Can We Breathe Easier Now That Sentiment Analysis is “Nailed?”

This is the first in a three part series examining sentiment analysis. Sentiment analysis – the use of computational linguistics to determine and categorize attitude – allows us to sort through mounds of customer responses and understand trends in consumer attitude on the large scale. Quick, accurate methods of analysis are vital for creating responsive improvements to products and better customer relationships.

According to an article in GigaOM, Andrew Ng, the founder of Coursera, has been working alongside Stanford PhD student Richard Socher to develop a program that can accurately assign sentiment 85 percent of the time. Socher’s team used an evolution of recursive neural networks, or “Deep Learning”, to achieve this breakthrough. Previously, the state-of-the-art computer model for this task peaked around 80 percent. The Stanford team’s 85 percent accuracy rate is a solid step toward establishing consistency between machine learning outcomes and human consistency in judging sentiment. Socher and his team believe that they can close the gap even further, and may reach 90 or 95 percent accuracy. The team should be recognized for their accomplishments in developing machine learning algorithms that are modeled after the same processes the human brain uses, and for extending their research into the analysis of sentiment.

We all use our own advanced form of sentiment analysis every day in our interactions with others, whether they are at home, at work or in social settings. We read and interpret facial expressions and body language. We listen to the tone of other’s spoken words. We search for context in writing to provide additional meaning. It is also clear that our perception of sentiment can be subjective. Our own experience has shown that there may be others that interpret things differently. In everyday life, we seem to be okay with this lack of consistency, and often embrace it as a form of diversity. This seemingly natural inconsistency in perceptions of sentiment is just one more trait that makes each of us unique.

Interpreting sentiment from written text is much more difficult than evaluating sentiment from a social conversation. Much of this difficulty is simply due to the absence of influence from visual cues and tone of voice. However, written language is also very rich, and full of apparent contradictions like satire or sarcasm that act to disguise true underlying sentiment. Add to this the complexity of our modern languages and the brevity of some of the mediums we use to express ourselves. Just as we have difficulty agreeing on sentiment when all of our senses are operating, humans also seem to suffer the same problem when it comes to evaluating sentiment of the written word. Mike Marshall of text analytics firm Lexalytics states, “Experience has also shown us that human analysts tend to agree about 80% of the time.”

The secret sauce for sentiment researchers is maintaining consistency in judgment while also dealing with the subtle nuances of language. Most approaches to sentiment analysis use a combination of classification approaches, annotated word lists and machine learning to attach sentiment “scores” to social posts. The hope is that sentiment analysis from an individual can be extended to gauge the overall sentiment of a “crowd” of online users. Over the last few years, the applications and use of automated sentiment analysis has grown significantly. Businesses use sentiment analysis to monitor the tone of product reviews and brand mentions across various social media platforms. The Obama campaign, for example, used sentiment analysis to test various campaign platforms on issues and Topsy published daily sentiment metrics for the two presidential candidates in the 2012 election. Sentiment analysis is even being used to try to predict individual stock and larger market fluctuations based upon the tone of news and social media feeds.

If all we are looking for is the ability to determine positive or negative tonality, then we may be satisfied by these advances. We, at 113 Industries, believe that it is time to move beyond the emotional measures and into the underlying meaning, reason, and rationale behind sentiment. After all, knowing that a crowd is angry does not provide you with any actionable information on how to turn their anger into joy. Likewise, knowing that someone is happy does not ensure that you can take actions in the future to return that person to that same euphoric state. As businesses strive to understand their customers better and discover new ways to relate to them and influence them, sentiment analysis will continue to play the role of the canary in the coal mine. However, the real advancements are going to come from those that apply “Deep Learning” techniques to truly transform the customer experience.

Thanks for reading. Until next time…

– Anupam Singh, President and Co-Founder of 113 Industries