Amanda says in a forum: “I have an iPhone 4. It sucks.” Her post, along with many others, is gathered by “online sentiment” robots trying to make sense out of them and give the clients (the companies and brands) some insights, like “30% of the online talk about the iPhone 4 is negative.” The question is, does this make sense? What real information can companies get out of statements such as Amanda’s? In our case, it is possible that Amanda is dissatisfied with Apple’s iPhone, but it is also possible that she is waiting to get an iPhone 5. Meantime, she has to “endure” her iPhone 4, and it sucks.

That companies want to understand what their customers think is old news. Companies have been conducting surveys among their customers and among the general public for years now, trying to get an insight into what people think about the company and its products. However, these surveys were not proven to be a real predictor of customers’ actions. For example, a customer may complain in the survey about an airline and say that the service is horrible and he’s never going to fly with it again, yet remain loyal since it’s the only company that has a direct flight on the routes he normally travels.

Extracting Brand Sentiments from Online Conversations – Challenge #1

So, the next step was to eavesdrop on consumers when they are having conversations online, and are not necessarily trying to communicate with the brand. The web is full of conversations, many of them about specific brand – on Facebook, Twitter, blogs, forums, reviews, and more. When a brand is very small, this “eavesdropping” can be done manually – a person or a team inside a company can try to constantly monitor what people are saying about the brand by simply searching for the brand name on search engines and on social websites. When the brand is larger, however, the online conversation can consist of hundreds or thousands of mentions daily – making it impossible to track them manually.

This quest to harvest the sentiment towards the brand brought on the emergence of services that claim to be able to collect large amounts of “online conversation” data about a brand and algorithmically analyze it. These services then output a number representing the brand sentiment – for example, by checking the percentages of negative and positive mentions out of all mentions in total – and giving the client reports such as: “30% positive, 20% negative and 50% neutral.”

This analysis is a challenge, since unlike human beings, who can pick up sentiment easily even from not-so-obvious sentences, algorithms have to follow a predetermined set of rules and look for specific sentiment-oriented words that are not always present in the conversation. Check, for example, this sentence: “why would Apple have this application in the iTunes Store anyway??” Put in the right context, many humans will accurately detect this to be a negative comment about a certain application. To an algorithm, however, this can appear to be a neutral statement, owing to its lack of overtly negative words.

Extracting Brand Sentiments from Online Conversations – Challenge #2

In an MSI insights article called “measuring online brand sentiment,” Julia Hannah writes about the current hot issue of trying to extract meaning out of large quantities of word-of-mouth. The paper describes a large-scale academic study by David Schweidel and Wendy Moe examining word-of-mouth sentiment across different online venues. Schweidel and Moe do not focus on the problem of an algorithm’s limited ability to understand context, they are interested instead in inherent differences in sentiment that exist across venues, products, and brand attributes.

For example, conversation venues that are more ‘one to many’ in their communication style (such as Twitter and blogs) tend to have more positive posts than venues that are more ‘many to many’ communication style (such as forums). They suggest that this “forum negativity” happens due to the fact that “when people interact with each other and respond to someone else’s opinion, there’s a tendency to differentiate and try to outdo one another by being more negative.”

Moreover, if we look at the different attributes of a product, it is very common for fans to criticize some aspects of the product while still maintaining an overall positive sentiment and positive purchase intent. One example could be “I don’t like the interface of the new speech recognition in the Galaxy 4,” or “I hate it that you can’t record your calls on the Galaxy.” The writer of this comment, even though expressing a negative opinion about the Samsung Galaxy 4, is actually only criticizing one attribute of it while still liking the overall product, and potentially being loyal to the brand.

Conclusion

Listening to online conversation is definitely necessary for successful brands these days. However, jumping to conclusions based on these conversations is dangerous. Both the theories and the software supporting extraction of online conversation sentiment are, at the moment, far from being predictive. Human involvement in these processes is still highly effective and required, and the overall best measurement of customer sentiments remains sales figures.