Andrea Santilli, Universita’ degli Studi di Roma Tor Vergata,
Continual Language Learning with Syntax-based Episodic Memory in Neural Networks.
Board judgment: The thesis enhances a state-of-the-art transformers, BERT, with a Syntax-based Episodic Memory module. The new module improves BERT performance both in a single task and in a continual learning scenario. Moreover, the syntax-based model mitigates catastrophic forgetting. The methodology is sound as well as the experiment set up whose results are further enriched with an in depth discussion. The thesis is an important contribution to the field of Computational Linguistics and has already received recognition at international level: part of it has been published at EMNLP 2021. We unanimously consider it an excellent thesis and are happy to assign the “Emanuele Pianta Award for the Best Master Thesis” to Andrea Santilli, Universita’ degli Studi di Roma, Tor Vergata.