Invited Talk
Table Host
Chair: Mirco Giacobbe
From ε-Balls to Semantics: Towards Verifiable NLP Pipelines
Abstract
As language models enter safety-critical domains, verifying their behaviour is vital. This talk introduces a principled approach to NLP verification that bridges the gap between semantic meaning and geometric embeddings. We propose new metrics for generalisability and falsifiability, and show how semantically-informed training and verification can provide stronger, more meaningful guarantees. Our tool ANTONIO demonstrates how these methods can be applied in practice to safeguard LLMs.