Invited presentation at IRIT in Frebruary 2025 (part of Sébastien Mosser’s invited prof visit)
Title: Justifying Machine Learning Pipelines: Applications to Natural Language Processing and Agility
In modern software engineering, user requirements are often captured as user stories, as an easy way to model (personal, action, benefits) relations. As stories are stored using plain text in backlogs, one might consider using machine learning (ML) / natural language processing (NLP) pipelines to extract the captured concepts from a textual backlog. In the era of ChatGPT and LLMs, implementing such a task is trivial. But, surprisingly, it is also extremely easy to do things in the most wrong possible way. The talk will present an NLP experiment to extract conceptual models from a reference user stories backlog and then show how safety-related techniques such as justification modelling (using the jPipe language) can help software engineers improve their ML pipelines.