- admin
- 0 Comments
- 606 Views
Abstract:
The advent of powerful natural language processing (NLP) models has recently captured the imagination of the general public. In the legal NLP domain, models have exhibited increasingly strong performance on various benchmarks. However, there is surprisingly little legal expertise employed in solving these tasks. Instead of considering law as just another machine learning dataset, this talk will present two directions in which we can bring Law and NLP meaningfully together.
In the first part of this talk, we will consider how neural models allow us to learn more about the law itself. Specifically, we will take a lesson from information theory to shine some light on legal precedent. In other words, we will bring NLP to Law.
In the second part of the talk, we will consider how law can help us design better NLP models. Specifically, we will look at the outcome prediction task – the premier task in legal NLP – which is non-sensical from the legal perspective. In response, we will use legal domain knowledge to design more powerful and legally plausible neural models. In other words, we will bring Law to NLP.
About our speaker:
Josef Valvoda is a final-year PhD student at the University of Cambridge supervised by Professor Simone Teufel (Cambridge) and Professor Ryan Cotterell (ETH). Before joining the Natural Language and Information Processing group he completed the MPhil in Advanced Computer Science also at the University of Cambridge. Before that, he obtained a Bachelor of Law at the University of Exeter.
His PhD work focuses on Artificial Intelligence and Law. He thinks about what deep learning models can learn, have learned and should learn with respect to law. While Legal NLP is the focus of his thesis, he is also very much interested in broader NLP research. Among his other work is the recently published paper on testing the compositional behaviour of neural networks or developing (probing) methods that could better our understanding of what the neural models learn.
