In late 2019, Invoca announced the launch of Signal Discovery, the latest addition to the Signal AI conversational analytics suite. While the product is amazing and the technology groundbreaking, its true awesomeness lay in the people who helped create the algorithms that make it work. Since there’s no way my catchphrase-focused marketing mind could explain that process to you, we interviewed Invoca Data Scientist Mike McCourt and Director of Data Science Victor Borda to explain what goes into developing the AI that makes Invoca the technology leader in call tracking and analytics. Without further ado, let’s turn this over to Victor and Mike!
Victor: When we were first designing Signal AI, we wanted to provide a breakthrough experience for our customers that would provide the accuracy of customized predictive modeling in an easy-to-use platform. We needed to make sure that it didn’t require our customers to gather huge masses of training data while providing them with maximum transparency and control. We set the standard in our industry when we initially released Signal AI, doing all of the above with real-time Signal AI predictions tagging calls within milliseconds of a call ending.
It’s a great feeling to see how integral Signal AI has become to many of our customers. Our newly released Signal Discovery product really ties all the pieces together by automatically analyzing our customer’s full set of calls, showing them what sets of conversations are taking place, and then allowing them to turn those conversations into new predictive models with just a few clicks.
Victor: The Signal AI “Custom Signals” feature is a form of predictive modeling. It takes user-labeled training data and builds a model to automatically label new phone calls. Those training labels can be wrong, and most AI techniques are very sensitive to mislabeled data. We invented a way to automatically detect and remove mislabeled data. It works great, and it’s something we’re quite proud of.
We also found that the traditional methods for measuring predictive model accuracy produce scores that don’t intuitively match our customer’s interpretations of the numbers. That’s why we improved our performance scoring tool in a way that is mathematically pure and matches customer expectations.
For Signal Discovery, the latest feature in our Signal AI product, we not only had to figure out how to greatly extend the state of the art in topic modeling to apply to phone calls, but then we had to invent a technique to automatically partition the results into training data for our customer’s predictive models. This enables our customers to review the Signal Discovery topics and automatically create new signals from any of the topics.
Mike: Phone calls are idiosyncratic in the world of natural language processing. Phone calls can be repetitive, can contain both recorded messages and human speech, and often suffer from bad connections. Phone calls also contain both real conversations and sequences of simple yes/no answers. Needless to say, it’s very hard to design AI software that can juggle so many competing needs!
Most well-known AI models for language are designed for either long, carefully edited texts such as Wikipedia pages or news articles, or for very short, spontaneous texts such as tweets. In our experience, none of these well-known models work well on phone calls, which are both long and spontaneous. Since our needs differ from the rest of the research community, we have to do our own research and development.
Victor: I am inspired on a regular basis by my team members. We’re all-in on explicitly using Bayesian statistics to power our AI, which gives us tremendous power and flexibility. It doesn’t matter what specific technique, or more typically, an ensemble of techniques, we are using for a given problem — our depth with Bayesian thinking allows us to dissect, customize, and optimize in a way that would otherwise be impossible. It allows us to move freely and make the leap to take results from okay to stunning.
All of our data scientists are also coders, and they know more about the details of the Intel instruction sets than most programmers. That’s a powerful combination, and it allows us to ask “what if”, and then go do it quickly and effectively.
Mike: Our needs are a bit different from mainstream applications of AI to language. Surprisingly, we typically take our inspiration from the genomics research community: finding gene patterns in DNA data has a lot in common with extracting outcomes from phone conversations. In particular, no two DNA sequences will be alike, and you must always be prepared for patterns you haven’t seen yet — phone calls are similarly unpredictable.
This analogy led us toward nonparametric Bayesian models, which are common in the genomics community, rather than word vectors and recurrent neural networks which are most common in natural language processing.
Victor: We have many savvy enterprise-level marketers who know that optimizing their martech strategy translates to millions of dollars of revenue. We continually seek feedback from them. Our new Signal Discovery feature took a cue from customers asking us to show them what sets of conversations were taking place on inbound phone calls — to show them what they didn’t know was happening. They were already excited about what our existing Signal AI predictive models were doing for them, and now we’re helping them to understand new critical aspects of their business that they previously didn’t have a view of.
Mike: To run a business, you need to understand what motivates your customers. The good news is that you already have this information — your customers are picking up the phone all the time to tell you. Unfortunately, phone calls are messy and difficult to analyze. That’s why we provide a range of AI models to help you analyze your calls: we have Signal AI, which listens to your calls and automatically buckets them into categories you’ve defined, and Signal Discovery, which passively analyzes your calls and produces a “conversational map” outlining recurring thematic structures.
Victor: Over the last year, we’ve built a world-class AI engine powered by a Bayesian statistics view of data. Signal Discovery is the first customer-facing feature using it, and the results have been tremendous. The engine seamlessly and automatically handles the intense mathematics required, which enables us to go from idea to working code in minimal time. It’s like upgrading to a supersonic jet. It’s going to be a game-changer for us and our customers.
Mike: The engine we developed for our new AI models is in fact extremely powerful. We created our own proprietary deep learning platform, along the lines of Google’s TensorFlow, but better suited for language data. We also implemented a programming language for specifying our AI models. Our new tools significantly lower the technical barrier to inventing new products, and we’re only limited by our imagination at this point!
