Abstract Current question answering systems are usually based on a knowledge base which is populated with domain specific knowledge and managed through Unstructured Information Management Architecture (UIMA). But drawback in this approach is that knowledge base may be grown with knowledge which is not relevant to the users connected with the system. In order to address this drawback we propose unsupervised knowledge accumulation algorithm which can monitor user preferences and acquire knowledge without any supervision of the system management unit. Basically, this algorithm learns domain of interest of each and every user connected with the system and extract knowledge from the web or from a given corpus. We have also adopted several Natural Language Processing algorithms to design this high-level algorithm. Knowledge modelling is done through a conceptual graph based knowledge base. This novel paradigm is evaluated with the help of several connected users and with more than 280 questions. We have achieved excellent accuracy during the evaluation phase. It shows our novel approach is effective and can be used to address the drawback decently.