Languages & the Media Berlin, November 23rd 2012 Davor Orlic davor.orlic@ijs.si Knowledge for All Foundation Ltd

zoomzurichAI and Robotics

Oct 16, 2013 (3 years and 7 months ago)

76 views

Languages
&
the Media

Berlin
,
November 23rd

2012

Davor Orlic

davor.orlic@ijs.si

Knowledge

for

All

Foundation

Ltd


VideoLectures.NET


Content
,
Statistics
,
Licenses
,
Partners


Education
,
MOOCs
, OpenCourseWare
Consortium
,
Opencast

Matterhorn



The

idea

behind


Who
,
What
,
Why
,
When
?




transLectures


Pillar
s,
Current

status,
Results

and

Demo



WHAT IS IT?


VideoLectures.NET is the largest OER free and open access digital library
of academic talks. The lectures are given by distinguished scholars and
scientists at conferences, summer schools, workshops.




WHAT IS THE CONTENT?


Content built up via European research projects based in Computer
Science fields. Other content from OCW partners.



WHAT ARE THE STATS?


732 events, 10512 authors, 13726 lectures, 15965 videos


Visits: 9,626,639


Page views: 26,011,939


Signed in users: 23560


Licenses: CC
-
NC
-
ND




I enrolled in the MOOC “Intro to Databases”
winter 2011 at Coursera


108,000 accounts


475,000 assignment submissions


3,150,000 video views (heavy use of video)



Wouldn't it be awesome if all such content
and future options would be multilingual?


Language personalisation for millions of students


Video, audio, papers, coursework
-

all multilingual





WHAT WAS THE REASONING?


Huge set of
Hig
h
erEd

users (undergrads, MA, MSc, PhD)


Huge collection of videos


Videos are made of audio and video


Audio and video are data


Data can be harvested, changed and remixed



WHAT IF?


We capture the audio


Transform it into text



WHAT THEN?


We can have subtitles, transcriptions, translations, personalisation,
contextualisation, descriptions, time alignment, fragmentation,
recommendations, for
15965
academic

talks

Same for
:


Speech Processing
,
Text Analysis
,
Speech and Text
Resources


Most of

Europe's Languages are apparently unlikely
to survive in the digital age.

(META
-
NET
white

paper
)





LEARNERS PREFER VIDEO
?


YouTube

(78
hours

per

minute
upload
)


MOOCs

(3 mio
accounts
)




INITIATIVES AROUND VIDEO
?


Open content: OCW

(20.000
courses
)


Massive

le
cture

capture

system
: Opencast Matterhorn
project

(700
Universities
)


Massive

p
ortals

specialized in video lectures: VLN,
Polimedia

(25.000
academic

videos
)


SPECS?


Cost: 4,5
mio EUR


Project ref no. ICT
-
287755


Project acronym
:
transLectures


Project full title
:
Transcription and Translation of Video Lectures


Instrument
:
ICT
-
2011.4.2 Language Technologies


Thematic Priority
:
STREP


Start date / duration
:
01 November 2011 / 36 Months



WHO?


Universidad
Politecnica

De Valencia, Xerox, Knowledge 4 All Foundation
Ltd., RWTH, European Media Laboratory
Gmbh
, Deluxe Digital Studios
Ltd


OpenCast

Matterhorn
,
VideoLectures.Net
,
Polmedia







WHAT

IS

THE

AIM?



To

develop

innovative,

cost
-
effective

solutions

to

produce

accurate

transcriptions

and

translations

in

VideoLectures
,




To

deploy

those

tools

across

other

Matterhorn
-
related

repositories
.



For

translation,

we

consider

the

language

pairs
:

en

es
,

en

sl
,

en

fr

and

en

de
.




WHAT

IS

THE

IMPACT?



A

big

step

in

making

educational

repositories

truly

accessible

both

to

speakers

of

different

languages

and

to

people

with

disabilities
.




ADDITIONAL

VALUE?



Imagine

having

1
6
000

lectures

in

most

of

the

world`s

languages
.









KEYWORDS?


language technologies, machine translation, automatic speech
recognition, massive adaptation, intelligent interaction,
education, video lectures, multilingualism, accessibility



WHY TRANSCRIPTION & TRANSLATION?


There are accessibility issues that can be solved by transcription


Non
-
native
speakers understand
better by reading than by
hearing


At least 1,300 different languages with more than 100,000 native
speakers


No language with more than 20% of the world population





TRANSCRIPTION (EML)


the complete transcription of English lectures took 45000 hours
(2 months running parallel)


TRANSLATION (XRCE, UPV, RWTH)


different segmentation strategies for transcription and
translation being considered


INTELLIGENT INTERACTION WITH USERS


experimental protocol to evaluate intelligent interactive
approaches for users


INTEGRATION


f
irst

steps
on i
ntegration

software

into
VL
, Polimedia,
Matterhorn


EVALUATION


h
uman

evaluations for the second round of evaluation


Technology is good enough for transcription & translation


We are going to develop open tools for transcription and translation


Deploy the tools in the Opencast Matterhorn system


Think

of

a
business

plan
and

ideas

on
a spin
-
off


Provide optimisations for existing languages



Ideally extend the language set to Chinese, Hindi and other


Is intelligent interaction a realistic concept?


More focus on English into Slovenian translations to improve
them.


Work

on
building

a
community

of

students

for

evaluation



Languages
&
the Media

Berlin
,
November 23rd

2012

Davor Orlic

davor.orlic@ijs.si

Knowledge

for

All

Foundation

Ltd


Accuracy estimation for each transcription
and translation.


Adjustable computational behaviour.


Output constrained to user preferences and
corrections.


Fast learning from user corrections.



WHAT IS IT?


K4ALL is a Foundation based in London (2010) with the goal of
providing the legacy of the PASCAL2 Network of Excellence (machine
learning), part of this legacy is also the VideoLectures.NET website
and strong connections in Opencast Foundation (creating the
Matterhorn software) and Open Courseware Consortium.


WHAT DOES IT DO?


I4All: Provision and distribution of infrastructure that supports the
K4A mission


S4All: Online Science video journals and conference special issues


E4All: Organization and access to educational material


R4All: Research that facilitates the mission of K4All


A4All: Ensuring accessibility for as wide an audience as possible