Data Science – What’s It Got to Do With Text Messaging and Customer Sentiment?

From the Rosetta Stone and the Meroitic script to the Indus Valley seals, scholars have an insatiable desire to decipher these texts so they can discover the secrets from that time.

In day-to-day interactions, language gives us insight into people’s feelings in the moment and what is important to them. Some companies have started to realize this important lesson and use machine learning, specifically natural language process (NLP), through their scholars (the data scientists) to better understand their customers and unlock hidden value to thrive.

Insurers should take note of this lesson. The insurance industry has become fiercely competitive. Over the last decade, carriers have turned to IoT (Internet of Things) and usage-based insurance as the primary driver of data-oriented differentiation, resulting in some carriers outperforming others. However, for late-moving carriers, the early movers’ advantage in usage-based insurance is hard to overcome because data capabilities are self-reinforcing. To remain competitive, late-movers must look to other untapped or underutilized sources of data, like text-based conversations. There are multiple use cases for text-based conversations, including:

• Sentiment: Word usage, sentence structure, punctuation, etc., can tell you a lot about how your customers feel. Add in how they respond to your messages, and you can begin to identify what drives their behaviors. You will gain insights from these insights on how to position your products and services and better understand how to deliver loveable policyholder interactions.
• Fraud: Text data offers a critical tool in dealing with insurance fraud. Analyzing conversational trends and specific conversation behaviors across different claims can help detect potential leakage points. Additionally, carriers can look at text patterns in medical bills and social media posts to understand sources of fraud.
• Risk: NLP models can use data from case law and claim files to chart the right course on litigation and settlement. Additionally, carriers can evaluate if new innovations have intellectual property potential by analyzing current patents in progress. Insurers can also build new risk assessment tools using NLP models to track contract language in policies across their book of business.

While there are more use cases out there, the most impactful is the ability to understand policyholders’ behaviors to improve underwriting strategies, especially if you can tie in data from usage-based insurance to truly understand the risk profile. Before we get deeper into pursuing these NLP uses cases, let us establish some ground-level definitions of artificial intelligence (AI) first to give a foundation for what to do next.

What is AI?

The basic definition of AI is computers or machines performing functions typically associated with intelligent beings. These computers use highly advanced mathematics and statistics to imitate cognitive and advanced motor functions such as recognizing a picture of a four-legged domesticated animal as a dog or cat, which is surprisingly complex when you think about each task that goes into making that assessment. AI can be narrow—the ability to do one dedicated task—or general—the ability to do multiple tasks (more of what you see on sci-fi shows and movies). For our purposes in insurance, let us divide AI into four major categories.

1. Vision: The ability to interpret and act upon pictures and videos
2. Speech: The ability to interpret and act upon sounds
3. Natural Language: The ability to interpret and act upon or generate written words
4. Decision-making: The ability to approximate models of human choice and actions

Underpinning all these categories is machine learning, which takes enormous datasets and discovers the underlying rules and relationships between known answers from the past to extrapolate future or unknown observations.

What are machine learning and NLP?

While AI has existed since the 1950s, the scope and expanse of practice areas under its umbrella have grown. Machine learning and NLP have been two of the most talked-about AI components since the last decade.

2012 was the defining year that created this exponential growth. The convergence of high-speed data expansion and availability, increased hardware and infrastructure performance, the proliferation of advanced algorithms or technical recipes out of the research labs to industry, and more company investment in research teams contributed to unbelievable advances in machine learning and NLP.

We are now at an age where data volume, complexity and velocity have surpassed human-level comprehension. We are inundated by so much data that we can easily miss true insights with less-than-optimal intuition or not make decisions fast enough to keep pace with changing data that risks are not easily mitigated.

When humans learn through experience, our capacity, even as an aggregate-whole, will not be enough. The traditional approach of working with and translating business requirements into solutions is no longer easy given how fast everything changes. This is one of the main reasons why machine learning was born.

If humans learn through experience, machines learn from data. Lots of data. As more data exists and more of the “right” data is available, the higher the likelihood that patterns can be found and leveraged to drive the complex mix of business rules that dominate our world. It is not about directly acquiring business rules to drive solutions; it is about using data to surface those rules. The equation has flipped around.

Technical specialists are now called upon to write instructions to “teach” machines to learn from data. Data scientists and machine learning engineers are needed to create critical recipes for a computer to know how to extract and surface patterns from data. This is how the term “machine learning” came to be.

NLP has also come a long way since the early days decades ago. It started out as matching words or a combination of words from a dictionary to drive basic insights from unstructured text. For example, word clouds are commonly used to reflect how frequently words appear and provide a sense of “popularity” of what those words represent. These matching algorithms built the foundation of the common sentiment analysis that you typically see in AI solutions.

However, this lexical approach to NLP takes a lot of human effort, and word matching is not a scalable solution for understanding that one critical component of language – context. Therefore, researchers have searched for ways to teach a machine to decipher human language with the proper context for a more profound way of getting to the nuances of how human communication and articulation are expressed.

The decipherment tool that researchers stumbled upon was transformers technology.

Why use transformers?

We often hear the term “deep learning” in many blogs, vlogs, news articles and websites nowadays. Deep learning is a type of machine learning that follows some of the principles of how our human brain works but is expressed and actualized in extraordinarily complex and sophisticated mathematical recipes and formulas. It is used to analyze data beyond the structured format and recognize patterns of unstructured data such as free-form text, images, video and audio. Deep learning is still a growing and expanding field, with advances happening at a rapid clip.

If we drill down on deep learning, we find a dizzying maze of components and building blocks to dismantle the data, in whatever form, into pure mathematical and numerical representations and formats. Transformers are one of these frameworks.

While transformers can work with various data, advances in this field make it especially effective in analyzing unstructured text in ways that have never been done before. Transformer technology converts and encodes copious amounts of unstructured data into numbers representing human context, partly by “reading” through scores of English language source materials spanning from books, novels, scripts, websites and news articles covering countless genres.

Transformers establish context through “encoding,” mathematically formulating relationships between how words are used, the proximity of words used with each other and remembering connections to other words. In addition, it will mathematically determine the propensity to generate certain words and phrases when given only a few words in the same way that when I mention “to be or not to be,” most people would know to follow that up with “…that is the question.”

The analogy is equivalent to training a human subject matter expert with a photographic memory to absorb all the English language’s unstructured text and materials and have near-perfect recall of all the words and sentences in context of each other. The only difference is that we are training a machine. That is the power, promise and potential of what transformers can achieve in the realm of natural language understanding (context) and generation.

How can transformers elevate what you do?

Businesses and companies can now harness the capabilities of transformers to create and innovate toward an unbelievable array of applications, predictive and prescriptive models around all that unstructured data. We can tap into and analyze these data sources as assets with a level of accuracy and precision never seen before. Rather than spending an inordinate amount of time and effort trying to maintain a dictionary and vocabulary of words, we can now focus on solving the business challenges we face.

For example, traditional sentiment analysis is not useful in the context of claims, which inherently have negative words like loss, damage and death. Instead, we could focus on feelings of frustration from persistent delays or lack of understanding because those often lead to poor claims experiences for policyholders. We can also look for adjuster biases across different interactions or adherence to the service levels required for the carrier brand to escalate when coaching might be required. Ultimately, carriers can hyper-personalize interactions so that policyholders are delighted at every moment of truth.

Additional resources 

To leverage transformers, ensure that you have a team of highly qualified and experienced data scientists who can help get you the actionable insights and foresight that can move the needle for the business. These teams can learn more about Transformer Models here:

• Attention is All You Need (Research Paper)
• “The Illustrated Transformer” (Article)
• “The Deep Natural Language Papers You Need to Read — BERT, GPT-2 and looking forward” (Article)