AI-supported sentiment analysis

Futuristic Graphics for decoration only.
Just what is Capt’n Butler trying to say here?

Well! Tell us how you really feel!

OK, so it is not a well-kept secret what his feelings are. At least on the surface anyway. It would appear that he could not care less. Of course, the fact that he stopped mid-stride, and turned to address her and did not just continue walking away might leave the crack in the wall open a bit. Did she win him back? I will leave that unspoiled for you to find out.

We are here to talk about determining the sentiment of what is said. In the previous paragraph, we were looking at the line from the movie. But what about trying to determine the sentiment of the people with whom you are interacting? Your customers and clients.

How would you adapt your decision-making if you had reliable sentiment analysis at your fingertips on demand? This automatic sentiment analysis is exactly what AI can provide those of us who have to work with people on a regular basis. Whether in customer service or support trying to ensure that the customer is always happy (whether they are “right” or not) or whether in education trying to ensure that knowledge transfer and learning is taking place by getting feedback on the delivery of the instructional materials and how well they are received, AI-supported sentiment analysis can be a marvelous aid.

But what exactly is “sentiment analysis”?

Well in short, it involves analyzing text (written or spoken) in context, tone, and nuance to determine whether the sentiment (opinion or thought) is positive, neutral, or negative.

What is your sentiment analysis (SA) of Cap’n Butler’s reply to Scarlett?

Is it: Positive (+1), Neutral (0), or Negative (-1)?

Natural language processing (NLP) and machine learning (ML) applications of artificial intelligence (AI) allow for the analysis of text, like Rhett’s statement, within context, like his walking out on a crying Scarlett after a tiff, to try and determine and interpret his sentiment towards her question of what to do next.

Likewise, when you might ask your stakeholders (customers, partners, students, etc.) a question, AI-based SA will attempt to determine their sentiment based on their reply when given within the situational context the question was asked.

There are different types of SA

IBM describes the top three in use (in no particular order) as fine-grained (or graded), aspect-based SA (ABSA), and emotional detection (IBM, 2024).

Graded SA looks to determine the level, or strength, of the emotion expressed by the responder on a scale running from 0-100. A strong reaction, like Rhett’s, should rate in the 90-100 range.

ABSA is looking for a more focused or targeted aspect of something (e.g., product first impression, or learning experience in a lesson). This type of feedback can help the pollster better understand what is/is not working and what might be needed for the next encounter or interaction. Think of it as a sort of Start/Stop/Continue exercise.

Emotional detection seeks to understand the psychological frame of mind of the responder at the point of answering the poll’s question. This can certainly be a much more complicated report than either of the other two types and instead of trying to determine whether a response is weak/strong, or positive/negative, it instead tries to identify what emotion is present.

From a decision support standpoint, ABSA looks to be the most useful from an organizational standpoint.

Think about this… if you were a teacher, and could ask your students at the end of a lesson what they thought (+,0,-) about the lesson overall and/or what they thought about the delivery of the lesson (not the lesson itself), known as the user experience (UX), that could be VERY beneficial to what sort of adjustments need to be made to account for the individualization that students are begging for (because they can get it on their phone/game/whatever) and learning technology can now start to provide. In this case, you would not necessarily have to change the lesson content, just customize the delivery.

Taking this thought a step further, in addition to using ABSA to personalize the lesson delivery, an aggregate ABSA (called Net Sentiment Score or NSS (Konopka, 2024)) could be used to take the pulse of the entire class and feed that real-time information to the instructor/teacher to tailor the course of the delivery for everyone anonymously rather than individually if that level of granular refinement isn’t available.

Going further still, an ABSA NSS across the school or district could begin to inform school administrators about what is working and what is not allowing them to do things like heatmap for problems and inform decisions about where to pilot reforms. Overlay that heatmap about the UX with grades from the lessons and it could be possible that coincidental occurrences feed each other in ways previously unseen or even considered.

While the concept of AI-based ABSA can quickly get the wheels of thought spinning, those wheels need a regulator on them as well. Just throwing AI at something will not magically generate the answers we seek. The AI, like a well place fastball, must be thrown with precision and strategic intent.

Without proper planning, the SA can suffer from poor question formatting (the SA may need to know the question and context in addition to the responses to glean meaningful insights). The responses may need to be multiple choice instead of free text to keep negation (“I wouldn’t say that”), sarcasm, irony, slang, or anything otherwise potentially problematic (like responding with Rhett’s line).

To be reliable and trustworthy, the responses from the SA should be dependable and consistent. As Det. Joe Friday said, “Just the facts.” With some planning to narrow the variables into expected responses that can be interpreted at face value, then the results of the SA should be trustworthy. With trustworthy and reliable decision-support information that can be updated in real-time directly from those that are impacted the most (the end users – i.e., customers, clients, students*), the best decisions for the advancement of both the school(s) and the student(s) can be made.

*Remember, in academia, your ultimate customers are the people hiring your students and your intermediate customers are your students. Just like a company that produces poor-performing products will cause end-user consumers to shun the brand, a school that produces poor-performing students will cause end-user employers to avoid your graduates. When your graduates cannot get a job, not only are they upset, but word gets out and enrollment will drop because who wants to go to a school for nothing?

References

IBM. (2024). What is sentiment analysis? Retrieved from IBM: https://www.ibm.com/topics/sentiment-analysis

Konopka, K. (2024, February 15). Net Sentiment Score – A powerful metric in sentiment analysis. Retrieved from Responsly: https://www.responsly.com/blog/net-sentiment-score/

Comments are closed.