Why Data Storytelling is Essential for Conversational AI
By: Nate Nichols
I joined Narrative Science almost 10 years ago, while I was in the midst of finishing my Ph.D. in artificial intelligence from Northwestern. At the time, I was dead set on becoming a professor and had verbally accepted a tenure-track position. But then my advisor approached me about a really interesting problem he was solving—automatically writing stories from data. He was starting a company and wanted me to come help out for a few months. What started as a consulting engagement has resulted in a decade-long career that has been one of the best moves of my life.
So why did I do it? There are a few obvious things—the people, the culture, the perks. But my real answer? Data storytelling—writing stories and having conversations with your data—is an incredibly complex, difficult, and exciting problem to solve. Maybe more importantly, I felt deeply that it’s a problem worth solving, and one I was eager to spend a good chunk of my career pushing forward.
Language is Hard for Computers
There is an obvious difficulty in support of artificial intelligence (AI)-powered conversations: conversations involve natural language, and human language is hard for computers. This may be surprising to some people, particularly those not well-versed in computer science.
Computers have always struggled with the complexities and ambiguities of human language. Humans communicate through stories and conversations, and AI-powered conversation requires solutions for a number of problems in the language space. Text-based stories and conversations require solutions for natural language understanding (NLU) and natural language generation (NLG). Voice-based conversations also require solutions for speech recognition and text-to-speech. Let’s look briefly at the current state in each of these areas.
The First Frontier: Data Storytelling
In order to enable truly conversational artificial intelligence, we first need to start with data storytelling. The ability to truly automatically write stories from data, without any configuration or building from the user, is incredibly complex. At Narrative Science, we’ve spent the last decade building technology that truly understands, analyzes, and writes like a human.
Our products utilize proprietary analytical frameworks, our natural language generation engine, and an AI-powered relevancy framework in order to write personalized stories from data.
Why Lexio’s Data Storytelling Capabilities Enable True Conversational AI
That list of requirements for AI-powered conversations can seem daunting, and indeed, many of those capabilities are extremely challenging to develop.
Fortunately, Narrative Science has already developed these required capabilities within our new product, Lexio. Lexio empowers everyone in your organization to understand data by automatically analyzing your data and transforming the analysis into a plain-English, data-driven story that naturally articulates the most important and interesting information to the intended audience.
While there’s more work to be done in extending Lexio’s conversational capabilities to make AI-powered conversations truly pervasive, the foundational infrastructure is already built, and wholly new capabilities do not need to be created. Truly conversational AI must start with data storytelling.
Let’s look at five requirements of a good conversation to see how Lexio’s existing data storytelling functionality aligns to the requirements for what is required to support AI-powered conversations.
1. An Understanding of the Conversational Domain
Lexio is domain-agnostic and has the ability to write data-driven stories on any topic. Lexio can be taught a new domain (such as “sales performance” or “franchise restaurants”) by specifying the entities in the domain (salespeople, restaurants, etc.), relationships between the entities, typical values, important thresholds, sentiment (is a large value good or bad?), and the language to use when expressing that domain.
This is precisely the type of domain knowledge needed to tell great stories—and, conveniently, also for a conversational system. A deep understanding of the domain, the ability to access data as needed, and a vocabulary for discussing the domain will allow Lexio to comprehensively answer questions and surface valuable information in a conversational setting like a domain expert.
2. An Understanding of the Reader’s Goals
Lexio’s model of data story writing is focused on answering the user’s question, and then structuring the resulting ideas and expressions into a coherent, cohesive story. For example, a user may want to understand their sales pipeline, know what’s driving an uptick in visitors to their site, or learn about the throughput of their engineering team. Once Lexio understands what someone is looking to learn by reading a particular story, it knows:
- The types of data necessary to answer the question
- The kinds of analysis that should be performed
- What constitutes an especially interesting or unusual outcome
- What is most important in answering this question
- The follow-up questions a reader is likely to ask
This foundation of a goal-oriented approach in data storytelling lays the groundwork for AI-powered conversations, as it enables the system to have the context necessary to engage in relevant and valuable dialogue. For instance, understanding the reader’s actual goal for the conversation helps the AI know when and how it should go beyond the specific question being asked. “Did we hit our bookings target last month?” is technically a yes or no question.
Understanding, though, that the reader has an overall goal of digging into their recent sales performance gives the system enough guidance to answer even a simple yes or no question with a response like, “You just missed your goal for last month because of a few key deals slipping into this month, but assuming no more slips, you’re on track to exceed your goal for this month by 22 percent.”
3. An Intelligent Natural Language Engine
Unlike basic NLG systems, which queue off of templates and can sound robotic, Lexio writes its stories the same way a person does. It has a set of ideas it wants to convey, and it works to convey them in the most natural and easy-to-understand way possible. It has a deep model of how humans communicate and how natural language works, and it uses those at runtime to generate stories that sounds like a person.
To support conversational interactions, the Lexio NLG engine will need to be extended to support dynamic settings such as “verbosity” and “tone.” These types of settings will “steer” Lexio’s NLG engine toward capabilities such as generating shorter sentences or using more colloquial language.
4. An Ability to Articulate the Most Valuable Thing to Say
Lexio understands a wide array of different ideas and situations needed to articulate what’s most important in the form of a story—and many (if not all) of these are also common to data-driven conversations. These ideas include both quantitative statements (“Mary was the top salesperson by bookings this quarter”) and qualitative assessments (“Mary carried the team this quarter”). These valuable insights that drive Lexio’s stories today will constitute the makeup of Lexio’s conversations tomorrow. And as we continue to expand the stories Lexio can write and develop our library of ideas and assessments, they’ll naturally serve in both storytelling and conversational use cases.
5. A Memory of Previous Conversations
As Lexio’s knowledge base grows, so too will its memory, with the ability to reference historical conversations and deliver even more contextualized dialogue. To write a great story, Lexio needs to know a lot about what it said previously in the same narrative. This memory is used in a variety of ways, ranging from the simple (e.g. using a pronoun to refer to an entity that’s already been introduced) to the more complex (e.g. using contrasting words such as “however” to contrast what Lexio just wrote with what it’s about to write).
These capabilities will need to be extended to support conversational interactions, and the scope of Lexio’s memory will have to grow from “within a narrative” to “between conversations.” These advances will build on Lexio’s existing framework and also benefit Lexio’s ability to generate narratives (e.g. reporting that includes information such as, “Like we predicted last week”).
The ability to create truly conversational AI is extremely exciting. In fact, we’ve even started to test this out within our Incubation team. Check out demos of our latest projects here.
The Future is Now
At Narrative Science, we will continue to push the limits of data storytelling and conversational AI by always questioning what’s next. What if your data told you the story you want to know every single day? What if your data talked to you? That’s what we are building every day. The culture of innovation is why I have been here for 10 years, and why I’m more excited for the future than ever before—both for Narrative Science and all of the people who will be affected by this amazing technology.
Data storytelling is coming—and we are so excited to bring it to you.
Follow Our Story
Like the what you read here? Share your email to get blog alerts for more content like this.