During the past decade, neural networks have become prominent in Natural Language Processing (NLP), notably for their capacity to learn relevant word representations from large unlabeled corpora. These word embeddings can then be transferred and finetuned for diverse end applications during a supervised training phase. More recently, in 2018, the transfer of entire pretrained Language Models and the preservation of their contextualization capacities enabled to reach unprecedented performance on virtually every NLP benchmark, sometimes even outperforming human baselines. However, as models reach such impressive scores, their comprehension abilities still appear as shallow, which reveal limitations of benchmarks to provide useful insights on their factors of performance and to accurately measure understanding capabilities. In this thesis, we study the behaviour of state-of-the-art models regarding generalization to facts unseen during training in two important Information Extraction tasks: Named Entity Recognition (NER) and Relation Extraction (RE). Indeed, traditional benchmarks present important lexical overlap between mentions and relations used for training and evaluating models, whereas the main interest of Information Extraction is to extract previously unknown information. We propose empirical studies to separate performance based on mention and relation overlap with the training set and find that pretrained Language Models are mainly beneficial to detect unseen mentions, in particular out-of-domain. While this makes them suited for real use cases, there is still a gap in performance between seen and unseen mentions that hurts generalization to new facts. In particular, even state-of-the-art ERE models rely on a shallow retention heuristic, basing their prediction more on arguments surface forms than context.