We have all had an experience with a chatbot....Good or bad, these robots do not leave us indifferent.
The impression can be excellent or disappointing, and transition from one state to another can happen very quickly. Here is how to design a chatbot conversation. Let's see what the tips are to overcome the limits of our chatbots.
Define the role and knowledge base of your chatbot in advance
Most chatbots are based on a set of rules that dictate the answer to a specific question by drawing the necessary resources from a knowledge base. The richer this knowledge base, the more complete the range of questions your chatbot can answer.
It is therefore essential to have considered most of the issues to which the user may be exposed before launching a chatbot project. To do this, the operational teams that manage the customer service must participate in the strategic framing and development of the knowledge base to provide the most accurate feedback on the various subjects raised during customer exchanges and create natural conversations.
The easiest way to set up a chatbot project is to start small and develop it according to a structured schedule.
Most often, we set up specific use cases on which we train the chatbot and make it evolve so that it can reach high comprehension rates, that is, above 90%. This learning phase is essential; it allows us to launch the chatbot in the best conditions.
The chatbot, your company's spokesperson
Ideally, and to respond to the widest scope, your chatbot should be integrated with your different tools: CRM, information system..., to provide personalised services in line with the company's positioning.
The chatbot represents your company when you are not available. It is therefore important from the very beginning of your project to decide on the tone, the degree of friendliness and humour that you want to give it. It's about giving it a real personality improving the interaction experience and its adoption.
It is possible that sometimes your chatbot cannot answer, don't panic... Think about setting up a livechat solution in addition to your chatbot, which will allow the bot to leave the hand to a human agent to bring an answer to the user. If you don't have a livechat solution, it is possible to consider a first step with the customer service phone number or email.
It is also possible to set up an email escalation directly via the conversation: the chatbot will then ask for the customer's contact details - if not connected - the problem encountered and send the email directly to the CRM or dedicated service.
How to maintain good performance?
Misunderstandings, loops, frustrations and blocked transfers can occur during chatbot conversations. These accidents are a common reason for chatbots to fail from a User Experience perspective (UX). Indeed, users often report that the chatbot does not understand and is far from being "smart". Has this ever happened to you?
There can be several reasons why a bot does not work:
- A knowledge base that does not cover the most important topics according to the brand's positioning
- A lack of training on past conversations that were not correctly answered by the bot
In both cases, the lack of time of the bot's administration team often triggers these misunderstandings.
And here is the key that will determine whether the bot will perform well in the long term or not: supervised learning! Although this is essential, it is often long and tedious for the teams. The bots' interfaces are not necessarily designed to facilitate training. Often, to reach correct levels of understanding, the amount of data that teams have to integrate is very large.
We give you our tips, so that your chatbot becomes and remains efficient:
1. It is necessary to enrich your bot regularly, to obtain good results quickly, and then a few hours a week to check and control that everything is working properly.
This control phase allows you to :
- detect misunderstandings and remedy them;
- fix errors;
- create new topics in the knowledge base.
2. All misunderstandings of the bot must be filtered and the topics and the reason for the misunderstanding must be identified:
- If the topic does not exist in the database, either it may not have been identified as important by the teams, then by using the bot we have the opportunity to discover new topics which are important for users and that we have to add in the knowledge base.
The solution is to create a new intention, associate the understanding linked to the user's sentences and train the bot.
- If the subject exists in the database, but the intention has not been recognised, then the training database must be enriched.
The solution is to train the bot on new sentences to optimise its understanding.
This follow-up is essential for the bot to function properly and is totally linked to its performance over time.
How can you optimise the training of your Chatbot with our solutions?
The advanced Worldline solution, integrated into the WL Conversational Platform, uses topic modelling and clustering. What can we use these tools for?
Thanks to machine learning, these tools allow the administrator to collect the questions asked by users, analyse them and group them into specialised clusters of intentions.
Once identified, each phrase group/cluster is divided into training and test data. The key phrases, central to the cluster and therefore most relevant to the identified intent, will be used as training phrases. The sentences further away from the centre of the cluster will constitute the test set. The training data will be used to enrich the bot and the test data to evaluate the bot after training.
This automation represents a very interesting time saving for the bot administration team. The analysis time is divided by 3 because it is automated and all the administrator has to do is check and correct in the clusters directly if necessary. This functionality allows both to process large volumes of sentences, which is essential for highly exposed bots, and to maintain high performance in understanding.
At the end of the day, monitoring conversations is less boring and processing is much faster. Monitoring your chatbot's performance becomes easy and enrichment only takes a few hours per month to maintain a level of excellence.
At Worldline, chatbots are sustainable solutions, maintainable over time and beneficial from a user perspective.
This is why we pay particular attention to the aspect of enrichment and training, which is the key to the misunderstandings and frustrations encountered in conversations.
To find out more, contact us for a demo!
About the author
Valentine Horstmann has been working in digital and eCommerce since 2007. Passionate about innovations and digital trends, Valentine has joined the digital banking team as product manager for WL Conversational Platform. In her current role, Valentine focuses on new offers, customers’ needs & user experiences to fit the evolution of conversational experiences.