masterclass san francisco: data-driven analysis of social conversations using natural language...

22
NLP + Brandwatch Analytics Deriving insights from social conversations using Natural Language Processing and the Brandwatch Analytics API

Upload: brandwatch

Post on 14-Apr-2017

691 views

Category:

Social Media


1 download

TRANSCRIPT

Page 1: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

NLP  +  Brandwatch AnalyticsDeriving  insights  from  social  conversations using Natural  Language  Processing and  the  Brandwatch Analyt ics  API

Page 2: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

What  we  will  cover  today

What  do  we  want  to  answer?  (and  why)

Our  approach  to  social  data

Leveraging  the  Brandwatch API to  extract  data  

Deriving  insight  from  personas  

Identifying  key topics  of  conversation

Segmenting  on  those  topics  to develop  personas

Replicating  back  into  Brandwatch

What  we’re  working  on  next

Page 3: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

3 |

§ Provides  an  in-­‐situ  portrait  based  on  exhibited  behavior  not  on  elicited  feedback

§ Highly  relevant  as  it  can  be  updated  in  near-­‐real  time

§ Enables  research  budget  to  be  focused  on  insights  rather  than  data  collection

Social  intelligence  enables  new  ways  of  answering  traditional  business  questions  and  driving  data  driven  actions

What  are  the  sort  of  questions  we  want  to  answer?  

How  can  a  financial  services  company  reach  out  to  cyclists?

How  can  we  get  small  business  owners  to  engage  with  their  cell  phone  

provider  online?

What  is  the  customer  journey   for  a  motorcycle  enthusiast?

Page 4: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

4 |

There  are  seven  stages  to  the  analytical  process  of  developing  utilizing  personas  with  social  data

Our  approach  to  working  with  social  data

ExtractDevelop  the  dataset

Linguistic  model Segment Analyze TrackQuery  data

Prepare

§ Need  to  truly  understand  your  data  before  any  analysis

§ Iterative  query/dataset  development  through  virtual  ethnography

§ Use  the  Brandwatch API  to  extract  the  full  text  mentions  

Model

§ Employ  Natural  Language  processing  to  model  how  people  talk

§ Use  either  qualitative  methods  or  clustering  algorithms  to  segment

Understand

§ Through  visualization  and  analysis  we  can  understand  thoughts,   feelings  and  preferences

§ Replicate  back  into  Brandwatchas  sub-­‐categories  to  monitor  on  an  ongoing  basis

Page 5: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

5 |

• Provide  the  basis  for  a  ‘corpus’  in  NLP  jargon  from  which  to  model

• We  have  built  a  library  of  functions  using    python   to  retrieve  and  format  the  data

• The  output   format  of  the  API  is  in  JSON  so  there  is  some  work  to  turn  it  into  a  table  we  can  read  and  use

Extracted  BW  data  has  many  use  cases,  today  we  will  be  primarily  focused  on  full  text  mentions

Leveraging  the  API  to  extract  the  dataset  

Example  API  function:def get_mentions_query_URL(  startdate,enddate,project_id,  query_id,access_token,fullText):

query_def =  "data/mentions”  end_date =  "endDate="  +  end_date +  "T00:00:00.000Z”start_date =  "startDate="  +  start_date +  "T00:00:00.000Z"request_URL="https://newapi.brandwatch.com/projects/"  +str(project_id)  +  "/"  +query_defif fullText ==  True:

request_URL =  request_URL +  "/fulltext"

request_URL =  request_URL +  "?"  +  "queryId="  +  str(query_id)  +  "&"  +  start_date +  "&"  +  end_date +  "&pageSize=5000"  +  "&access_token="  +  access_token

return request_URL

Read  more:  blog.tahzoo.com/tech-­‐thursday-­‐brandy-­‐py-­‐a-­‐python-­‐library-­‐for-­‐brandwatch/

Github:  https://github.com/BillmanH/brandy.py/

Page 6: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

6 |

Linguistic  model  -­‐ Identifying  the  topics  in  a  conversation

pumpkinsugar HEALTHY  LIVING

PUMPKIN  SPICECONVERSATIONS

TEXT  ANALYSIS

TOPIC  MODEL

1 Break  down  each  conversation  into  the  words  and  sentences  to  probabilistically  assess  each  word’s  relationship  with  each  other  word

2 Analyze  to  uncover  the  most  common  “topics”  of  conversation

3 Run  clustering  analysis  to  segment  on  topics4 Iterate  on  topics  until  we  develop  a  solid  segmentation

Four  steps  to  targeting  personas

Page 7: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

7 |

("pumpkin  spice  latte")  NOT("vue pack"  OR "value  pack"  OR  "how  to  make"  OR  "win  free"  OR "latte  cake" OR  "black  friday"  OR "pack  of"  OR "My  TL  right  now  iOS7  Hump  Day  iOS7"  ORsite:(twitter.comOR kdvr.com OR fox59.com  OR  news.google.com))

An  example:  who  discusses  Pumpkin  Spice  Lattes?

Our  query…Excluded  because  of  irrelevant    recipes  

Purposefully   broad  query   to  capture  full  

range  of  conversations

Exclude  Twitter  as  it  would  overwhelm  the  results  and  we  couldn’t  export  full  text  mentions

Page 8: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

8 |

Do  it  Yourself Starbucks Nutrition Healthy  

living Style Urban  living restaurants PS  recipes Amazing  treat

Pumpkin  Spice  

ingredientsPS  Flavor Coffee  at  

home

people pumpkin grams squash fall city binary milk love food pumpkin home

make spice fat healthy wear place victoire pumpkin time hari spice inch

things latte calories recipe fashion food restaurants coffee day babe pie coffee

life starbucks sugar food boots park options sugar good ingredients latte green

time fall registers recipes style street time recipe back sugar flavor set

thing psl data copycat color local pst spice week science flavored keurig

feel drink saturated favorite wearing free visit cup great cancer seasonal count

find coffee carbs soup dress art trading make home organic year price

years today sodium paleo black event restaurant cream made found taste mountain

world lattes pos version top restaurant september syrup work chemical food make

The  topics  

Page 9: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

9 |

DIY  Example:“I  get  annoyed  when  a recipe  calls  for  pumpkin  pie  spice.    It's  not  that  people  use  it  that  annoys  me,  it's  the  mere  existence  of  it  as  a  single  spice.    …  I  guess  I'm  just  a  purist  at  heart.    Since  I  haven't  seen  pumpkin  pie  spice  here  in  France  I  now  need  to  make  our  own  pumpkin  pie  spice  mixture,  and  then  figure  out  the  right  proportions   for  my  dreamboat  pumpkin  spice  latte.    Nothing  that  a  Google  search  won't  solve,  but  annoying  nonetheless.    And  don't  worry,  when  I  do  I'll  be  sure  to  share  it  with  you.    Maybe  you'll  even  get  some  rainbows.    Fingers  crossed.  

An  example  of  how  this  analysis  works

Treat/Reward  Routine  Example:“I  thought   splurging  on  a  venti pumpkin  spice  latte  would  make  me  feel  better  this  morning,  (or  maybe  even  the  three  cups  of  green  tea  with  lotsa honey  in  it!)  ...but  as  my  ears  pop,  my  nose  runs,  and  my  throat   feels  like  somebody  took  sandpaper  to  it  last  nite,  I  guess  it's  time  to  finally  suck  it  up  &  take  some  meds  ó¾Œ®ó¾�‚  I  blame  you!  Rodney  Deal!!  Haha kidding  kidding  ;  )

Below  are  two  pieces  of  verbatim  content  that  we  used  in  our  model.  The  first  post  is  connected  with  the  DIY  (62%  relevant)  topic  and  the  second  with  Treat/Reward  (73%  relevant)

62% 73%

DIY TREAT  /  REWARD PS  FLAVOR FALL  (SEASON) PSL  RECIPES HEALTHY  

LIVINGFILLER/  

INFREQUENT  WORDS

TOPICS:

Page 10: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

10 |

• K-­‐means  highlights  clusters  of  conversations  based  on  the  topics  they  discuss

• This  creates  a  segmentation  that  reflects  how  people  discuss  a  subject

• Keys  in  on  the  pattern  of  topics   in  a  conversation  

We  use  the  k  means  clustering  algorithm  to  segment  the  conversations  based  on  the  topics  in  order  to  create  the  personas

Segmenting  on  the  topics  

Page 11: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

11 |

UrbanLiving

Fall  (season) Dessert

Starbucks  drinks

Pumpkin  flavor

Treat  /    reward

Pumpkin  (recipes)

DIY

Fall  (season)

Treat  /  reward

Pumpkin(flavor)

Dessert

StyleDesserts

Treat /  Reward

Being  Healthy

Urban  Living

Fall  (Season)

Starbucks  Drinks

Fall  (season)

DIY Desserts

Pumpkinspice  recipes

Treat/  Reward

LESS  IMPORTANT

MORE  IMPORTANT

Grouping  the  topics  that  are  core  to  each  segment  we  can  see  where  differences  break  down

Mapping  topics  to  personas

Page 12: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

12 |

Plotting  continuums  to  understand  the  personas

Why  they  like  it

What  it  stands  for

NOVELTYNOSTALGIA

GUILTY  PLEASURE

DAILY  RITUAL

OPPORTUNISTICTRADITIONAL

PERENNIALSEASONAL

OFTEN

OCCASIONAL

EXPECTED

EARNED

Page 13: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

13 |

What  we  found

22%

PSL  PAMPERER A  pumpkin spice  latte  is  a  treat  to  be  savored  

after  it’s  earned  or  after  a  tough  a  Monday  morning,

“What  a  weekend.  Hello,  slow  Monday.  Oh  what's  that?  I  should  get  a  pumpkin  spice   latte?  Well,   if  

you  insist...”

34%

LATTECHEMIST

They make  their  own  lattes  in  the  comfort  of  their  own  home  or  tinker  with  the  official  version  

“Here   is  an  awesome  home  version  of  Starbucks  Pumpkin  Spice  Latte.  Very  simple  to  make  and  alotcheaper… personally  I  like  it  better  because  you  control  the  amounts  of  ingredients  you  put  in  it  

according  to  your  taste.”

38%

FALLFANATIC Pumpkin  spice is  part  of  what  makes  fall  special  

for  them,  a  pumpkin  spice  latte  is  one  part  of  their  fall  tradition

“Pumpkin  Spice  Latte  at  Panera.  Oh  yeah,  I  need  one  of  those!    Bring  on  fall!    Looking  forward  to  bonfires  in  my  fire  pit  and  my  newly  refinished  

fireplace.”

6%

PUMPKINTRADITIONALIST Loves  everything pumpkin  from  pumpkin  pie  to  

lattes,  fall  is  just  an  excuse  to  get  their  fix  of  pumpkin

“Are  you  ready  for  a  Pumpkin  Spice  Latte!?!?!    Or  how  about  a  Pumpkin  Bar????  Well   tomorrow  they  

both  will  be  available!!!!”

Page 14: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

14 |

Replicating  back  into  Brandwatch

PSL  PAMPERER

“morning  treat”  OR “Savedmy  morning”  OR ((rough  OR bad  OR terrible*  OR awful  OR  stressful))  NEAR/4 (morning  ORday  OR  week))

LATTE CHEMIST

((myOR I  OR Mine  OR  “made  a”)  NEAR/2f (organic  OR make  OR recipe  OR mixture))  OR homemade  OR “the  perfect”  OR  ((coffee)  NEAR/3  (dessert  OR “sweet  tooth”)

FALL FANATIC

(I OR MY)  NEAR/3  (“love  fall”  OR “finally  here”  OR “the  season”  OR  autumn)  OR  ((making  OR  made  OR  bake  OR  baked)  NEAR/4f (cake  OR pie  OR pastry))

PUMPKIN TRADITIONALIST

((pumpkin) AND (candle  OR products OR  cake  OR pie))  OR“pumpkin  flavor”  OR ((“I  need  a”  OR “must  have”  OR “must  get”)  NEAR/3f  (latte))

We  conduct  a  careful  qualitative  analysis  of  persona  mentions  to  translate  the  topic  model  into  Brandwatch rules

• Allows  us  to  visualize  and  track  in  Brandwatch• Create  each  persona  as  a  sub-­‐category• Creating  the  persona  rules are  iteratively  written

Hypo

thetical  ru

les

Page 15: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

15 |

Custom  geo-­‐mapping  for  DMA’s

Persona  use  cases

Typing  tools

Scoring  conversation  relevance

IDENTIFYING  TARGET  SEGMENTS

Commentator

DIY PS  Flavor Fall  (Season)

Treat  /  Reward

Focused  on  others

Traditional

Page 16: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

16

Next  level  – what  we’re  working  on  now

§ Ability  to  use  the  model  to  tag  incoming  mentions  in  Brandwatch

§ Determining  demographic  characteristics  from  language

§ Utilizing  topics  to  predict  outcomes

Page 17: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

17 |

Introduction  to  Tahzoo

Page 18: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

18 |

What  we  do:  data  driven  customer  experiences

Business  OrientedFrom  business  process  to  change  management,  

Tahzoo  helps  enterprises  become  organizationally  ready  for  transformation

Customer  Data  CenteredUnderstanding  customers  from  their  own  point  of  view  is  the  foundation   for  successful  transformation,  Tahzoo  compiles,  curates  and  analyzes  all  the  data,  outside/in  and  inside  out

Digital  Content  InspiredCustomers  have  transformed  the  way  they  

engage  with  companies  using  every  channel,  device  and  platform  available,  Tahzoo  addresses  

engagement  through  content  marketing

Technology  EnabledArchitecting,  implementing  and  integrating  the  right  technical  solutions  determines  the  transformative  nature  of  the  experience,  this  is  Tahzoo’s  core  expertise  and  experience

Frictionlessly  DeliveredBoth  the  customer  and  the  business  win  when  the  experience  is  delivered  as  efficiently  and  effectively  as  possible,  whether  that’s  delivered  in-­‐house,  as  a  managed  service  or  in  the  cloud  as  an  integrated  

technology  and  marketing  operations  services  model

Page 19: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

19 |

• Founded  in  2010• Privately  Held• 300+  employees• Primary  Verticals

• Financial  Services• Retail• Automotive/Manufacturing• Business  Services  

A  fast  growing  CX  agency  focusing  on  digital  transformation  across  a  number  of  verticals

Company  overview

• Washington  DC  -­‐ HQ• Seattle,  WA• San  Francisco,  CA• Richmond,   VA

• Milton  Keynes,  UK• Delft,  NL• Borlange,  SE• Novi  Sad,  RS

Locations

Page 20: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

20 |

Past  &  current  clients

Page 21: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

21 |

Technology  partners

Page 22: Masterclass San Francisco: Data-driven analysis of social conversations using Natural Language Processing & the Brandwatch API

Thank  you

Bil l  Harding  – Data  Scientistbil [email protected]

Colin  Rogers  – Direction  of  Content  [email protected]