4. Understanding the meaning of results and how
to use monitoring tools so they work for you
Filtering the information
–Using a top down approach to filtering often wastes time, money
and effort
–Try a bottom up approach to define filters – it’s quicker and it will
help understand the nature of the social media and how your brand
interacts with it better – particularly useful when you need to ‘expand’
the search
–Use as many specifics as you can (that may be none, but usually isn’t)
–Be prepared to spend some time planning how you will use filters to
categorize and quantify data (every hour you spend at the planning
stage can save you days in execution – and can be especially valuable
when you are briefing vendors)
–Make sure the filters make sense to your business and monitoring
objectives
4
6. How much do you already know about
your audience?
•Take the time to establish which social media channels it
makes sense for your audience to use
•Better still, choose a tool which allows you to check your
assumptions
•Experiment and test your assumptions before you go setting up a
whole lot of details in specifications and filtering
•Use ‘exceptions’ to exclude unwanted audiences
•Make sure you understand the vocabulary your audiences use
6
16. Framing questions
•The most important rule is to make sure the answers are
answering YOUR questions
•Either make sure you test ‘logical’ questions ahead of time
or…
16
17. An end to keywords?
•Look at the new generation of analyzers which combine natural
language and context to generate results which ‘mean’ what you
‘said’ in your question
•The benefit is that missed ‘hits’ are minimised simply because
your vocabulary missed a word or variation, or the time tense is
different, or one or more of the key elements isn’t language (such
as an emoticon, or slang expressions)
•This approach allows you to work in a more natural way in both
framing questions and evaluating outcomes
17
18. The trouble with automated sentiment analysis
•Is often what it is expected to accomplish
•If you have designed in the filtering, contextual mapping and the
end user of the information it is perfectly practical to expect a
result in the 90-95% accuracy range
•Beware of claims over 95% (in a ‘live’ language, subject to fashion
and the re-use of terms in new ways there are very real and
tangible barriers
•But don’t make the mistake of assuming a human reader can do
any better – a 2009 study of 2,000 people saw them score an
average of 84% – and 88% was the top mark
18
19. The trouble with automated sentiment analysis
•There is a ‘semantic’ trap’ in most analysers in that they use
language as it is defined – not as it is actually used – and most
rely on words (social media posters sometimes don’t) – a lot of
language isn’t traditional linguistics!
•Black box solutions, however good they appear, are hiding things
from you – insist on transparency
19
21. Why automated analysis is NOT about
replacing people
•The role of automation is not about replacing people from the
process – rather it should be about allowing you more time to
think about what is important rather than speed reading
•Take the Apple iPhone 4 – social media DISCUSSION THREADS
not posts were ruining at 30/second during the launch day –
leaving anyone trying to read and make sense of the sentiment in
a passive, or reactive state.
•Context-based analysis can reduce the results to actions and
insights which are both manageable and insightful – without
increasing the errors from volume – without getting tired and it
will read every blog or post all the way through.
23. And finally…
•Feedback loops – both in the human and automated sense are a
valuable mechanism to ‘shrink wrap’ results ever closer to your
brand objectives – use them!
23
24. 1st
Generation 2nd
Generation 3rd
Generation 4th
Generation
Automated identification of
keywords
Manual sentiment mark-up
with article level values
Simple totals positive, neutral,
negative
Business Drivers
Deliverables
2010
Identification of keywords and
Boolean logic for product,
names and keywords
Automated sentiment for
keywords using dictionaries
• Identification of keywords and text
strings in context using Boolean
logic for products, names and
keywords
• Automated sentiment for
keywords/strings using dictionaries
• Article level values but based
on keyword / string values
• Natural language tracking for
phrases, names and
equivalents (antonyms and
synonyms) in contact entity
• 'Phrase level' analysis
including anaphora.
• Social media multiplies
media outlets into millions
• Business reaction time reduced
to real-time responses
• MarCom and press office
merges
• Fully interactive 'right time'
delivery of filtered and
categorised materials
• Innovation in removal of 'Black
Box' allows brands to obtain
'best fit' results
Tool providers include:
Glide Technologies
— We don’t use humans for analysis or dictionaries
— Ability to dynamically handle high volume up to 10,000 per hour
— Sentiment Analysis is computer generated using Natural Language
Processing (NLP),able to score slang, irony and even symbolic
language
— Accuracy levels are high, transactions to provide overall article
scoring with ‘transparent’ sentiment scoring.
— Use of anaphora to identify “its”, “their”, “the company” and
maintain reference
— Fully integrated with all other media types – no silos – full cause and
effect mapping