Dialogue Systems for Academic Research: A Guide to Create Humanly Academic Chatbots
Matthew Dawit and Dr. Manas Gaur, Department of Computer Science and Electrical Engineering
Dialogue systems offer a powerful medium for academic dissemination, yet creating chatbots that effectively translate complex research for non-experts remains a significant hurdle. This guide addresses the lack of a comprehensive framework for developing ‘humanly’ academic chatbots that can present a researcher’s work intelligently and empathetically. We introduce a methodology for structuring academic work into a conversational flow, guided by core principles of natural language understanding (NLU), dialogue management, and personality design. The framework also includes a system for evaluating chatbot effectiveness based on user engagement and knowledge transfer. By following this guide, researchers and developers can create more engaging academic chatbots, thereby improving public understanding of complex topics and pioneering a new method for scholarly dissemination. This work seeks to advance the future of academic communication by making research more accessible and interactive.
Building trustworthy NeuroSymbolic AI Systems: Consistency, reliability, explainability, and safety
Dr. Manas Gaur, Department of Computer Science and Electrical Engineering
Building trust in AI involves the essential elements of explainability and safety, which demand a model to consistently and reliably perform. To achieve this, a combination of statistical and symbolic AI methods, rather than relying on either alone, is crucial for analyzing relevant data and knowledge in the AI application. We advocate and aim to illustrate that the NeuroSymbolic AI approach is better suited to establish AI as a trusted system. For instance, despite having safety guardrails, ChatGPT can still generate unsafe responses.
At the Knowledge-infused AI and Inference Lab at UMBC, the students will be dedicated to constructing the CREST framework. This framework will showcase how consistency, reliability, user-level explainability, and safety can be attained in NeuroSymbolic methods, using data and knowledge to meet the requirements of critical applications such as health and well-being.