Abstract: LLMs have been in limelight of AI researchers for some years and after the advent of ChatGPT they have brought an AI storm amongst masses. These tools are having there moments in the hype cycle. It is important to explore how these generative AI tools can make a leap from consumer applications to real world enterprise grade applications. To be useful in real world, one needs higher precision and consistency. These LLMs have great understanding of the natural language and its semantics. They can well understand what task is being asked from them and they can obtain relevant information present in their pre-trained corpus with good accuracy. In general, for an enterprise grade product, it is not about getting the job done, but getting it done well. To fulfil a task with good precision LLMs need knowledge, this is where they lack the most. They have built a good linguistic and a general sense of understanding on the massive publicly available data. But they lack business knowledge or domain knowledge of any area, for which the product is intended to be designed for.
On the other hand, Ontologies provide a structured representation of domain knowledge, defining concepts, relationships, and properties in a closely inter-connected graph format. Since LLMs excel at understanding and generating unstructured natural language data and Ontologies excel on relational structured data. So, when we combine the two, the unified AI system can take advantage of both structured and unstructured knowledge, leading to a more comprehensive understanding of a given domain. Leading to more precise and consistent results to build a real-world enterprise grade product.
In this tutorial, we will do a hands down walk through of techniques and concepts that will help in effectively using LLMs for our respective real world use cases with acceptable precision levels. We will be picking use case from Healthcare domain to illustrate the concept, we will specifically choose precision ICD-10 code detection using LLMs.
The flow of tutorial will be in following order:
• Examples of hallucinations for chat based LLMs (chatGPT)
• Techniques on giving LLMs time to think – Prompt engineering
o We will re-run the earlier examples to see how more context improves precision
• Observing increased precision with relevant context
• Ways to add context: {Fine-Tuning, Ontologies}
• Quick example showing increased precision using a Fine-Tuned LLM
• Examples of passing knowledge graphs as part of context to improve precision
• Using a hybrid version to use embedding vectors from LLMs and anchors in Ontologies to do precision matching
o We will focus on cases where we have to do precision matching in an Ontology around an anchor
Bio: Kuldeep Jiwani is Head of Data Science for HiLabs, a US Healthcare MNC. He has been driving research and innovation in the Healthcare sector using state of the art AI technologies like LLMs, Medical Ontologies, NLP, Predictive Analytics in multiple areas, Bayesian modeling, Statistical modeling, Time series forecasting, etc. Built 6 products in a year with a team of 50+ data scientists, where each product gathered multi-million dollars for the company.
Prior to this he was building machine learning applications at massive scale for the telecom sector. Discovering telecom subscribers behavioural patterns via mining and modelling billions of daily records, for various use cases like Churn prediction, Network congestion, Service experience, etc. He has been a Performance architect designing high scalable Big Data solutions over distributed systems. Then designing ultra-low latency trading solutions for the Financial trading tools industry. He has been a researcher all along, publishing papers and practically finding new ways to solve real world problems. He has also been an Entrepreneur and founding member of a startup that was successfully acquired by Oracle.
