Natural Language Interface to Enterprise Data Management

This 2-part talk will explore the development and use of LLM-powered technology to enable data consumers to easily explore and use business data and to automatically design business processes.

The first part will look at the core problem of text-to-SQL translation, generating SQL code directly from natural language utterances to query a database. We'll discuss a number of challenges associated with the problem, particularly for enterprise use-cases.

We'll explore techniques for addressing such problems, including pre-training, fine-tuning and prompting techniques for Large Language Models (LLMs) and state-of-the-art models, including some from IBM Research.

In the second part, we'll discuss a text-to-SQL pipeline that we have been developing at IBM Research, which can handle a range of client databases - each with different structures, sizes, and expectations. This means making the pipeline highly configurable, as well as being able to run efficiently on smaller models, since large-scale LLMs are not always an option.

We will also talk about how we are incrementally improving performance by benchmarking and identifying failure points through error analysis. While Text2SQL is far from a solved problem, our goal is to better understand its practical limitations and see how useful we can make it in real-world scenarios.

Speakers
Speaker 1: Dr. Ndivhuwo Makondo
Dr. Ndivhuwo Makondo is a research scientist and manager at IBM Research | Africa, leading a team of scientists and engineers conducting basic and applied research in the integration of knowledge, reasoning and learning and natural language processing (NLP) for various scientific and business use-cases. He is also a visiting researcher at the School of Computer Science and Applied Mathematics at the University of the Witwatersrand. He has a PhD in Computational Intelligence, MSc (Robotics) and BSc (Control Engineering) in Electrical Engineering. He joined IBM Research | Africa in 2019 and was previously a robotics researcher at the Council for Scientific and Industrial Research (CSIR).

Speaker 2: Richard Young
Richard Young is a Research Engineer at IBM Research Lab in Johannesburg, where he has been since 2018. He has over 14 years of experience as a professional software developer and has worked on projects in a wide variety of industries, including banking, health insurance, e-commerce, geospatial analytics and education. Richard is part of the Data and AI Platform theme at IBM Research, where his current work involves leveraging foundation AI models to extract meaningful insights from data, with a particular emphasis on developing a text-to-SQL pipeline for enterprise use cases.