Enhancing Reliability and User Experience in Conversational Agents

Conversational agents are in place in a variety of domains and tasks such as sales and customer support services in business, student counselling in education and medical services in healthcare. There is abundant data available for modelling dialogs because online chat has been a popular way of communication between humans already for several decades. There are also innumerable other valuable digital resources that can be exploited when building a conversational agent, including pretrained large language models. We tested and evaluated several ways of preprocessing and modelling of chat dialogs in Finnish. As a result, we found out that the best accuracy is achieved using uncasing and spell-checking in the preprocessing phase and a BERT model pretrained with Finnish in the modelling phase. Despite the extensive use of conversational agents, there are still many open research questions. One example is the effect of the interaction style of the agent on user experience and emotions. Our initial study suggests that chatbots including small talk are less likely to elicit negative emotions, whereby emojis and emotional statements issued by chatbots do not play a significant role on the user’s emotional responses. We also discuss how medical expert work may be partially automated and made more interesting as input for routine conversation is handled by a chatbot. Special attention is paid on the requirements for trustworthiness and reliability for conversational agents acting in different tasks and domains.