Sentiment Analysis, Hard But Worth It!

Last week our CTO David Bean joined analyst and industry expert Esteban Kolsky (@ekolsky), the CTO from ScoutLabs, Jochen Frey and architect Franco Salvetti from Microsoft Bing (and formerly from Powerset) on a panel about sentiment analysis. One of my favorite topics to talk about with Dr. Bean – it was great to hear the whole panel discuss this topic both from an academic and practical point of view. (Here’s a picture of the guys on the panel getting into the discussion!)

So, in a nutshell – here are some of the learnings from that event:

1) Sentiment analysis is hard!   Yes, sentiment analysis is hard for at least a few reasons.  First, in some cases, even humans have a hard time understanding the sentiment of what someone else is saying.  Commonly used statements when humans interact: “What do you mean?” “What are you trying to say?”  And that’s when they ask – a lot of times people just don’t understand the sentiment and don’t ask for clarification.  Second, beyond the issues of ambiguity, for computers, being able to pull out the tone and meaning in a statement or set of statements is hard because people express things in different ways and finding the sentiment in a sentence is hard using certain statistical approaches. Many applications that try to understand sentiment use keywords or clusters of keywords to understand sentiment.  If “happy” is used in the sentence then it is positive sentiment.  Well – one can see why this can be brittle. First, users need to list every word that they can think of that is indicative of positive or negative sentiment and track for it – this takes time and effort to create and maintain the rules needed to do this. Second, what if “happy” is negated somewhere in the sentence? Proximity analysis helps with this “look for any negations in the sentence within two words from happy.”  The problem is more rules to write and the sentences where the negation is more than a few words away are misclassified.

To accommodate for these issues Attensity invented (and has patented) a completely unique method for getting to accurate sentiment – we call it Exhaustive Extraction. This technology enables us to parse sentences and accurately look for instances of sentiment expression and get it right more so than any other vendor that attempts to do this.  The image below depicts how we parse sentences to find both sentiment and meaning. The process is analogous to what we learned in grammar school English classes. In this example we parse the sentence: I really love my iPhone, but the reception here is very bad.

Next we extract and aggregate the facts from the content. This allows us to not only find the words that are indicative of sentiment, but to find the relationships between words so that we can accurately identify both words that modify the sentiment (even if they are not close to the sentiment) and what the sentiment is about. Below is an example of the facts that we automatically extracted from this example:

As you can see from this example not only do we pull out the facts – but we can also find the degree of sentiment not just “love” but “really love” and not just “bad” reception but “very bad.”

Another example that highlights our ability to both find sentiment and get it right is illustrated below. In this example the negation on the sentiment is not near the sentiment – but we are still able to get it right! The content in this one is: I have an iPhone, but I am not really feeling very happy about the iPhone.

2) Sentiment analysis is needed! While it’s a hard subject and in some cases impossible for both people and technology to tackle – the one thing that everyone agreed on is that understanding it is needed! It helps companies understand what buying customers think of their products, services, buying experience, customer service, even the competition.  It is a leading indicator of purchase intent and churn. It can help organizations identify cries for help and emerging issues. Wow!  Worth the millions of dollars of research that go into making this possible – hey?  We’ve recently blogged about a few examples of analysis that we’ve done – valuable insights for companies and based on understanding sentiment. You can find them here.

3) Sentiment analysis is most useful when connected to what the sentiment is about! The ever elusive “why.” When listening to customers or conducting market research the thing we want to know most is “why?”  I have already covered this somewhat above – but I wanted to emphasize this again.  The key thing about sentiment analysis is being able to go beyond the sentiment analysis to find out what people were happy or sad about.  So why did they like the iPad and say they were going to buy it? (below is an excerpt from analysis we did on the iPad launch.)

Why didn’t they like about the Nexus?

You get the picture…….We can show you how and why understanding your customer’s sentiment is valuable to you.  Let me know if you want to see how!  Signing off for now…..