Over the last several years, the topic of automation has gotten a great deal of attention. It’s fairly widely accepted that we are in the beginning stages of what has been called the Fourth Industrial Revolution. AI’s ability to automate many routine tasks is at the heart of the notion of the Fourth Industrial Revolution.
In parallel, I’ve been having some interesting conversations with clients and contacts about the use of predictive analytics, in the form of sentiment analysis, in chat channels. One way to think about this is trying to tap into the power of prediction by using AI to pick up on weak signals of sentiment or preferences that customers might indicate within a conversation (in this case via chat or messenger services).
What I find interesting about the ongoing work in this area is that what we are trying to get our algorithms as ‘listeners’ and ‘interpreters’ of chat-based conversations is to do something where most humans struggle too. As humans, we’re generally good at picking up on sentiment and communication cues when we are engaging with someone face-to-face. However, when you remove the subtle non-verbal cues that we pick up with our ears and eyes in face-to-face communication (e.g. email, texting, chat, messenger), this is where misunderstandings occur.
Here’s a question that we are trying to understand and test with our education and corporate clients: can natural language processing (NLP) replicate in chat what we humans are good at in face-to-face communication. Can we deploy or re-purpose algorithms to interpret the subtle cues that another person is giving off in their written communication? We know that this is already possible where the text is a bit more structured—emails, articles, research—but what about chat? Grammarly is an excellent example of a company that is doing some exciting work in this space. I use Grammarly (primarily for academic writing), and it has gotten good at helping to shape the ‘tone’ of my messaging based on what I assume is a form of sentiment analysis using NLP. What we are trying to do in our Noodle Factory platform is to think through how we turn the problem of sentiment analysis in chat into a problem of prediction.
One of the challenges that we are working through (like other companies in the chat space) is how to predict the sentiment, needs, wants, and unspoken messages in a medium that is notoriously informal. In this way, the language used in chat channels is a lot like a unique dialect that must be re-learned. A great example of this is the type of chat-based communication that happens informally, where I live in Singapore. The kind of chat that is used informally in Singapore is hyper-contextual because of some factors common in most countries (e.g. the use of emojis, slang, abbreviations, acronyms, etc., etc.). Sentiment analysis for chat runs into particular difficulty in Singapore because of ‘Singlish’. Most people recognise Singlish as the mix of (mainly) English, various Chinese dialects, Malay, and Tamil words plus many acronyms and abbreviations (a lot of them from the army). This hyper-contextual language in chat channels is exceptionally challenging to classify in NLP.
Here are the questions that I’m interested in exploring. Can Singlish, as it is typically used in chat and messaging services, act as a ‘high bar’ to test sentiment analysis? And can we use sentiment analysis of these chat conversations (primarily with customer service or customer loyalty agents or bots) to predict correlation with Net Promoter Scores (NPS)? NPS is an attractive metric for marketers because of its foundational simplicity. We have all been asked the ‘golden’ question that feeds into NPS metrics: “How likely are you to recommend our product or service to a friend or colleague?” The answers to that question, and subsequent Net Promoter Scores, help marketers to gauge sentiment. My contention, and the work we are aiming to do, is that chat-based conversations with agents and bots should be rich sources of patterns and independent variables that allow us to do a better job of predicting movement in NPS.
Context matters here. In this way, we might be able to separate out a chat-based sentiment analysis of something like Singlish as a problem of 1) pattern recognition, 2) contextualisation, and 3) the prediction of, or correlation to, NPS.
If we can better predict NPS through sentiment analysis in chat-based conversations, this is just the first step. Understanding these patterns could also allow organisations to map and predict causal links between earlier stages in the customer journey, likely sentiment in chat-based conversations, and ultimately metrics like NPS. Most importantly, if we can better understand these signals of correlation or causality across the customer journey, we can encourage a positive customer experience by getting ahead of the right actions and activities. This will allow us drive positive sentiment and a better outcome for our customers.