I dreamed about a device called the emotitonimeter. One inputs a bit of web or social media text and the emotitonimeter instantly determines whether the text is relevant to your brand and generates a smiling, frowning or neutral face indicating the emotional tone of the text by reaching back through the ether and tapping directly into the brain of the sender. Perhaps the deluxe version indicates whether and how we should respond.
Until someone invents the emotitonimeter, we’re tasked with sifting through mountains of available text documents (such as tweets, posts and comments) to uncover what’s relevant to our brands. Then, to guide our actions, it’s beneficial to decipher, among other vectors, the sentiment of the text.
Welcome to the science and art of sentiment analysis.
I was fortunate to be invited to second annual Sentiment Analysis Symposium organized by Alta Plana founder Seth Grimes. Over the next few posts, I’ll presenttakeaways most relevant to brand marketers and the agencies with whom they collaborate. I’ll touch on the state of the underlying science; the role of sentiment analysis in overall marketing strategies; factors to consider when selecting sentiment analysis tools; as well as perspectives from industry experts on practical applications.
Let’s get practical
Sentiment analysis is a component of larger text-analytics efforts. Text is turned into data — manually, using machine processes or some combination of the two — to discover, extract and analyze sentiment, opinions, passions, emotions or some other near-synonym of “sentiment” chosen by the analyzer.
Practical applications of sentiment analysis for brands include marketing, responding to customer voice and monitoring for events. Sentiment analysis is used by the financial industry to anticipate market changes, as well as by political strategists and the CIA. There is growing interest in using sentiment analysis to detect opinion spam, whether presenting hype or defamation.
A tutorial was conducted by Yongzheng “Tiger” Zhang, Catherine Baudin and Nitin Indurkhya of eBay Research Labs the afternoon before the symposium. The presenters expressed that properly conducted sentiment analysis provides more usable information than surveys and focus groups, where participants may want to complete the process quickly or express less-than-honest opinions due to the influence of others. eBay’s implementations include refining shopping research tools using consumer generated reviews and providing early warnings for events such as outages, when published complaints may precede triggering of system alerts.
Opinions — what are we looking for?
These points come primarily from the excellent eBay Research Labs tutorial.
Find all the facts and delete them — Well, not quite. Sentiment analysis may appear to start by separating objective expressions about entities, events and their attributes from subjective, targeted expressions of sentiments, attitudes, emotions, appraisals or feelings. However, not all subjective sentences contain opinions (“I want a phone with good voice quality”) and not all objective sentences are opinion-free (“The earphone broke in just two days”).
A quintuple of elements comprise most text documents with opinions — Not all elements are always needed for sentiment analysis, as some information may be implied by pronouns, context or language conventions and some will be indicated by document attributes. These factors can also obfuscate analysis. The five elements are:
- Object — Product, service, individual, event or topic.
- Attribute — Objects generally have two types of attributes that may or may not be part of an opinion portion of a text statement: Components (scoop of ice cream, sugar cone) and properties (weight, smell, temperature).
- Opinion holder — Person or organization that expresses the opinion. This could be the “speaker” or a person referenced in the text — it’s important to discern which is the case.
- Opinion orientation and strength — An opinion with a positive or negative polarity (as opposed to being neutral) may be placed on an intensity scale (contented -> happy -> joyous -> ecstatic)
Opinions exist on two vectors — These are not always easily discerned:
- A direct opinion (“The voice quality of this phone is great”) versus a comparative opinion (“The voice quality of this phone is better than my brother’s phone”).
- An opinion explicitly expressed (“This new search box is hard to use”) versus implicitly expressed (“Please bring back the old search engine”).
Sentiment can be identified at several levels — The three levels are the overall document (e.g., product review, blog, forum post), a sentence or a specific object attribute. For each level, search and analysis operates under somewhat different assumptions.
Some current and future challenges for sentiment analysis:
- Domain dependency — You may want to see an unpredictable movie, but not drive an unpredictable car.
- Handling negation with care — Sentences such as, “There is not one thing I hate about this product” should be judged positively.
- Irony and sarcasm.
- Social elements — Mining opinions only from similar people.
Awareness of the multifaceted nature of opinions gives us a foundation upon which to conduct our search and analysis. The proper means of conducting semantic analysis is not without controversy; I’ll take a look at those in the next post in this series from the presentations and conversations at the Sentiment Analysis Symposium.
Neil Glassman is principal marketing strategist at WhizBangPowWow, with a track record of success across linear, digital and social media. Join his conversation on Twitter or email Neil to talk about marketing or swap recipes.