There are oceans of information hidden in text. Most of the time, this information goes unused. As part of SAS' work on text analysis, one of our interns, Kenneth Enevoldsen, has been working on a model that analyses whether text sentiment is negative or positive. This could be particularly useful for customer satisfaction analysis but also has wider implications. I caught up with Kenneth before he leaves to start his PhD to find out more.
What opportunities do you see with sentiment models?
Positive and negative loading of text and speech has always been important. However, the increase in social media use, and the rise of emojis and blogging, means that the amount of text has increased. It has also become more normal to express explicit opinions in text. I see sentiment analysis as a way of managing this explosion of data.
I think it has three primary uses. The first is to prioritise actions. For example, customer service teams can prioritise emails from particularly frustrated customers, so they get a response before they have time to write a negative review. The second is trends over time. How has public opinion of our product or company evolved over time? What did our latest product do for our public image? Finally, it is useful for fine-grained analysis: to identify what customers particularly like about our product or company, and which areas need action.
How do you think sentiment models will develop over the next few years?
I think we will see a general move towards machine learning rather than rule-based models, but also two other main developments. First, I think we will see multidimensional sentiment analysis. Current methods focus primarily on positive and negative views. In the future, I think models will include more descriptive emotions, such as frustration or anger. This approach has been around for a while but has never really taken off.
The second area – which I think is much more interesting – is multimodal sentiment analysis. At the moment, sentiment analysis is usually based on text, but this is expanding to include images, voice and video. This will increase the usefulness of sentiment analysis and capture newer trends like the use of GIFs, Instagram and TikTok, which are text-poor.
How do you assess the possibilities of text analysis in general?
Text analysis is extremely useful. Google uses it for search engine optimisation, and Cambridge Analytica used it to influence the US elections. Even clothing brand Zalando has its own text analysis department. The benefit of using off-the-shelf text analysis tools like SAS Visual Text Analytics is that they are ready for use and contain the most up-to-date and useful technologies. I can see that it would be possible for a company that was new to text analysis to get up and running very quickly.
Are there any examples of this?
Theoretically, a telecoms company, say, could use topic modelling to look at product reviews divided into themes like internet speed, coverage and customer support. The company could then see whether views on each of these themes were positive or negative. It wouldn’t even have to worry about technical details, like automatic sentence analysis, spell checking and text cleaning. A more advanced setup could also include what words customers use to describe the coverage, or why customers are not happy with particular products. This could therefore have both immediate value for branding and long-term value for product development.
The sentiment model is a natural application of a customer satisfaction solution, but many other fields have interesting text analytics applications. See, for example:
- The free tool for medical COVID-19 literature search developed by SAS.
- A colleague in Germany also used it to look at the problems considered in a recent mass COVID-19 hackathon to get a sense of the issues covered and their relative importance.
- How text analytics is used to monitor and predict the risk of hospital-acquired infections.
- How text analysis has been used to look at player value in sport as a way to predict performance.