Social media has given brand value a whole new meaning. Millions of customers a day discuss brands, products and even CEOs online. Valuable insights can be extracted from these conversations, but it is difficult to capture all the relevant information that’s locked in the over 340 million daily tweets:

• Twitter messages are very short, only 140 characters, which makes it difficult to capture the expressed emotion.
• Twitter messages are relatively unstructured compared to longer texts.
• A lot of abbreviations & slang are used, which are often classified wrong by generic models.

Enjoying free wifi courtesy Southwest airlines one of the great and under appreciated American brands and companies $luv

— howardlindzon (@howardlindzon) June 11, 2012

Facebook is just like jail, you sit around and waste time, you write on walls, and you get poked by people you don't know.

— TED (@ComedyTed) September 24, 2012

Is there anything more wretched in this world than a Goldman Sachs banker?

— John Lennon (@JohnXLennon) June 4, 2012

Our solution

In order to optimize our own stock price predictions, we have built algorithms that are specifically trained to deal with unstructured context, abbreviations & slang. We use a combination of semantic models, SVM's and NLP classifiers in a three step process:

Tweet preprocessing	Feature extraction	Machine learning model
Enriching unstructured content (hashtags, slang, abbreviations)	NGrams	Trained on 60000 labeled sample tweets
Stemming	Dictionaries	Support vector machines
Detecting negations	Emoticons, pre-defined hashtags & product names

Our Brand Sentiment API has an accuracy of 84.7% on a binary scale and an accuracy of 86.9% on a three point scale. This is an outperformance of roughly 10 percent over the best generic models that we know exist.

Our Brand Sentiment API is currently available in the following languages:

• English
• Dutch (coming soon)

It is possible for us to set up new languages if you have a need for it. Our consumer sentiment classifier is available through our API or can be licensed via an on-site install. Feel free to contact us with any questions you may have. We're more than happy to discuss which solution suits you best.