Talk: When Neural Networks Meet the Law

Title

When Neural Networks Meet the Law

 

Abstract

Legal outcome prediction models, the hallmark of legal NLP, have a problem – they ignore half of the law there is. This talk will look at how neural models of law have deviated from the reality of the domain they are meant to model, how domain knowledge can help us with designing better models of law and what lessons this teaches us about aligning AI with law.

 

Bio

Josef Valvoda is a Postdoctoral Researcher, working jointly at DIKU and JUR.

He holds a PhD in Computer Science from the University of Cambridge, where he was co-advised by Simone Teufel (UoC) and Ryan Cotterell (ETH). Before that, he obtained a Bachelor of Law at the University of Exeter.

Josef’s PhD focused on Artificial Intelligence and Law. He is concerned with how automation can lead to widening access to justice and what deep learning models can, and should, learn. He is on the program committee of the Competition on Legal Information Extraction/Entailment (COLIEE), Natural Legal Language Processing Workshop (NLLP) and the International Workshop on Juris-informatics (JURISIN).

While Legal NLP is the focus of Josef’s thesis, he is also very much interested in broader NLP research. He has published work on testing the compositional behaviour of neural networks, modelling ethics using LLMs, and participates in developing (probing) methods that could better our understanding of what the neural models learn.

Outside of his work in Cambridge, Josef has been frequenting Japan’s National Institute of Informatics (NII), worked with startups (TechWolf), and interned as a researcher at Apple and Amazon.