Every day, an uncountable amount of documents are received and processed by companies worldwide. In an effort to reduce the cost of processing each document, the largest companies have resorted to document automation technologies. In an ideal world, a document can be automatically processed without any human intervention: its content is read, and information is extracted and forwarded to the relevant service. The state-of-the-art techniques have quickly evolved in the last decades, from rule-based algorithms to statistical models. This thesis focuses on machine learning models for document information extraction. Recent advances in model architecture for natural language processing have shown the importance of the attention mechanism. Transformers have revolutionized the field by generalizing the use of attention and by pushing self-supervised pre-training to the next level. In the first part, we confirm that transformers with appropriate pre-training were able to perform document understanding tasks with high performance. We show that, when used as a token classifier for information extraction, transformers are able to exceptionally efficiently learn the task compared to recurrent networks. Transformers only need a small proportion of the training data to reach close to maximum performance. This highlights the importance of self-supervised pre-training for future fine-tuning. In the following part, we design specialized pre-training tasks, to better prepare the model for specific data distributions such as business documents. By acknowledging the specificities of business documents such as their table structure and their over-representation of numeric figures, we are able to target specific skills useful for the model in its future tasks. We show that those new tasks improve the model's downstream performances, even with small models. Using this pre-training approach, we are able to reach the performances of significantly bigger models without any additional cost during finetuning or inference. Finally, in the last part, we address one drawback of the transformer architecture which is its computational cost when used on long sequences. We show that efficient architectures derived from the classic transformer require fewer resources and perform better on long sequences. However, due to how they approximate the attention computation, efficient models suffer from a small but significant performance drop on short sequences compared to classical architectures. This incentivizes the use of different models depending on the input length and enables concatenating multimodal inputs into a single sequence.