How we did and what we learned from building a Covid-19 chatbot
When hit by an unexpected crisis like the global spread of a new infectious disease, people are in dire need of trustworthy and reliable information, especially since there is no scenario or action plan ready to tackle the unknown. The global spread of the coronavirus has caught millions of people off-guard, frightened and in despair. Thousands and thousands were trapped abroad in a situation where borders where abruptly shut, flights and other means of transportation cancelled.
This new crisis created many rumours and fake information about a multitude of topics, from “Is this just another flu?”, “How do I know whether I have it?” to “What shall I do now?”, “How can I protect myself and my family?” or “Where can I get tested?”. And governments around the globe had very different reactions to this new danger, often leaving their citizens in frustration on what to do and how to act. This was the context in early March when Tilde joined an initiative by our long-term client and business partner Tet, Latvia’s largest national telecommunications operator, technology and entertainment company. We jointly decided to create a chatbot that would provide trustworthy information about the disease, its symptoms and the actions to be taken to prevent getting infected among other important questions that people were asking at the time.
Building the Covidbot
Tilde is a leading European language technology company specialising in developing custom solutions with AI powered technologies like Neural Machine Translation, Speech Recognition, Speech Synthesis and Natural Language Processing tools. Our Language Technology team embraced the idea of a chatbot to help people in these difficult times with great enthusiasm. It took our joint team a week to build the first version of the informational chatbot, create a website, and start internal testing.
The first chatbot release was developed and deployed on Tilde’s Conversational AI platform Tilde.ai. We decided to name it Covidbot (Covidbots in Latvian) to clearly establish its purpose and limitations. It was easy to define our bot’s persona: it had to be friendly but formal and serious, to reflect its mission and context. Initial scenarios where created based on reliable local and international media and national government sources as well as by analysing frequently asked questions and discussion topics on social media.
Our experience with building conversational agents and virtual assistants proves that it is very important to get the bot out for user testing as soon as possible since our understanding of what and how people ask questions about different subjects is limited. Logically, this rule applies mostly to free-form dialogues. Tilde’s Natural Language Understanding models are customised by using examples: each intent (representation of what the user is willing to achieve) must have at least 5 example questions or queries (inputs) to learn to detect respective intents in user requests. One can ask for a location of a point of interest in multiple ways, e.g., “What is your address? or “Where are you located?”, but the same intent should also be detected in user inputs like “How can I get to you?” or “I cannot find your office”. Identifying appropriate intents based on topics of interest and collecting eventual user input samples is time-consuming. And this is in addition to the task of building a knowledge base of chatbot responses which contains not only ready answers to user questions but also pictures, infographics, tables, webforms, and useful links.
Covidbot scenarios and knowledge base
The Tilde.ai platform enables chatbot developers to use a wide range of functionalities and features. It is equipped with Natural Language Understanding technology that supports free-text inputs and context-handling features. Our Covidbot was designed using a blend of guided dialogue and free dialogue scenarios. As already mentioned, we started by collecting data about FAQ, relevant and reliable sources, global statistics, and related national and international regulations.
The most important dialogue scenarios where consolidated into a starting menu (see below). It covered the following topics:
- current official information (national and international),
- list of relevant contacts and information hotlines,
- information about the disease and what to do,
- information about entering and leaving the country,
- other related questions.
Our team kept adding new data – intents, input samples and responses – based on user inputs. Each piece of information was carefully vetted and was checked by at least four eyes before being accepted and added to the chatbot’s knowledge base. Currently, Covidbot’s knowledge base contains more than 240 intents with single or multiple answers to various topics and the information provided by Latvian public institutions: the Cabinet of Ministers, the Ministry of Health, the Centre for Disease Prevention and Control, the Ministry of Education and Science, the Ministry of Culture, the Ministry of Economics, and many others. We have been paying special attention to the visualisation of information, e.g., using infographics, tables, videos, etc. Covidbot’s responses are brief and, in the event that replying to the user’s inquiry requires a more extensive answer, it provides links to the respective sources. The dynamically changing content such as global and local statistics is renewed automatically by using open source APIs.
Unlike with other chatbot launches where our experience has been that people tend to test each new bot by asking silly and spurious questions (especially during the first month or two), Covidbot users have been polite and focused on the topics relevant to the chatbot’s purpose: asking questions about variety of topics related to the pandemic. Here is the list of most popular queries:
- statistics (number of infected people, death rates, etc., in Latvia and globally),
- news and official information (rules, restrictions, etc.),
- disease and personal health issues (i.e., symptoms, testing, medical help),
- isolation and distancing related issues (e.g., rules and regulations, coping with stress and psychological advice),
- information hotlines and other relevant contact information,
- repatriation and crossing borders.
Focussing on a rather narrow range of topics and the given circumstances probably are major factors affecting the successful performance of our Covidbot. The customised Covidbot’s Natural Language Understanding module can detect a user’s intent with a high level of confidence in almost 90% of all user inputs. This, in turn, is one of the main reasons of positive user feedback.
Going viral with our chatbot
Our initial idea was to make our chatbot available for deployment on different national and regional institution websites where people are looking for information related to the coronavirus pandemic. Since we opened the website www.covidbots.lv to the general public and started spreading information about our social initiative, Covidbot has gained wide publicity and positive popularity. As of April 28, Covidbot is deployed on the official websites of 13 municipalities including our capital Riga and other larger and smaller Latvian cities, as well as several NGOs, e.g., the Latvian Municipality Association, the Latvian Student Association, and others. We have received very positive feedback from different municipalities and NGOs regarding our bot as it has reduced the workload of their staff responsible for customer communication and support, providing instant and trustworthy information to incoming citizens’ queries 24/7.
Latvians are also getting used to communicate with chatbots in the public domain. Since 2018, when the Latvian National Registry of Enterprises in cooperation with Tilde introduced the first virtual public servant – chatbot UNA – more than 10 public institutions have “joined the club” launching their chatbots on the public sector language technology platform www.hugo.lv (developed by Tilde and the Latvian Cultural Information Systems Centre). This prior experience of using chatbots as a communication channel with state and local institutions has had an impact on our Covidbot’s popularity – over 20,000 communication sessions with thousands of users and counting.
Our joint social initiative introducing the new alternative communication channel using AI empowered language technologies provided us with a unique experience and created value for its users, the people of Latvia. Here are some lessons learned:
- It is very important to diversify communication channels during a crisis so that people can get access to reliable and trustworthy information.
- Getting to the end users early on is key to learning about the topics that are of real interest and testing dialogue scenarios as well as the quality of intent detection and responses.
- Users concentrate on relevant topics when they are eager to find specific information.
- Visualisation of information using different tools, e.g., infographics, videos, useful links, etc., helps to build an exceptional user experience in virtual communication.
- Domain adaptation of the underlying AI enabled Language Technology models is critical for better quality user intent detection.
- Use of APIs to connect with third-party information sources is an effective way of managing dynamically changing content.
- Engaging local municipalities and NGOs fosters quick dissemination of new means of communication.
- Working on meaningful projects facilitates positive engagement of the team and raises efficiency.
We will continue developing and propagating our Covidbot while there is a need for this modern communication channel and reliable information.
Business Development Director, Head of Language Technology
Developer survey: Since you are here and interested in our project, could you please spare a moment to share your concerns and answer 12 questions related to developing voice-enabled apps.