Subjectivity and sentiment add richness to human communications, whether in conversations or posted online or to our social networks. When captured electronically, customer sentiment — expressions beyond facts, that convey mood, opinion, and emotion — carries immense business value. We’re talking the voice of the customer, and of the prospect, patient, voter, and opinion leader.
Listening — for brand mentions, complaints, and concerns — is the first element of any credible social engagement program. Businesses that listen can uncover sales opportunities, measure satisfaction, gauge reactions to marketing campaigns and message themes, uncover root causes behind events, and detect and respond to reputation and competitive threats. So we have monitoring and analytics solutions — the better of them apply text and sentiment analysis technology — targeting online and social media as well as enterprise feedback in surveys, e-mail, and contact center notes, aimed at discovering business value in complex, expressive, and sometimes-confusing human language.
My aim here is to explore modalities — how information technology helps us get at affect, at attitudinal information. (These are points that will be covered in depth at a social media analysis and engagement conference that I organize, the Sentiment Analysis Symposium, October 30 in San Francisco.) I’ll start with a key definition —
Sentiment analysis systematically rates human ‘affective’ states according to positive or negative polarity or a neutral or mixed value, or according to mood, emotion, or feelings (angry, happy, sad, proud, disappointed, etc.) and to use sentiment data for business purposes.
— and then explore two questions**:
What types of Sentiment Analysis are there?
In what directions is sentiment analysis evolving?
Six Types of Sentiment Analysis
The question “What type of person is she?” may have many, many answers. Each of us has many types, according to: Demographic categories (e.g., sex, age, race, income), Personality, Interests, Occupation, and so on. We are many types simultaneously. All questions of the form “What types of X are there?” have many answers, and X = Sentiment Analysis is no exception. Of course, we’re most interested in the most important types, and here they are as I see them:
Coarse-grained to fine-grained: Some analyses discern sentiment at a corpus or data-space level (e.g., for a set of reviews or survey responses); others score particular documents or messages; and others resolve sentiment at an entity (e.g., person, place, or company), topic, or concept level. (Coarse-grained analysis is fine for some business applications, but you need fine-grained for others.)
Individual versus aggregate: Analyses might look for individual cases (“mentions”) or for aggregates over populations or sources or trends over time. (If you’re managing customer support, you need to get at each mention, while if you’re studying market pulse, you’re looking for the big picture.)
Metric: Analyses may rate sentiment on an absolute scale or they may look for relative/comparative sentiment — “I don’t much care for sports, but I do prefer basketball to ice hockey” — and/or measure such as variation, intensity, and change. (I think those latter 3 are under-appreciated, under-utilized measures. Too many tools will rate a review with 8 positive points and 7 negative as mildly positive, where the high degree of variation should flag the review as interesting, even more if you have both strong positives and strong negatives, that is intense versus mild opinions. Sentiment change is always notable. It invites the question, What triggered the change?)
Focused or integrated: Do you need to go beyond simple figures to get at breakdowns, root causes, and predictions? Integrated analyses seek to link sentiment to psychological profile, behaviors, demographic characteristics, transactions, events, and/or other data.
How It’s Done: Functionally, there’s 1) analysis by trained humans, 2) crowd-sourced analysis by untrained humans, 3) automated analysis of information extracted from “unstructured” sources such as text, audio, images, and video, for instance for text via natural-language processing (NLP), 4) analysis of categorical poll or survey questions, e.g., “Rate your hotel stay on a scale of 1 to 5,” and the equivalent, star ratings, and 5) inference of sentiment from numerical statistics, for instance, commercial inventories, consumer spending, investment levels, etc. Automated NLP may apply linguistic, statistical, and/or machine-learning techniques.
Finally, ROI: There’s sentiment analysis that delivers business value, and there’s eye candy. We see lots of social-media analytics and BI dashboards that convey sentiment via pie charts, trends lines, color-coded word clouds, and other graphics. With most of them, I’ve concluded, you face a decision gap: They tell you what (sometimes accurately, sometimes not), but they don’t convey meaning or suggest how you can use the information visualized to improve your business decision making.
The type of sentiment analysis that will work for you is analysis that is aligned to business goals. Work back from your goals to understand the type and form of insights that will best help you make better business decisions. Decide what data sources you need to tap, the analytical techniques and analysis granularity that are appropriate, exactly what you’re going to measure, and how you’re going to link sentiment to the wealth of other types of data available to you. Then you’ll be positioned for return on investment, for the only type of sentiment analysis that really matters, sentiment analysis that delivers business value.
Let’s conclude with a look ahead. In what directions is sentiment analysis evolving? Four are particularly important:
- As an industry: Adoption is growing, in a spectrum of business domains and applications. The anti-automation backlash continues but should fade as sentiment analysis providers and users move toward semantically infused analysis, with feature-level, business-need-aligned sentiment resolution, and away from simplistic, keyword-based solutions.
- Considering sources: Detection and exploitation of emotion in speech and images (in facial and “body language”), and implied by video-captured behaviors, will increasingly come into play, including actively in meeting commerce and security needs. Availability of solutions for smaller-market languages will remain demand-driven.
- A focus on intent: We seek to understand not just how people feel, but what feelings, linked to data from the variety of relevant, associated sources, say about plans.
- Predictive modeling: Sentiment analysis becomes fuel for efforts to shape opinion, attitude, and emotion.
The end result is sentiment analysis as a contributor to sense-making, to intelligent automation that enables machines to understand and act on the spectrum of signals present in the human world.
** This article builds out my response to a question posted on Quora and responds to a second Quora question.
Seth Grimes is an analytics industry observer — an analyst, consultant, writer — who helps organizations find business value in enterprise data and online information. Seth consult via Alta Plana Corporation, works as an industry analyst, organizes the Sentiment Analysis Symposium, and tweets at @sethgrimes.