17/1/2018

Chatbots Gone Wild

 

The data science behind making a chatbot worth talking to.

Heard of Ken Schwencke? He's the pseudonym for one the world's first journalism bots, famous for breaking the story of a 4.4 magnitude earthquake in Los Angeles back in 2014. 

Three years on, bots have emerged as a unique tool for newsroom to promote audience engagement. Known as 'chatbots', this technology allows audiences to ask questions and participate in conversations about particular stories.

For example, French newsweekly L’Obs followed four undecided voters during the 2017 presidential election, using a chatbot to share their thoughts. Over in Spain, PolitiBot also helped voters keep up with their elections. And, for something a bit lighter, The Huffington Post's Felix assists users to find the perfect movies to watch on Netflix.

But how do you build a chatbot that audiences want to talk to? This is the question posed by Figure Eight's most recent e-book Chatbots Gone Wild.


Available for free download, the e-book provides a good starting point for newsrooms thinking about using a chatbot. Rather than going into the technical details of development, it looks at the core of what makes a chatbot successful: conversations.

"...because conversations are so commonplace, we don’t often think about how complex they really are. They’re full of subtlety, convention, and nuance. And that makes building a chatbot that actually works incredibly difficult. After all, users expect their questions answered and, if your chatbot can’t do that, they’ll go find an experience that works," explain the authors.

To this end, the e-book unpacks how a conversation works, and applies this to a chatbot setting. 


Once you understand these discrete components, the e-book guides you through how to train, test and finetune conversations using four key elements:

  • Utterance: How many ways are there to say the same thing? Chatbots need to understand as many as possible to avoid getting perpetually confused. 
  • Relevance: Is a particular response relevant to a particular question? 
  • Intent detection: Does the chatbot understand what is being asked?
  • Entity extraction: There's a difference between "this burrito is so bad" and "I want a burrito so bad". Entity extraction is a good way to teach an algorithm the subtleties of language. 

For non-coding journalists, the e-book is a valuable resource to get across the intricacies of building a chatbot, without becoming overwhelmed by the complexities of development.

Explore Chatbots Gone Wild here.

Image: James Royal-Lawson.

Comments