Maslow’s hierarchy of needs is an invaluable tool in the study of human motivations, with higher “self-actualization” needs rising above lower level “physiological” requirements such as food, water, sex, and sleep. These needs are expressed in our everyday lives, in our words, actions, and interactions, forms of expression relayed online, on-social, and enterprise-feedback sources such as customer surveys.
Whether you work in customer experience, market research, product management, or financial markets, you need a guide to making sense of the attitudes, emotions, and opinions—and associated transactions, behaviors, and networks—that make up so much of today’s Big Data. So why not fashion a hierarchy, in homage to Maslow’s, of functions involved in mining and exploiting human sentiment?
From Mentions to Meaning
The first step is to listen, but don’t stop there. The analysis journey will take you from mentions to meaning, from simple monitoring to engagement to optimization. Your technical goal is to tap the spectrum of sentiment sources and link sentiment to customer transactions, behaviors, and profiles. Your motivation? To (re-)personalize business and consumer decision-making. We’re aiming for machines that understand humans.
The analysis journey will take you from mentions to meaning, from simple monitoring to engagement to optimization.
So I’ll take a shot, now, at organizing the variety of approaches into a sentiment analysis/tools hierarchy, 7 levels, 24 elements: Digital measurement, optimization, and beyond.
Let’s start with basic capabilities and build up from there. Consider this a How-To guide that you can apply to evaluate your own organization’s needs, strategy, and progress.
Level 1: Listen
- Relevance: Monitor likely sources, but ask, Is a given status update, message, or post useful? Effective filters are your best tool in combating “information overload.” Typical approaches start with keyword search; more advanced ones add topic- and concept-based selection and filtering. Taxonomies are an asset here.
- Subjectivity: We’re interested in subjective customer (and patient, voter, and market) voices. Is sentiment expressed, any form of opinion, attitude, mood, or emotion? It’s often actually more important to know about the presence or absence of feeling than what that feeling actually is. Stronger tools will limit hits to updates and messages where the sentiment is expressed about the entity or topic of interest.
- Sentiment: Are views positive, negative, or neutral in tone? Weak tools look at messages; stronger tools associate sentiment with entities and topics; and the most advanced tools go beyond tonality to classify sentiment according to emotional and mood categories that are better aligned with your business needs.
Level 2: Measure
- Intensity: How intense is the expressed sentiment, based on word choice and pattern, based on capitalizations, emoticons, and other clues? Weak tools, operating at message level, record mixed positive and negative sentiment as cancelling out. Strong tools help you measure variation.
- Extent: What’s the aggregate view, for multiple opinion holders, across channels, over time?
- Trend: How do views change over time?
- Share: What’s the distribution and share-of-voice for each topic of interest and each view?
- Influence: Which voices are influential, on which topics, and how does influence relate to message diffusion?
Level 3: Analyze
- Root Cause: What are the sentiment root causes? Here, you definitely need to get beyond dashboards and summary stats to the underlying messages that are the sources of the numbers. Ability to extract and summarize opinions can help.
- Impact: What’s the business impact? To assess impact, you need to match sentiment to performance statistics or transactional records. This is not an easy task, and it’s at precisely this step that many, many commercially available tools break down.
- Action: What decisions and actions are indicated? Measurement will tell you What and root-cause analysis will suggest Why. It’s the unusual tool or model, in today’s market, that will tell you What to do about it.
Level 4: Engage
- Identity: Who are the opinion holders, individually? If all you can get is a social handle, use it. You’re aiming for conversations, not disjointed one-off messages.
- Interaction: How do interaction threads track across time, across channels? (You are managing your customer relationships, including on social channels, right?)
- Profile: Whom are you talking to? If you can pull profile information and understand the demographic categories a person fit in, so much the better: This understanding will, or should, inform your response. And who are they, the people you’re interacting with, talking to? Understand cross-platform social networks to understand message impact.
- Effectiveness: How do you tell if a response was effective? How do you systematize, response-effectiveness measurement?
Level 5: Predict
- Correlation: Let’s take measurement, engagement, and analysis to a new level: How does sentiment correlate with profile characteristics—demographic categories such as age, sex, and cultural background, and location—and with measureable behaviors?
- Genre: What’s the genre of a message? In practical terms: Is someone posting to complain (expecting a response), to inform others (no response expected), seeking information, or for some other purpose?
- Intent: What intent can you read from a message, or in more-technical terms, how do genre, object, and sentiment mesh-up? Do topic and choice of words suggest that someone is shopping for a new car, plans to switch mobile phone providers over frustrating customer service, or is simply making noise?
- Signal: What signal can you glean from a series of messages? Is Microsoft share price going to take a hit, based on social opinions? What volume of sales is likely for a new product, based on social chatter?
Level 6: Align and Optimize
- Conditions: A level-4 point addressed response-effectiveness measurement. Take it up a notch. Now look at response-effectiveness given the set of different individual and market signals and given business impact (a level-3 point).
- ROI: How should you measure online and social sentiment Return On Investment? What are your goals and how are you doing reaching them?
- Scenarios: You have signals, effectiveness, ROI: Use them to improve… everything: Your information collection (and the material you ignore), your measurement methods and analyses (including the information you factor in to customer engagement techniques and how you categorize it), your predictive models and their applications. You have a model: Use it. Evaluate different scenarios and approaches.
Level 7: Integrate and Extend
- Multi-source: The underlying assumption is that you’ll be working (first) with text, pulled from social and online media, surveys, e-mail, and the like. Yet there’s immense sentiment content in other forms of “unstructured data” including audio, images, and video. In a call-center, recorded speech may even be the primary sentiment source. Whatever source you start with, extend your analyses across the set of linked data sources. You’ll gain insights beyond what you can learn from any single source.
- Synthesis: Most analyses remain siloed; the biggest Big Data challenge I see isn’t Volume, Velocity, or Variety (the “3 Vs”), it’s integration and synthesis that aim to deliver insights possible only when you link across Big Data sources.
The Meaning of It All
What have I left out? A lot.
I’ve provided a road map with a lot of questions but no tool pointers. Let me just suggest that you look into tools with deep statistical capabilities, able to handle disparate information with the flexibility to integrate closely into key business processes.
And by the way, my 24 steps aren’t a linear sequence. Skip a few; undertake the ones that seem most important, doable, and highest return.
I haven’t written about analysis methods, although anyone who’s read anything I’ve written knows that I’m a huge fan of semantic-analysis methods, of text and content analytics and related techniques. These technologies uncover meaning in data, meaning that (per the last of my points) is found not only in ‘unstructured’ content but also in behaviors and profiles.
I wrote about that our technical goal is to tap the spectrum of sentiment sources and link sentiment to customer transactions, behaviors, and profiles. Our business goal goes beyond, to automate sense-making from the spectrum of sources available in the Big Data era. Sense-making finds the meaning of it all.
I haven’t addressed particular applications, because the approach I’ve describe should apply broadly, whether your task is market research, customer service, media analysis, counter-terrorism, or financial-markets trading. I have a lot to learn about applications myself, which is why I’ve organized the Sentiment Analysis Symposium. Join us to learn about customer sentiment in a multi-channel world; integrating survey and social data; and emotional vs. rational consumer decision-making.
Seth: Thanks for putting both perspective and order around customer analytics and the technologies that provide it. Sometimes I’ve felt like the data tools that are now widely available to the everyperson are akin to plopping us in the cockpit of the most advanced 747, and saying “here kid, go ahead and fly this thing.”
Happily, for now, the FAA doesn’t allow that, but that’s not the case with analytic tools. Lots of people make odd, granular pronouncements about consumer behavior, with all the authority and gravitas of a PhD researcher. Your article will be helpful for adding rigor to the process, so that people don’t go bonkers giving ‘insight’ from their analytics.
You pointed out that you’ve left out “lots,” so I’d like suggest on Level Three adding Consequence before assessing impact. Consequence can be one of the more difficult dimensions to ferret out, but it’s critical to pair with Impact. Related to that on Level Six, I recommend not attempting to use insights to improve everything, but rather to look at the range of project possibilities through a risk lens to identify which risks and opportunities portend the greatest impact for the enterprise, and then implementing those initiatives that will likely produce the greatest results.
Hello Seth,
I enjoyed reading this article and the framework you have offered. It can offer a way of thinking as applied to qualitative buyer research. In particular, I believe that predictive buyer modeling which analyzes buyer behaviors and decisions involves a good amount of sentiment analysis as well. I like the predictive aspects you mentioned – especially intent.
Whether the analysis takes place quantitatively or qualitatively – understanding meaning is the end goal. As you point out – integrating the many aspects of both worlds can lead to powerful meanings about customers that are – well, meaningful!
Thanks,
Tony Zambito