Advancing beyond current sequential rules, towards multi-signal memory to generate reponses more tailored to the users need.
State of art: Advances in natural language processing have enabled the development of a new generation of conversational agents (or “chatbots”), more capable in understanding natural language and retrieving or generating more natural and engaging answers. As a result, AI chatbots are becoming increasingly popular in many areas, including business, healthcare, and learning. Based on goal, chatbots are categorised into task-oriented and conversational (“open ended”). Based on technical approach, they are categorised into rule-based, retrieval based, and generative-based chatbots.
Empirical evaluation suggests that task-oriented chatbots are best served with rulebased approaches, while conversational chatbots require the use of generative (and based on the sophistication level, retrieval) based approaches.
Socratic method as a means to inquiry the user has a significant presence, especially in areas like mental health, and education. In this context, UNIVDUN has been amongst the first who have applied the Socratic method in the field of fighting disinformation, with the “News Immunity Chatbot” .
Challenge: The vision of TITAN is to develop an open-domain coach, able to guide the user in as many situations as possible. Based on the current state-of-art, this requirement suggests the need for a generative based approach. State-of-art chatbots like DialoGPT, Meena and BlenderBot employ large neural models (typically Transformers) in a sequence-to-sequence architecture, frequently selfsupervised on large amounts of text and trained on crowdsourced dialogues. However, TITAN’s conversational coach requirements go beyond what this standard approach in state-of-art can achieve, due to the “ingredients” that must be taken into consideration for response generation, including:
Disinformation signals stemming from integrated AI tools and services for the content at hand;
Citizen’s critical thinking level assessment;
Long-term memory of previous engagements with the user; and
The Socratic method for guiding the user to reach a logical conclusion.
Going beyond: The need for augmenting generative models with external signals or knowledge has already been identified as a challenge in the literature, constituting a recent area of research . TITAN will investigate how multivariate signals can be incorporated during response generation, following successful attempts to incorporate external knowledge, (like search engine results), into generative chatbots, like BlenderBot, enhancing approaches like “Fusion in Decoder”. Special focus will be placed on the various methods that the Socratic method can be integrated into generative chatbots, through research on two dimensions, incorporating it either as a signal, or as a “personality” skill, which has been successfully used in BlenderBot to acquire the skill of empathy.