Dan Somers of Warwick Analytics explains how machine learning is improving the lack of accuracy and relevance associated with sentiment analysis.
Sentiment Analysis is widely used to supplement the analysis of text data in surveys, complaints, reviews and other contact centre data. In theory, Sentiment Analysis categorises opinions expressed in a piece of text just as a human might.
However, as the volume and complexity of customer data grows, growing issues with Sentiment Analysis are providing flawed information and preventing companies from getting the full picture from their data. Key customer feedback that could drive positive business change is being missed.
For contact centres that have used sentiment analysis it is likely that you have experienced these inaccuracies – from skewed sentiment percentages and incorrect tagging, to completely missing key business insights. Furthermore, the accuracy of items sold as sentiment analysis that are billed as 90 percent accurate are sometimes as low as 40 percent accurate. This is a result of low recall and low precision, or multi-sentiment records being inaccurately classified.
Driving Business Change
So why is sentiment analysis such a poor business driver amongst contact centres? Usually it features basic output such as ‘neutral’, ‘bad’, or ‘good’ with a score that hardly contributes to driving business change. There is also the key issue that if there are multiple sentiments within a piece of text, Sentiment Analysis is unable to select both or identify the most important one, so key data is always being missed.
For example, a review along the lines of ‘the food was brilliant, and I loved the atmosphere, but the service was terribly slow!’ would have only one sentiment considered. Therefore, the actionable insight, ‘slow service’ is ignored, leading the business to miss out on change that could potentially have turned this satisfied customer into a huge promoter.
Sometimes feedback is missed as sentences with either a negative of positive word do not express any sentiment at all. For instance, the sentence ‘can you recommend a good tool?’ doesn’t actually express any sentiment, but it the word ‘good’ generally infers a positive sentiment. This leads to inaccuracy in both the categorisation and overall sentiment score, as well as the customer request potentially being missed.
There is also the fact that topics and sentiment are separable. Labelling a topic as “about billing” with the sentiment “unhappy” is not as useful or faithful as the output “customer expected pricing to be lower”. Companies are listing topics so that people only understand the ‘sentiment’ of the part of the customer journey, but it is losing all context with little insight into what the actual issue is.
The latest machine learning is now being applied by many contact centres instead of Sentiment Analysis with great effect. It is able to categorise and label multiple sentiments and label granular topics within feedback and reviews. In the case of the review ‘the food was brilliant and I loved the atmosphere, but the service was terribly slow!’, both sentiments will be picked up, in near real time, allowing both to drive business change. PrediCX from Warwick Analytics is one machine learning solution that is improving the lack of accuracy and relevance associated with sentiment analysis.
Whilst Sentiment Analysis can identify single sentiments and speed up the human categorisation and interpretation processes, we need to consider the opportunity cost of missing key data from our contact centres.
The overarching fact with Sentiment Analysis is that there is still considerable human effort required to discern from the data what is driving positive or negative sentiment. Not only is this time and resource heavy, but it often provides insufficient context, and lacks actionable insight. With machine learning you no longer have to read things, you simply act on the insight, a much better place to invest time and resources.
About the author
Dan is a serial entrepreneur and CEO of Warwick Analytics, A highly-respected speaker and widely published in the predictive analytics arena, Dan has also founded other successful IT businesses, including the managed services provider VC-Net. He holds an MA from Cambridge University in Natural Sciences and a Diploma in Business Studies.