=Paper=
{{Paper
|id=Vol-1383/paper24
|storemode=property
|title=The Voice of the Customer for Digital Telcos
|pdfUrl=https://ceur-ws.org/Vol-1383/paper24.pdf
|volume=Vol-1383
|dblpUrl=https://dblp.org/rec/conf/semweb/BenjaminsCAVG14
}}
==The Voice of the Customer for Digital Telcos==
https://ceur-ws.org/Vol-1383/paper24.pdf
The
voice
of
the
customer
for
Digital
Telcos
V.
Richard
Benjamins
Antonio
Valderrabanos
David
Cadenas
Josu
Gomez
Pedro
Alonso
Bitext,
Spain
Telefonica,
Spain
Abstract
In
the
midst
of
the
digital
revolution,
the
telecommunications
industry
is
undergoing
major
changes.
One
of
the
changes
affecting
telcos
is
the
increase
in
data
sources
from
which
to
get
customer
feedback
–
big
data,
social
media.
Where
this
used
to
be
fully
controlled
by
companies
through
their
call
centres,
websites
and
shops,
today
much
feedback
is
expressed
in
social
media,
blogs,
news
sites,
app
stores
and
forums.
Telefonica
has
taken
up
this
challenge
and
opportunity,
and
is
now
systematically
listening
to
the
voice
of
its
customers
online.
The
problem
Only
a
few
years
ago
(around
2010),
telcos
received
a
hard
wake-‐up
call
when
Whatsapp
started
to
significantly
decrease
SMS
revenues.
Since
then,
the
choice
for
over-‐the-‐top
(OTT)
products
and
services
has
multiplied
by
orders
of
magnitude,
and
telcos
run
the
risk
to
loose
the
end-‐customer
contact,
and
to
be
forced
into
a
connectivity-‐only
offering.
The
answer
of
the
telecoms
industry
to
this
threat
has
many
aspects
(beyond
the
scope
of
this
paper),
including
the
launch
of
digital
services
such
as
financial
services,
security,
video,
etc;
leaner
working
methodologies
(lean
start-‐up);
and
much
more
customer-‐driven
development
and
in-‐live
management.
Listening
to,
understanding,
and
acting
on
customer
insights
are
key
for
launching,
growing
or
(rapid)
killing
of
customer
propositions.
In
this
work,
we
present
our
approach
to
systematically
and
automatically
listen
to
customers
on
the
Internet
as
soon
as
a
product
has
gone
live1.
Our
approach
is
built
around
three
main
concepts:
(i)
crawling
the
Internet,
(ii)
concept
identification
&
sentiment
analysis,
and
(iii)
visualisation,
and
is
set
up
in
such
as
way
that
any
person
with
“advanced
excel
skills”
is
able
to
self
serve
the
needed
dashboards
in
a
matter
of
hours.
Crawler
We
use
a
commercial
crawler
of
Sysomos
(www.sysomos.com),
which
crawls
social
media,
blogs,
news,
media
and
forums
on
a
continuous
basis.
The
input
is
any
Boolean
combination
of
keywords
(and,
or,
not).
Selecting
the
right
search
terms
is
important
to
avoid
inclusion
of
noise.
The
output
is
a
set
of
posts,
tweets,
and
articles
containing
the
specific
search
terms.
Where
possible,
1
Notice
that
customer
insights
are
also
key
for
the
conception
of
new
propositions,
but
that
is
outside
the
scope
of
this
paper.
1
geolocation
is
provided.
We
also
incorporate
reviews
from
app
stores
(e.g.
Google
Play
and
iTunes)
if
appropriate.
Concept
detection
and
sentiment
analysis
We
use
Bitext
(www.bitext.com),
which
provides
an
API
that
receives
as
input
the
set
of
retrieved
“items”
and
as
output
provides:
• The
concepts
mentioned
in
the
items
• A
set
of
possibly
multiple
opinions
that
constitute
the
items.
E.g.
a
tweet
or
news
item
may
contain
several
opinions
about
different
concepts
• The
neutrality
or
degree
of
tonality
of
the
opinions
(how
positive
or
negative)
• The
concepts
the
opinions
are
about
• The
phrases
used
to
express
the
tonality
of
opinions
(sentiment)
Bitext
applies
semantic
and
linguistic
technology
to
perform
those
tasks.
We
currently
use
it
for
Spanish,
English
and
Portuguese.
Sentiment
is
assigned
to
opinions
based
on
dictionaries
annotated
with
tonality,
which
we
have
tuned
for
“digital
products”
(e.g.
in
the
world
of
digital
products,
“cheap”
is
usually
something
positive,
whereas
in
general,
it
can
be
both
positive
and
negative.
Out
of
the
box,
Bitext’s
technology
is
about
70%
accurate,
and
after
tuning
to
the
digital
domain
this
is
increased
to
80%-‐90%.
Most
of
the
time,
irony
is
not
interpreted
correctly.
Figure
1
Concept
cloud
representing
how
sentiment
about
objects
is
expressed.
Size
represents
number
of
opinions;
colour
represents
tonality
as
in
next
figure.
Visualisation
For
visualisation,
we
use
Tableau
(www.tableau.com),
which
is
an
easy-‐to-‐use
(both
for
development
and
for
viewing)
tool
for
building
interactive
dashboards.
Interaction
allows
users
to
filter
for
viewing
only
negative
or
positive
comments,
or
for
different
languages,
to
drill
down
into
more
detail,
and
to
always
review
the
original
content.
For
each
product
or
service
we
monitor,
we
use
the
following
dashboards:
Mentions
(by
date,
source,
location,
language,
most
active
users
&
influence
factor,
most
shared
content),
locations
(geography
of
mentions
–
about
60%
of
the
mentions),
concept
clouds,
sentiment
(concepts
triggering
opinions,
phrases
expressing
sentiment,
degree
of
tonality,
date),
app
store
review
analysis.
Figure
1
shows
a
concept
cloud
of
how
tonality
is
expressed,
and
Figure
2
shows
the
breakdown
of
tonality
of
the
opinions.
2
Figure
2
Distribution
of
tonality
of
opinions
detected
in
the
mentions.
Conclusions
&
learnings
We
have
learned
that
–apart
from
the
typical
reputation
tracking
that
social
media
analytics
is
used
for-‐
it
is
a
valuable
tool
for
getting
quick
and
economic
customer
feedback
and
insights
for
products.
The
tool
is
able
to
detect
specific
issues
people
complain
about,
such
as
for
example
customer
care
quality,
price
and
pricing
issues,
registration
process,
and
specific
product
features
like
crashes,
unclear
interfaces,
battery
drain,
etc.
One
thing
business
users
see
as
very
positive
is
the
fact
that
the
tool
is
available
from
day1
of
launch,
which
enables
quick
responses
to
typical
overlooked
product
issues,
and
complements
the
internally
available
product
KPIs
such
as
downloads,
registrations,
active
users,
etc.
In
general,
we
see
that
in
the
early
days
after
commercial
launch,
comments
are
mostly
positive
reflecting
the
fact
that
most
are
announcements
and
promises
of
the
great
features
of
the
product.
Over
time,
more
and
more
feedback
comes
in
based
on
actual
usage
of
the
product.
Regarding
the
70%-‐90%
accuracy
of
the
sentiment
analysis
software
of
Bitext,
we
have
learned
that
when
there
are
thousands
of
mentions
per
month,
this
does
not
cause
a
major
problem.
The
dashboards
mostly
show
aggregated
information
and
the
main
trends,
concerns,
issues,
etc.
come
out
clearly.
However,
when
there
are
less
than
100
mentions
a
month,
especially
false
positives
(e.g.
something
seen
as
negative
which
in
fact
is
not
or
vice
versa),
harm
the
credibility
of
the
tool
towards
business
users.
A
final
learning
is
that
for
some
business
owners,
it
is
not
easy
to
deal
with
a
lot
of
negative
feedback.
And
the
fact
that
it
is
so
easy
to
get
feedback
and
that
it
is
based
on
publically
available
knowledge
makes
it
harder
to
“hide”
the
insights.
This
is
however
above
all
a
cultural
issue.
In
the
lean,
digital
world,
negative
feedback
should
be
embraced
and
taken
as
an
opportunity
to
quickly
improve
products
based
on
real
customer
insights.
In
this
brief
paper,
we
discussed
the
opportunity
for
telcos
(and
in
general,
large
enterprises)
to
take
advantage
of
social
media
analytics
as
a
valuable
and
economic
tool
for
obtaining
customer
insights.
This
is
however
just
one
step
in
the
journey
to
become
a
full
digital
telco,
which
eagerly
listens
to
any
relevant
external
and
internal
data
source
about
customer
and
markets,
including
internal
product
data,
call
centre
data,
open
data,
paid-‐for-‐data,
analyst
data,
screen-‐scraped
data,
and
APIs.
3