A framework for cognitive chatbots based on abductive–deductive inference

This paper presents a framework based on natural language processing and first-order logic aiming at instantiating cognitive chatbots. The proposed framework leverages two types of knowledge bases interacting with each other in a meta-reasoning process. The first one is devoted to the reactive interactions within the environment, while the second one to conceptual reasoning. The latter exploits a combination of axioms represented with rich semantics and abduction as pre-stage of deduction, dealing also with some of the state-of-the-art issues in the natural language ontology domain. As a case study, a Telegram chatbot system has been implemented, supported by a module which automatically transforms polar and wh-questions into one or more likely assertions, so as to infer Boolean values or snippets with variable length as factoid answer. The conceptual knowledge base is organized in two layers, representing both long- and short-term memory. The knowledge transition between the two layers is achieved by leveraging both a greedy algorithm and the engine's features of a NoSQL database, with promising timing performance if compared with the adoption of a single layer. Furthermore, the implemented chatbot only requires the knowledge base in natural language sentences, avoiding any script updates or code refactoring when new knowledge has to income. The framework has been also evaluated as cognitive system by taking into account the state-of-the art criteria: the results show that AD-CASPAR is an interesting starting point for the design of psychologically inspired cognitive systems, endowed of functional features and integrating different types of perception.