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