Exploring Social Media Search ResultsRianne KapteinECIR Industry DayApril 5th 2012 Barcelona
BACKGROUND OXYMEFact Sheet Clients Est. in 2007 Consumer Goods FMCG Based in Amsterdam 24 client staff (9 nationalities) and 350 social media analysts around the globe Chemicals 13 languages covered by native speakers Transport & Logistics 125 projects completed at: Other sectors Dutch National Airport o 30 Multinational clients o In 12 countries Dutch National Railways
APPROACH | QUALITY DATA BY AUTOMATED & HUMAN ANALYSIS Inhouse built Human analysis Translate high specialized to remove not quality data into search software relevant results actionable insightsDefine a solid Collect opinions Smart analysis Classify only Analyze andsearch strategy with a webrobot relevant data report Total results WHERESearch strategy WHAT WHOSearch period # Not relevant MESSAGESearch terms SENTIMENT And thousands more! Etc. # Relevant
SOCIAL MEDIA SEARCH TYPES• Personal: What do my friends say about x?• General: What does the world say about x?• Commercial: • What are people saying about my brand, product or campaign? • What are people saying about my competitors? Relevant result: “Oh delicious! I just got Orangina at Albert Heijn. Orangina and I are friends for life”
SOCIAL MEDIA SEARCH CHARACTERISTICS• The date and time of search results is important. • Out of the temporal context, the meaning of the message might be lost • The query can include temporal restrictions, e.g. monitor a product launch• Ambiguity of search terms• Many search results are short such as tweets, or Facebook status updates. • Not a lot of noise in the text• Recall is important, you want to know how many people talk about you. • Large volumes of data• You want to know who is talking and how influential their messages are. • Identify promotors of your products • Identify complaints • Advertisements and spam
YET ANOTHER SEARCH ENGINE•Goal: First impression about the buzz around a brand•Exploit capabilities of human analysis: make it easy for the user to exploreand analyze the search results•Show basic statistics•Show differences over time•Show all results: raise awareness about what people will find when theysearch for your company
TWO DIMENSIONAL WORDCLOUDSTwo dimensions:o Size of terms represents relative term frequencyo Colour of terms represents novelty of termInstead of novelty other categorizations can be used as the second dimensionin the word cloud.Categorizations can include :• sentiment: positive vs. negative messages• brands: messages mentioning "Braun" vs. messages mentioning "Gillette"
WARNING SIGN 1Rubbish in, rubbish out!Be careful with your categorizations.If too many messages are not classified into the correct category, the quality ofthe wordcloud will degrade.Since in the interface its easy to go back to the original message, users will nottrust the system if they see too many misclassified messages
COMPARING DELONGHI AND SAECO BY OCCURRENCES IN TEXT
COMPARING TAGGED NEGATIVE POSTS SAECO AND DELONGHI
WARNING SIGN 2Be careful with small result sets:If at least one of the two categories only consists of a few messages (e.g. lessthan 25) the statistics on which the wordclouds are based are not so reliableThink about your minimum sample size!
OTHER POINTS OF ATTENTION• Words can have multiple meanings• Be clear on the fact it is an automatically generated summary• Heavily retweeted messages can have a large influence on the wordcloud
CONCLUSIONSWordclouds are a great search interface element for social media search• Navigational aid:Easy drilldown on search results• Summary of large number of search resultsShort, to the point messages