How do people really feel about your brand? ConveyAPI delivers sentiment analysis
The ConveyAPI is based on REST web service, is simple to use and provides programmatic access to a text analytics engine, allowing you to better hone in on what people are saying about your brand. The API employs natural language processing (NPL), statistical modeling and machine learning techniques to return content-specific values across sentiment, emotion, intensity, and relevance.
Data could be sourced through Twitter API, for example, so a tweet that reads: “I loveee my new Ford Mustang!!” would return the following:
- Polarity: Positive – with a confidence level of .76
- Emotion: Joy – with a confidence level of .65
- Intensity: High – with a confidence level of .73
- Spam: Not spam – with a confidence level of .45
Because I was at LAX when the United Airlines systems went down, I thought I would take one of my own tweets to see how I scored. My not so subtle tweet: “United Airlines system wide outage has everyone stranded at LAX. Manually checking in. Lines out the door – literally.”
- Polarity: Negative – with a confidence level of .56
- Emotion: Anger – with a confidence level of .27
- Intensity: Low – with a confidence level of .14
- Spam: Not spam – with a confidence level of .46
So where did the ConveyAPI get these emotional attributions?
The work with emotion attribution is based on the work of Robert Plutchik, in the '80s, who created a wheel of emotions consisting of “8 basic emotions and 8 advanced emotions each composed of 2 basic ones”.
Plutchik laid out the following basic emotions (joy, trust, fear, and surprise) and their opposites (sadness, disgust, anger, and anticipation), where joy and sadness are the basic and opposite emotions.
Then created the following chart to describe how they went together to create advanced emotions.
| Human feelings (results of emotions) | Feelings | Opposite |
| Optimism | Anticipation + Joy | Disapproval |
| Love | Joy + Trust | Remorse |
| Submission | Trust + Fear | Contempt |
| Awe | Fear + Surprise | Aggression |
| Disappointment | Surprise + Sadness | Optimism |
| Remorse | Sadness + Disgust | Love |
| Contempt | Disgust + Anger | Submission |
| Aggression | Anger + Anticipation | Awe |
Fascinating stuff!
The intensity attribution, as in the example above (high), might be comprised of punctuation+adjective+adverb. The rules are customizable, but there is a good deal of math that has gone into the existing rule-set, according to Mark Walz, VP of Product Management at Converseon.
The ConveyAPI has the ability to drill down from a page, to a sentence, to an Entity or specific keyword.
Being able to input a large amount of text is fine if you have an article that you wanted to analyze, but for a more granular analysis you will need to do a sentence-by-sentence analysis -- with each sentence returning polarity, emotion, intensity, and whether or not it is spam.
This way, instead of hearing that a blogger was 'neutral' on the latest Ford Mustang, for example, you would be able to categorize the article and car into its constituent parts. It may be that the blogger loved the new body style, power, and basic options, but was down on the gas mileage and telematics.
Likewise you are able to do Named Entity recognition – for example, Romney versus Obama; and Keyword recognition - Mustang, for example; going through the document/article and report the sentiment for only the 'Mustang' and not any of the other cars in the Ford line up.
The natural language engine understands slang, and odd phrasings with repeating vowels which tend to happen a good deal on social media sites. Consider the tweet, “I loveeeee my new Ford Mustang!” The ConveyAPI includes rules which state if word+last letter repeated, then count first repeating letter only. So that 'loveeeee' is correctly interpreted as 'love'.
Though the interface is Spartan, being developer focused, the tool could easily be used to analyze election content, Olympic coverage, and more. To ensure that the architecture can handle the traffic the product is hosted on the Amazon Cloud.
The interactive relevance model allows you to train the system for concurrence. In this way you have the ability to classify or tag relevant documents. You tell it what to keep and what to toss; with each iteration the tool improves. Once the training set is completed, you can quickly zero in on a keyword or topic across a large volume of data.
Soon the ConveAPI will be available as an add-on for Radian6.
If you want to know what your clients are syaing about your brand you had better be listening actively.
Let me know what you think. Start a conversation.