What are you thinking? Brands would like to know
Brands only sell when people like them. So how would a brand know, aside from those sales?
Opinion mining and sentiment analysis can provide that insight. Companies can monitor how well or poorly their brands are thought of, simply by monitoring social media, online vendors and their own internal data streams.
That sounds great, until you stop and think about how to do this. Data gushes like water out of a high-pressure fire hose. Just managing the data streams is a challenge. Even then, how many sources of data can you manage? And how do you draw inferences from the data you are tapping?
There are different approaches to handling the challenge. Sociality.io gleans data from social media, while Lexalytics and Diffbot rely more on natural language processing to draw data from many online sources.
Apps can provide the data streams and record the “likes” for your brand. But no app can draw inferences about what the data is really showing. That challenge falls on the marketer and not the machine, though some think the machine will catch up.
And remember, opinion mining and sentiment analysis, while used interchangeably, are not the same thing. With opinion mining, you are looking for insights, noted Paul Barba, chief scientist at Lexalytics. Sentiment analysis reveals the “polarity” of the opinions, like good or bad, he added.
Dig, dig, dig
Different domains yield different types of data. Amazon will provide lots of “star reviews.” Twitter may uncover complaints. Even the company call center has data that can provide insights. If people are calling to complain about a product, they are flagging a flaw that can be fixed. While the goal is “situational awareness,” there is “no strategy [that] says you have everything,” Barba said.
Volume is one key factor. Companies that sell B2B often have small customer bases, so a lot of opinion gets filtered through people. Human-driven intuition works here, Barba noted. “Where the machine wins is when the human can’t keep up.” Mass markets with high sales volumes are a good example, as they generate huge amounts of data that can only be analyzed by algorithms. The upside here is that sample sizes are large enough to yield statistically reliable conclusions, Barba explained.
Socialty.io focuses its apps on monitoring and engaging with social media. Sentiment analysis can be used to see if brands are registering positively or negatively with the public. That enters the realm of categorization, developing “word clouds” to see what characteristics are being associated with which products, explained Hasan Toprakkaya, Sociality.io’s founder.
Who is doing the talking matters. The “first circle” are those who actively follow a brand, and monitoring and engaging this cohort can provide a lot of useful information, Toprakkaya said.
“More data is not good data,” Toprakkaya cautioned. The volume of data is tremendous, so “be more specific. Be more focused,” he said. “If you are building any kind of product, you need one hundred loving customers…who are keen to share your product with their friends.” he said. “You should focus on those people, using opinion mining.”
Or you can dig everywhere. Diffbot will scour the web, but the analyst has to assess the value of the source. Say the product is a running shoe. Twitter can be a source, but it is like a firehose. “[It’s] not a real ID social network,” observed Mike Tung, CEO. “Natural language on Twitter is not fully, grammatical English.” There is probably a lot of noise. In contrast, a runner’s forum on Reddit may be a better source of data on that running shoe, as dedicated runners will be sharing their thoughts and opinions, he said.
The remarks on that web page can also be rearranged for clarity, broken down to show posts, names and times. This allows influencers to be more easily identified, which then creates an opportunity for user-generated content marketing, Tung explained.
Remarks can also be graded for salience, as either positive (good) or negative (bad). “It’s useful for getting feedback for product development.” Tung said. But in the end, it is about garnering useful information.
“Are you collecting enough signal?” Tung asked.
What do those blips mean?
It is one thing to pick up a signal on the Web. It is another to figure out what it means. That still requires human interpretation. “The data making it science from end-to-end does not exist,” Barba said. “If there are no sales histories for customers, you can’t teach a machine to make that [intuitive] jump.” Machine synthesis “is very much beyond the state we are in.” he said.
“Marketing as a profession has become more technical,” Tung said. The more data you can access, the quicker you can respond or the quicker you can model a situation, he said. More signal means more accurate modeling. “You can supercharge and augment the guy who is interpreting [the data]” he said.
Being a computer engineer, Toprakkaya is a little more optimistic about the technology. Data streams are being turned into structured data, which machines can read better than humans. “Some critical part will need humans to understand it.” he said. “Maybe in the future, it will change.”