Cybersmile wanted to develop a completely scalable and sustainable support solution for people affected by cyberbullying and online abuse on a global level. They envisaged a chatbot that could identify the needs of users and provide them with appropriate advice while signposting them to expert support organisations local to where they were in the world. They also told us about their ambition to have the system deploy on to partners websites and apps in the future.
We came up with a system architecture that would involve an embeddable chat interface that could be installed unobtrusively on any website, starting with their own. The interface would allow users to send text or audio through to an AI platform that we would programme to understand users’ requests and respond accordingly. We would also produce a user friendly content management system that would allow the Cybersmile team to curate helpful, relevant and safe content that could be returned to users once the bot had identified their needs.
We started by building the CMS to allow Cybersmile’s team to get to work building a library of content that would be delivered to users through the bot. It was important that they could use the system without technical knowledge, and without having to get involved with the complexities of the AI system. Content types were created for “articles”, “advice”, “organisations” and “helplines”, each with their own specific data fields. We also made it possible to restrict content to specific regions. This would allow their team to ensure that helplines were delivered that were appropriate to the user’s location and that articles would use localised spellings.
While Cybersmile were at work on their content library we built the interface through which users would communicate with the chat bot. Our design team came up with something that would be intuitive to use and would work well on both desktops and mobile devices. We developed this as a fullscreen overlay that would be triggered by a button that floated in the corner of the website. This approach worked great on the cybersmile site but was also chosen for it’s portability as it could be added with relatively minor customisation to Cybersmile’s partner websites in the future.
The last piece of the puzzle was to develop the conversational routes within the AI software. The main body of the conversational logic was focussed on identifying the core topic of a user’s query; be that “depression”, “anxiety” or even “crisis support”. We worked with the Cybersmile team to train the AI to understand a wide range of ways that a person may describe their problem when reaching out for support.
Once the bot could identify the topic of people’s problems we could then direct them ask them to refine what they were looking for by suggesting the range of content types available. By combining the topic, content type and the user’s location the bot could then find and return suitable content from the Cybersmile CMS.
Of course, to make the bot feel more life like, it wasn’t all business. We programmed the bot to understand and respond to hundreds of off-topic queries from users. These were designed to handle everything from the playful to the profane!
Until this point, our work on the AI was based on our linguistic skills and the Cybersmile’s expertise in both their topic and their audience. But once we were live, the focus turned to the real-world queries of users. By reviewing the conversation history of the system we could provide the bot with feedback that would help it respond with increasing accuracy. We could also identify and add topics that users were reaching out for that the system didn’t yet address.