Making the facts speak for themselves
Our texts are based on data which means we can guarantee accuracy
In journalism, a story begins with the facts. The facts, their interrelationships and the chronology are the constituent parts of a coherent and accurate story.
The same is true for the texts our robots write. They are based on data. Simply put – if a fact is in the data, it can become part of the story. Otherwise not.
When an editor asks a journalist to cover a story, facts and fact checking are at the core. The journalist investigates and analyses how the facts fit together and works out the sequence of events that form the story. What United Robots do is scale up human writing. Our NLG (Natural Language Generation) technology automates the process of fact based writing.
As early as 2017, Li L'Estrade, Head of Content Development at Swedish local media group Mittmedia, relied on United Robots to generate thousands of automated articles about house sales a year for their 22 local titles.
We work with the most demanding text clients in the world – journalists. Errors are out of the question.
In our process, a story starts with a set of structured data. A robot analyses the data to form the story, identifying the main angle(s) and events. The data analysis becomes the story.
With our data based NLG, we achieve:
Predictability. The quality guarantee is built into the technology. With correct data + correct conditions you always get a correct result. If something is missed, you can trace the error back, whether it’s a data point or a condition.
Variation. The goal for editorial purposes is to generate a variety of texts. With our technology we can generate dozens of texts from the same set of data points, all written differently.
Control. High end publishers demand full control over the content created.
Our approach enables human-like story writing techniques.
> The robot can consider whether something has already been mentioned in the text, instructing it in how to develop the story
> The robot can differentiate between how to write about different types of data points in the text
> The robot can provide texts with different points of view (home team vs away team etc)
> The robot can adjust its writing style and tonality depending on the publisher – journalists train the robots.
Wait, writing robots...hasn't Google already done that?
Sure, Google has developed tools which can generate text, and they are getting increasingly skilled. But Google’s NLG does not create text from defined data. Instead, the robot works out the most probable next word in a given sentence, based on existing text parameters. In other words, it bases its decisions on what has been written before.
The same methodology is used by the much-talked-about language generator GPT-3, which was used to write an opinion piece in the Guardian in September 2020. But although GPT-3 draws on a massive 175 billion parameters (existing text) and uses machine learning to form its sentences, the Guardian editors discarded 90% of the text generated (for eight different versions of the article) – and still had to edit what remained.
Crucially, the Google / GPT-3 NLG has no built-in way to guarantee factual accuracy, or even cohesive story-telling. At the core it’s about language probabilities. If you give GPT-3 a clear prompt for what it’s supposed to write, it may generate a story that’s linguistically correct and makes perfect sense. Until, suddenly, it doesn’t. The topic is the same, but the context may veer off. Because in text-to-text NLG, there are no predetermined data points to inform the story, and no structure for how it should develop.
"For a local newsroom, automation is necessary"
The story in a nutshell: At the tiny local newsroom of Bärgslagsbladet (BBLAT – part of Bonnier News Local in Sweden), automation of the routine reporting has freed up journalists' time to focus on the stories that drive reader engagement.
”I estimate we’d need two additional reporters to do the work the robots do for us today. I’m not sure that’s enough. The robot works day and night, it’s not off sick or at home looking after ill children. For a small newsroom, automation is necessary. We’re forever prioritising and sometimes I feel all we ever do is choose not to cover things. We know where to deploy our resources in order to make our readers happy. And if we can use technology and automation to perform tasks as well as we reporters would, there’s no doubt that’s what we should do.”
Helena Tell, Editor-in-Chief, Bärgslagsbladet, Sweden