Social media is notoriously difficult to interpret and analyze. Social media is unstructured text, which comes with the unstructured text interpretation problems. Qualitative research has always been harder to analyze than quantitative, because you can’t add and subtract words, find means and mediums, or measure their intensity on a Likert scale. In more cogent, longer form entries, such as blogposts and peer reviews, you can at least rely on somewhat intact grammar and sentence structure. However, in short-form entries like microblogs and tumblogs, brevity often forces the writer to forego proper grammar, spelling and sentence structure in favor of truncated words, shorthand (“frm” vs.”from” and “rly” vs. “really”), internet speak (“LOL” and “FTW”) and fragmented sentences. This makes the job of understanding and interpreting unstructured social media messages really darn difficult.
Even in longer form writing, such as blogposts, monitoring tools have a hard time figuring out how the sentences relate to each other and connecting them. For example if I built a keyword query based on “love” and “Diet Pepsi” to tease out Pepsi’s brand advocates, I would miss all the instances where the two are broken up into two sentences, such as: “Just bought a can of Diet Pepsi. I love the refreshing bubbles!” The system wouldn’t be able to look at the two sentences and figure out that they are related. This is an issue of precision. With advanced monitoring tools like Attensity360, you can specify proximity between terms and set it to “exact”, which will help you quite a bit to zero in on only those articles that have “love” and “Diet Pepsi” right next to each other. However it still doesn’t solve the problem of associating several sentences to each other. Moreover, unless you determine all the possible keywords that are synonymous with “love”, you will miss some, and especially those that are colloquialisms. This is an issue of recall. Because language is fluid in its structure and popular vernacular, you simply can’t anticipate all situations, inevitably falling into one of the above traps from time to time. This is precisely why we built theAttensity Analyze product on the tenets of natural language processing.
Another fascinating property of language is the ability of certain words to just pop into a sentence and completely change its meaning. If I had a penny for every time I misunderstood something on Twitter, then I would be a millionaire! Am I a millionaire? I wish! (I’m working on it… Evenutally 
What we have here is a conditional voice, which changes the whole meaning. The introduction of “if/then” is the signal that a tool like Attensity Analyze needs to receive in order to adjust the meaning that it gleaned from the sentence. Ok great, but what exactly is the business application? Actionability! If I can read this sentence: “I think [product x] is too expensive, but if they dropped the price, I’d totally buy it” and understand that 1) the writer thinks it’s too expensive, and 2) there’s an actual intent to purchase if something was done about it — that makes it 10x more actionable than if I walked away with “The product is expensive.” Although it’s useful to tabulate how many people think your product is too expensive, it’s only first step. True usefulness comes from the “so what” conclusions; actionability is key here.
To help companies dive into the nuances of conversational speech, we have identified 7 voice tags. We use these tags to understand unstructured text and add actionability to insights we provide to clients. They are:
- Question: This is self-explanatory; when the system detects a question mark, as well as the appropriate sentence structure, it gets tagged as a question. The business implication for this is pretty clear. If you know that someone is talking about your product / brand or product category, you can queue that message appropriately — most likely to product support who can answer that question.
- Conditional: As described above, conditional voice allows you to differentiate between the writer’s current action / feeling and how it would be different if something else was different. This is great for figuring out where your biggest hindrances are, and where change can reap the most rewards. It’s truly the low-hanging fruit of customer experience improvement (from a product, as well as service perspective).
- Intent: This is also a bit self-explanatory. Unlike conditional voice which points to intent if something was different, the intent tag helps you understand what the person wants / doesn’t want to do right now, under the current set of circumstances. Needless to say it’s a goldmine for the sales organization, if they know how to use this right — it can serve up intelligence about customers and prospects, allowing you to identify them at their time of need. Perhaps they are looking for your product? Perhaps they are doing research, or perhaps they are even ready to buy today! At the same time, you can uncover service liabilities if you hear someone say “I plan to leave company X in favor of company Y.”
- Negation: This allows you to become more precise with your verbs, when their meaning is negated. For example, if you were looking for prospects who were going to a certain event, you certainly would want to put people who are not going to the event into a different bucket.
- Augment: This voice differentiates between degrees and severity in sentiment, as well as identifies emphasis. There’s a big difference between someone saying “I dislike product X” and “I loathe product X”, as well as “Product Y is easy to use” and “Product Y is extremely easy to use.” This is helpful to a brand from the standpoint of more intelligent queuing and triage. If someone is really really frustrated or really really angry, you probably want to make helping them a priority! Moreover, correct identification of intensity is useful for sentiment measurement. If you can say that of the 63% positive sentimented tweets, 10% are absolutely in love with your product, 30% are pretty happy with it and 23% sort of like it, that’s a lot more powerful. What’s even more powerful is understanding the key drivers of these feelings. But that’s fodder for another blogpost.
- Recurrence: This voice allows you to understand if the mentioned event happened before, or if it’s an ongoing issue. For example, consider the difference between “I can’t get through to your support” vs. “I still can’t get through to your support.” From a support perspective, it can help you triage staleness of problems, and from a product development perspective, you can identify which product deficiencies still linger, whether or not you tried to fix them previously.
- Indefinite: This voice is literally the virtual “suggestion box” for your product / brand. Consider these examples: “You totally should create an iPhone app for your product” or “I wish you had a free trial”. Statements made with this voice are extremely actionable, and can help you understand what your customers actually want, as well as further engage with them to fine-tune and dig in: “Great suggestion! What capabilities would your ideal iPhone app have?”
Deciphering social media is no easy task, but with advanced text analytics, you can do a lot more than you would with a naked eye. Reading each tweet and taking notes on suggestions can make sense when you are a small shop with a small online audience. However, when you are a large enterprise with a popular consumer product — “fuggetaboutit!”
To learn more about advanced text analytics and how they can help you in social media intelligence, please consider joining us for the Attensity Engage conference in November!
Photo credit: geekadman
