Originally published February 19, 2008
Last month, I looked at Sentiment Analysis: Opportunities and Challenges, promising a follow-on focus on applications. It's the breadth of opportunities – promising ways text analytics can be applied to extract and analyze attitudinal information from sources as varied as articles, blog postings, e-mail, call-center notes and survey responses – and the difficulty of the technical challenges that make existing and emerging applications so interesting.
We will explore three applications – influence networks, assessment of marketing response and customer experience management/enterprise feedback management – via mini-case studies.
Aafia Chaudhry, a physician who calls herself “a passionate enthusiast in the science of opinion leadership in healthcare systems,” is president of 81qd, a New York company that consults on pharmaceutical life-cycle management. Chaudhry applies text-analytics software from Linguamatics to perform targeted influence-mapping studies. She seeks to understand the correlation of physician sentiment, mined from sources that include event and interview transcripts, presentations, media releases and PubMed biomedical literature abstracts, with her clients' scientific and promotional messaging about therapies. She has purposely concentrated on sources where large volumes of readily mineable information are available; she is exploring adding blogs to the mix.
Jeff Catlin, CEO of text-analytics vendor Lexalytics, describes similar work at Cisco, which he characterizes as his company’s best success story. Cisco “used the sentiment engine to determine which executives have the highest correlation to positively moving the stock price when they deliver positive news. They found that certain executives had a positive influence on the markets, while others actually had a negative influence because of the tone of their delivery.”
Aafia Chaudhry's 81qd clients are “looking to develop relationships with key opinion leaders,” and text-mining along with peer-to-peer network analysis facilitate the task. She has been able to apply I2E Interactive Information Extraction software from Linguamatics to the text-mining task without modifications or extensions, although Phil Hastings, Linguamatics' business development director, notes that specialized thesauri could be brought in to handle sentiment extraction from sources whose language is less formal, containing colloquialisms and slang, than that used by physicians and biomedical researchers.
Study of influence networks, which Chaudhry calls the “science of opinion leadership, of innovation adoption,” is one approach to understanding stakeholder communities. More broadly targeted marketing initiatives, which are far more typical, similarly benefit from the application of text technologies to analyze sentiment.
Business spends huge sums shaping brand image and promoting brand awareness. To gauge the effectiveness of particular campaigns, corporate marketers will study transactions, for instance sales made in response to direct mail or using coupons, web-page visits and ad click-through, etc. But study of past transactions is of limited use in understanding potential buyers who are not responding to market messaging, in understanding competitive positioning and in picking up on nascent trends. Surveys and social-media mining, especially for attitudinal indicators, can fill the gap.
Unilever is one among many forward-looking producers of consumer packaged goods (CPG) that has applied text technologies to understand consumer sentiment. According to consultant Tom H.C. Anderson of Anderson Analytics, the analysis process applied in studying the Dove-brand pro.age campaign starts with surveys and web scraping from online consumer forums. Sites such as www.campaignforrealbeauty.com and www.doveproage.com have many thousands of messages with potential value. Anderson’s company codes and characterizes data, looking for sentiment polarity – positive, negative and neutral – seeking to understand emotions and attitudes. They apply a “triangulation” process with a 43-attribute “psychological content analysis” and with human coding of random sampling of records that validates results discovered through automated text analytics performed with SPSS Clementine and Text Analysis for Surveys.
There are many online consumer forums similar to the Dove sites. See, for example, the independent iCompact.com site. Theses forums as well as e-mail and blogs and other social media can be fruitfully mined not only for attitudinal data that indicates market sentiment, but also to understand influence networks, per Aafia Chaudhry’s work, and the diffusion of opinion.
The difficulty of sentiment-mining is shown by the example I offered in my earlier sentiment article. A comment posted to Dell's IdeaStorm.com forum, reproduced verbatim packs misspellings, “conversational” punctuation and syntax, and irregular capitalization that is both used for emphasis (“REALLY”) and misused (“ram”) in way that makes it hard to disambiguate the subject of the comment, random-access memory (RAM), from an animal:
Dell really... REALLY need to stop overcharging... and when i say overcharing... i mean atleast double what you would pay to pick up the ram yourself.
Yet while Catherine Cardoso, Associate Insights Manager at Unilever, says “Text analytics is a new methodology for us," she adds, "We were very pleased with the results and the depth of insight. The results were helpful beyond understanding reactions to our campaign. We also gained an understanding of what motivates people on discussion boards, which issues are most important to women in our target group, and how to create better products and messaging for them.” Cardoso sees great potential for text analytics. She says, “We've been thinking about other ways to utilize this technology which would allow us to not only continue to listen to and understand our consumers, but to create a more real-time two-way communication.”
Customer experience management (CEM) and enterprise feedback management (EFM) initiatives seek to take organizations beyond measurement to active stakeholder engagement. These disciplines are not new; they are long-time staples of customer relationship management (CRM), product design and quality programs that have been given a significant boost by the addition of text analytics, including sentiment extraction, to the analytical mix.
CEM and EFM seek to discern the “voice of the customer” in the many millions of enterprise-customer contact points that may include, per Sid Banerjee, CEO of text-analytics vendor Clarabridge, “survey responses triggered by sale of product or services (either traditional brick-and-mortar or e-business transactions), surveys triggered by calls for support, verbatims triggered by the support process (e-mails, chats, conversation notes stored in a CRM application),” and other forms of interaction. They seek to shift the focus from markets to customers. Banerjee’s colleague, Clarabridge President Justin Langseth, explains, “One of our design goals for 2008 is to make CEM a lot more than just monitoring and analysis, but to actually turn it into a decision-making, what-if-enabled feedback loop where you can do analysis, support a decision, take action and then measure the results of the action later on in terms of better loyalty, higher satisfaction, more sales, retention/loss of customers, etc... i.e., management.”
Vendors such as IBM have been working for several years to apply text analytics to enterprise decision-support needs. IBM unveiled their TAKMI (Text Analysis and Knowledge Mining) system, designed to analyze call-center logs, in 2001; their systems have evolved toward a vision of a “contact center of the future” that relies on real-time extraction and processing of emotions and attitude to smooth customer interactions.
Michelle DeHaaff, marketing VP at Attensity, describes her company’s approach in this way:
“We've been doing sentiment analysis on survey data. We've even had cases where the coded scores indicated a high satisfaction rating and the actual comments indicated a much lower level of satisfaction. We've parsed out negative/positive comments in service notes and have even used our sentiment extraction to find "cries for help" in e-mail, service notes, and web forums. We work with our customers to go beyond sentiment to cause. In the "cries for help" example, our customers get the detail around the sentiment, not just general sentiment. [Understanding] the cause of it leads us to finding a specific action our customer can take to impact sentiment.”
The trend, clearly, is in directions that would enable organizations to reach the goal put forward by Unilever’s Catherine Cardoso: analytics-assisted real-time, two-way communication. A spectrum of researchers and vendors take part in these efforts, and there has been significant progress. Accuracy remains an issue, however, for some potential users and will be subject matter for a future article.
Recent articles by Seth Grimes