We are developing a chatbot to supplement the Idea Catalog to provide more knowledge on nuclear decommissioning. The ultimate aim is to assist users in matching (technological) needs with potential solutions by having a conversation with an automated computer program. There is still a long way to go, but this blog explains why we are doing it, how we have gone about it, where we are up to and where it will develop in the future.

What is a chatbot?

A chatbot is an automated computer program with the ability to simulate conversations with humans. There are two types of chatbots: rule-based and machine learning-based (Schlicht 2016). Rule-based chatbots are more limited, can only respond to specific commands, and are only as “smart” as they are programmed to be. Being “smart” for a chatbot refers to how well it can respond and hold conversation given any particular input by the user. Machine learning is when the chatbot has access to data that allows it to learn patterns and gain information as they are conversing with a human. Machine learning-based chatbots use natural language processing, or the ability to process human language and communicate back with similar language, rather than just commands, and they “get smarter” as they are used. We are using machine learning.

History

Chatbots originated in the 1950s and ‘60s when scientists began considering the idea of computers communicating like humans do. Alan Turing created the Turing Test that would challenge a player to guess whether the response to their questions was coming from a computer or a human. Soon after, Joseph Weizenbaum invented the first coded chatbot, called Eliza, to imitate a therapist. This time users knew they were communicating with a computer program rather than a human, but people still became attached to Eliza and the program became quite popular in its time (Bayerque 2016). However, due to limited technology, there was a lull in chatbot popularity following Eliza until the rise of smartphones in the early 2000s. Since then, chatbots have been rapidly growing alongside the increase in popularity of smartphones and messaging apps and the advances in technology and artificial intelligence (AI) constantly being made. For example, Facebook opened its messenger platform to chatbots in April 2016 and currently has over 300,000 bots on the app, up from 100,000 in April 2017 (Johnson 2018).

Benefits

The benefits to businesses of using chatbots include both convenience for customers and financial benefits for companies. One study based on a survey of over 1000 adults in the US asked what benefits users would enjoy most from using chatbots over phone calls, emails, and other lines of communication (Devaney 2018). It was found that the top three potential benefits of chatbots are:

24-hour service,
Immediate responses, and
Answers to simple questions.

These benefits are possible because chatbots can carry out many functions of apps, websites, or human representatives in a much quicker and cheaper way. By replacing the need to speak to a human representative, they can provide 24-hour service, accommodate a high volume of requests at once, respond immediately, and severely decrease the marginal cost and handling time of more requests to nearly zero (Accenture 2016). Chatbots eliminate that wait time. They are easier to develop and launch than apps and don’t require the user to download anything to their device or learn a new user interface. They also have an advantage over websites by allowing the user to ask simple questions that may be hard to find on difficult to navigate websites (Asher 2017).

There are few downsides to using chatbots. They can be very time-consuming to build: Amazon has over 5,000 people working on the Alexa voice technology platform (Kim 2017). Additionally, some customers prefer to speak to a human representative. However, often chatbots have the ability of directing the user to a human representative if it cannot handle the user’s requests. Because of this, customers that prefer a chatbot can use it to quickly answer their simple questions and customers that prefer not to use a chatbot can speak to a human representative instead.

These benefits, we believe, are as applicable to the nuclear decommissioning sector as any other. Sure, there are areas that will always be too complex for a chatbot, but there are many others that will be applicable and we can integrate into the Idea Catalog..

Steps for Building the chatbot

We used the following four steps for building our chatbot:

Planning the chatbot;
Designing and mapping conversations;
Implementing conversation flows;
Making the chatbot more human-like.

The first step is planning the chatbot and determining what problem the chatbot is trying to solve. It is essential to know the purpose of the chatbot and what features it must have in order to accomplish its primary goal before beginning to build it. We have started with a narrow scope because it is easier to make the chatbot successful and the conversations can be expanded on later.

Once the goal of the chatbot was established, conversations can be designed and mapped. We used a decision trees to map the directions different conversations can go and how a chatbot should respond to certain user inquiries. A decision tree is a model that begins with the user’s initial input and expands down different conversation paths as the user continues conversing with the chatbot.

When implementing conversation flows, the chatbot must be able to understand the user’s intentions. Even if the conversation is mapped, there are a variety of ways to say the same thing, people can use different syntax, and people can have misspellings in their text. In the implementation, a chatbot should be able to account for all of these differences and possible misspellings from the user. This is called synonyms are we have spent quite a lot of time on thinking about variants.

Lastly, a successful chatbot must also have human-like qualities and a “personality” that mirrors the brand (or is at least polite and helpful). By having a personality, the chatbot is more relatable, and users enjoy their interactions and are more likely to use the chatbot again. We decided upon a “professional” response rather than matey.

Where we are now

Our chatbot, Genie, is up and running. It is available in the Idea Catalog, of course, but also on Google Assistant. This did take some effort to get through Google’s approval process but we did it!

Currently we are concentrating on answering basic questions on nuclear. If you have a subscription for the Idea Catalog, just go to the Genie icon on the top navigation bar and type in, or say, your question. If you have Google Assistant on your phone of Home Hub, initiate Genie by saying “Hey Google. Talk to the Genie chatbot”.

“What do you know about [Site]”. For instance Savannah River or Sellafield.
“What do you know about [Facility]”. For instance MSSS or PFSP.
“What do you know about [Company]”. For instance Imitec or Longenecker.
“What do you know about [Topic]”. For instance Distinctive or Transcends.

Next Steps

We are gradually increasing the scope of content by adding more details on Sites, Facilities, Companies and Topics. So, if Genie can’t help at the moment, it will do in the future - it just needs time to grow!

We will also be introducing the next level in conversation that links to the Idea Catalog. We have completed one on “Pipeline Characterisation” that guides the users through identifying potential solutions for crawlers depending on the pipeline dimensions. This is in the Google approval queue and it should be released in a month or two.

We have a forward plan of other things to incorporate but please let us know if there is anything you would particularly like to see.

References

Thanks to Shannon Armold for much of the background and history on chatbots. She did a project for us as part of her London School of Economics MSc Operations Research and Analytics called “Building a chatbot to Facilitate Swing Weighting for Multi-Criteria Decision Analysis”.

Accenture. 2016. Chatbots in Customer Service. Accenture Interactive, downloadable from: https://www.accenture.com/t00010101T000000__w__/br-pt/_acnmedia/PDF-45/Accenture-Chatbots-Customer-Service.pdf.

Asher, Natali. 30 May 2017. Application of a Chatbot as a Facilitator for New Hires Onboarding. Linnaeus University, downloadable from: http://lnu.divaportal.org/smash/get/diva2:1116842/FULLTEXT01.pdf.

Bayerque, Nicolas. 15 August 2016. A short history of chatbots and artificial intelligence. VentureBeat, retrieved 2 August 2018 from: https://venturebeat.com/2016/08/15/a-shorthistory-of-chatbots-and-artificial-intelligence/.

Devaney, Erik. 23 January 2018. The 2018 State of Chatbots Report: How chatbots are reshaping online experiences. Drift.com, retrieved 10 July 2018 from: https://blog.drift.com/chatbotsreport/.

Johnson, Khari. 1 May 2018. Facebook Messenger passes 300,000 bots. VentureBeat, retrieved 20 Aug 2018 from: https://venturebeat.com/2018/05/01/facebook-messenger-passes-300000-bots/

Kim, Eugene. 27 September 2017. Amazon has 5,000 people working on Echo and Alexa -- more than Fitbit and GoPro combined. CNBC, retrieved 20 August 2018 from https://www.cnbc.com/2017/09/27/amazon-has-5000-people-on-echo-and-alexa-more-than-fitbit-and-gopro.html

Schlicht, Matt. 20 April 2016. The Complete Beginner’s Guide to Chatbots. Chatbots Magazine, retrieved 10 July 2018 from: https://chatbotsmagazine.com/the-complete-beginner-s-guideto-chatbots-280b7b906ca.

Seed, Ian. 2017. Guide to Multi-Criteria Analysis. Reading, UK: Cogentus Consulting Ltd.