When you think about some of the biggest company fails, what comes to mind? For me, it's Facebook’s Cambridge Analytica scandal.
Last year, news broke that the data consulting firm – hired by the Trump campaign in 2016 – was abusing the information of more than 50 million Facebook users (I had no idea that many people still used Facebook). The scandal led to a string of investigations into the social media platform’s data sharing practices and the ways these practices were enforced, shedding light on how the platform handles (or mishandles) privacy across the board.
Since the fallout, things at Facebook have seemed to settle down. Advertisers and users haven’t fully abandoned the platform, but the company’s percentage increase in revenue hit its lowest point in six years in Q3 of 2018.
This is evidence that customer satisfaction ebbs and flows, and it’s crucial to gauge sentiment from feedback in order to make necessary changes to influence customer retention, CSAT scores, and revenue.
In the world of CX and data science, sentiment analysis is a key piece of the puzzle in determining customer intent. In Stratifyd’s case, we take a piece of text to determine whether it’s positive, negative, or neutral, and leverage it to help companies determine how consumers feel about their products, services, and overall brands. Sounds simple enough, right?
The role of sentiment analysis is growing significantly due to the explosive amount of unstructured data from the rise of globalization and interconnectedness of systems. Customers like to give feedback in any way that’s convenient to them, whether it’s a phone call, chat session, or open-ended survey. Collecting this unstructured data is the most valuable way to uncover sentiment. With these insights, companies can better understand CSAT, campaign effectiveness, and competitive intelligence.
Sentiment can be extremely difficult to detect, especially given the nature of human language and sarcasm (arguably the best part about natural language). Idioms, or an expression that has a figurative meaning totally different from the literal meaning of its individual components (throwback to middle school English class), can also make it difficult to understand a customer's true sentiment.
It’s typically easier for machine learning models to detect linguistic aspects from data sources like surveys and chat transcripts, these typed sources tend to have more pointed feedback. When you attempt to digest call recordings, it’s an entirely different level of complication. So, we built our own speech analytics technology, complete with acoustic modeling to gauge energy levels, indicating excitement or anger based on tone, and instances of overtalk or silence.
Our customers are Fortune 1500 enterprises spanning multiple verticals, which means we have to be hyper-aware of vernaculars, or jargon, when analyzing sentiment. For example, the word “interest” is commonly used across the board, but it has different meanings and carries a different sentiment in each industry. In CPG land, you express interest in a new product, which carries a positive sentiment. While in the financial services industry, interest is a proper noun and carries a neutral, or potentially negative, sentiment.
There are two main approaches to analyzing customer sentiment: The lexicon approach and the neural sentiment model approach.
The first involves calculating the semantic orientation of words or phrases, evaluating its positive, negative, or neutral strength. This method makes it easy to upload custom sentiment dictionaries and is great if you want to target sentiment around a specific word or phrase. However, it’s extremely time consuming to train machine learning models around this approach, which is crucial for accurate results.
One of the biggest drawbacks of a lexicon approach to sentiment analysis is figuring out how to accurately score it. This method assigns numerical values to words and phrases on a scale of -5 to 5. This method tends to create a bell curve, which, by nature, will push a majority of sentiment into the neutral category. This is a huge problem because neutral feedback is rarely actionable and most customers aren’t actually leaving neutral feedback about a product, the scoring model is incorrectly pushing it there. This puts only the most polarized feedback into the correct scoring categories.
The second approach to sentiment analysis is the one we prefer. We’ve been looking at the neural sentiment model approach over the past several years, as it involves deep learning and neural networks, which essentially helps teach machines how to comprehend natural language. This method breaks down phrases character by character instead of word by word, increasing accuracy and determining the exact points shifts in sentiment took place. But this approach requires a lot of trained, prelabeled data in order to be accurate. Not to mention, it can be expensive depending on the models used and required hardware (looking at you GPUs). Using the wrong neural network can bring processing time to a crawl, making it impossible to deploy in production. Luckily, we’ve found ways around all of these issues and are extremely proud of the models we’ve created.
Sentiment plays an extremely important part in understanding the emotional state of customers, but at the end of the day, it really is just a piece of the puzzle.
Combining sentiment with customer satisfaction scores and net promoter scores allows your company to understand what most impacts your customers.
Sentiment analysis, combined with other customer analytics methods like topic modeling, taxonomies, and predictive analysis, can uncover unknown insights and tease out the most pressing issues to fix to make your customers happy.
To learn more about how your company can benefit from Stratifyd, schedule a demo today.
About the Author
Kevin O’Dell is the Chief Technology Officer at Stratifyd. When he’s not finding solutions to the hardest customer analytics problems enterprises face, you can find him wakeboarding or spending time with his family.
Most young adults associate the first of the month with three things: Pay day, bills, and rent.Learn More
Netflix launched in 1998 as a mail-order DVD service that would have more movies than were available...Learn More
Contact us to learn more
Stratifyd is a next-generation analytics platform powered by Augmented Intelligence™. Inside your data there are key signals to health of your business and the Augmented Intelligence engine helps you discover them:
Ready to see it for yourself? Let's schedule a short demo.