Invited Talk by Dr. Milad Alshomary on Tuesday, August 26, at 16:00

Title

Unlocking Authorial Style: Enhancing robustness and interpretability of embedding-based authorship attribution models

Abstract

Current embedding-based models excel at authorship attribution, but they often are black boxes, making it hard to trust their decisions, especially in sensitive applications. Furthermore, they struggle to identify authors when the writing style is different from what they were trained on (out-of-domain data). In this talk, I will first present an approach to interpret the latent space of these black boxes through identifying salient subspaces and characterizing their specific writing style. Then, I will discuss how to improve the effectiveness of these models when tested in out-of-domain scenarios through leveraging different levels of linguistic granularities.