Towards Reliable Conversational Data Analytics

Abstract

Conversational AI systems for data analytics aim to enable the extraction of analytical insights by means of conversational inter- faces. Such interfaces are powered by a mix of query modalities and machine learning methods for analytics, and are relying on Large Language Models (LLMs) for natural language generation. However, critical challenges hinder their adoption. The question we discuss is how to devise reliable Conversational Data Analytics (CDA) systems producing timely, consistent, and verifiable answers. To reach this goal, we identify five properties that impose a paradigm shift in the way systems are built and in the way they interact with users. To illustrate that shift, we describe a proto- typical CDA system. Realizing these properties involves either extending existing components, or redesigning components from scratch; both solutions require overcoming data management challenges and conducting a tight integration with advanced data management and machine learning techniques.