic Detection


Topic Association Extraction


Interestingness Scores of News Topics


Hyperbolic Topic Network Visualization


Personalized Topic Network Generation


Personalized News Video Recommendation


Algorithm Evaluation


Conclusions

News Topic Detection


Define a set of over 4,000 elemental news topics.


Three major sources are integrated


Audio, Video, Closed Caption


News Topic Detection

News Topic Detection


Closed Captions


Natural Language Processing(NLP) is conducted.


Closed Captions are segmented into a set of keywords.


Special text sentences are removed by syntax parser


e
x) “CNN’s Andrew reports from Seoul”


not related to the topic


TreeTagger

is used to extract the POS(part
-
of
-
speech)
information


POS : a linguistic category of words (lexical category)


LingPipe

is used to



extract keywords.

News Topic Detection


Closed Captions

News Topic

K
-
POP, Korean Music, PSY

LingPipe

(Keyword Extraction)

Korean, music, PSY, Gangnam Style, release,

sequel

TreeTagger

(POS Tagging)

Korean[NN] music[NN] sensation[NN] PSY[FW] will[MD] release[VD] his[PP$] much[JJ] anticipated[JJ]
sequel[NN] to[TO] "Gangnam Style”[FW] Friday[NN].[SENT]

Original Caption

Korean music sensation PSY will release his much anticipated sequel to "Gangnam Style" Friday.

News Topic Detection

News Topic Detection


Audio


Automatic Speech Recognition(ASR)

system is used to
translate the audio channel to a text transcription.


Audio


Text


processed in a similar way to closed
caption.

Hidden Markov Model for
speech
recognition

News Topic Detection

News Topic Detection


Video


Detect
v
ideo objects(text area, human face)

because they
provide important clues about news story.


Confidence map

is used to measure the importance of
video objects in video.

News Topic Detection

Contents


Introduction


Related Work


User
-
Adaptive News Topic Recommendation


News Topic Detection


Topic Association Extraction


Interestingness Scores of News Topics


Hyperbolic Topic Network Visualization


Personalized Topic Network Generation


Personalized News Video Recommendation


Algorithm Evaluation


Conclusions

Topic Association Extraction


Inter
-
topic association(inter
-
topic contextual relationship)





d
(
Ci

,
Cj

)
: The
length of the shortest path between the news
topics
by
searching the relevant keywords for news topic interpretation fro
m
WordNet
.


ψ(
Ci
,
Cj
)
:
the co
-
occurrence
probability between the relevant news
topics obtained in news topic detection process.


The frequency
of co
-
occurrence of two news topic keywords in the same video.


ex) In a news video, “PSY”, “Music” co
-
occurs.


Topic Association Extraction


WordNet

:
a
lexical database for the English language
.


Provide a graph representing semantic relationship between words.

Topic Association Extraction


Topic network can be generated from topic association


News topics are organized according to the strength of
their association


Allow news seekers to easily recognize global overview of
large
-
scale news videos at the first glance.

Contents


Introduction


Related Work


User
-
Adaptive News Topic Recommendation


News Topic Detection


Topic Association Extraction


Interestingness Scores of News Topics


Hyperbolic Topic Network Visualization


Personalized Topic Network Generation


Personalized News Video Recommendation


Algorithm Evaluation


Conclusions

Interestingness Scores of News Topics


Interestingness Score



m(
Ci
)
: the
number of TV channels or news programs which have di
scussed
the
given news topic
Ci
.


Popularity


k
(
Ci
) : the
number of news topics linked with the given news topic
C
i

on the topic network.


Importance (
similar to PageRank)



Used to
highlight

the mo
st interesting news topi
cs and
eliminate

the les
s interesting news topic
s for reducing the visual
complexity for large
-
scal
e topic network visualiz
ation
.

Interestingness Scores of News Topics


PageRank Algorithm

Summary & Answers


How to detect news topic from video?


Define a set of news topic.


Integrate multi
-
modal sources


Closed caption : Natural language processing


Audio : Automatic speech recognition


Video : video object extraction and classification


How to measure inter
-
topic association?


Keyword association : length of path between keywords in
WordNet


Co
-
occurrence : the probability of co
-
occurrence obtained in news
topic detection process.


How to
measure interestingness of news topic?


Popularity
: the number of TV channels or news programs which have
discussed the given news topic


Importance
: the
number of news topics linked with the given news
topic
on
the topic network.

THANK YOU

Q&A