Transformer-based models have pushed state of the art in many areas of NLP, but our understanding of what is behind their success is still limited. This paper is the first survey of over 150 studies of the popular BERT model. We review the current state of knowledge about how BERT works, what kind of information it learns and how it is represented, common modifications to its training objectives and architecture, the overparameterization issue, and approaches to compression. We then outline directions for future research.

BERT-based architectures currently give stateof-the-art performance on many NLP tasks, but little is known about the exact mechanisms that contribute to its success. In the current work, we focus on the interpretation of selfattention, which is one of the fundamental underlying components of BERT. Using a subset of GLUE tasks and a set of handcrafted features-of-interest, we propose the methodology and carry out a qualitative and quantitative analysis of the information encoded by the individual BERT's heads. Our findings suggest that there is a limited set of attention patterns that are repeated across different heads, indicating the overall model overparametrization. While different heads consistently use the same attention patterns, they have varying impact on performance across different tasks. We show that manually disabling attention in certain heads leads to a performance improvement over the regular fine-tuned BERT models.

The recent explosion in question answering research produced a wealth of both factoid reading comprehension (RC) and commonsense reasoning datasets. Combining them presents a different kind of task: deciding not simply whether information is present in the text, but also whether a confident guess could be made for the missing information. We present QuAIL, the first RC dataset to combine text-based, world knowledge and unanswerable questions, and to provide question type annotation that would enable diagnostics of the reasoning strategies by a given QA system. QuAIL contains 15K multi-choice questions for 800 texts in 4 domains. Crucially, it offers both general and text-specific questions, unlikely to be found in pretraining data. We show that QuAIL poses substantial challenges to the current state-of-the-art systems, with a 30% drop in accuracy compared to the most similar existing dataset.

Multiple studies have shown that Transformers are remarkably robust to pruning. Contrary to this received wisdom, we demonstrate that pre-trained Transformer encoders are surprisingly fragile to the removal of a very small number of features in the layer outputs (<0.0001% of model weights). In case of BERT and other pre-trained encoder Transformers, the affected component is the scaling factors and biases in the LayerNorm. The outliers are high-magnitude normalization parameters that emerge early in pre-training and show up consistently in the same dimensional position throughout the model. We show that disabling them significantly degrades both the MLM loss and the downstream task performance. This effect is observed across several BERT-family models and other popular pre-trained Transformer architectures, including BART, XLNet and ELECTRA; we also show a similar effect in GPT-2.

Current research in machine learning for radiology is focused mostly on images. There exists limited work in investigating intelligent interactive systems for radiology. To address this limitation, we introduce a realistic and information-rich task of Visual Dialog in radiology, specific to chest X-ray images. Using MIMIC-CXR, an openly available database of chest X-ray images, we construct both a synthetic and a real-world dataset and provide baseline scores achieved by state-of-theart models. We show that incorporating medical history of the patient leads to better performance in answering questions as opposed to conventional visual question answering model which looks only at the image. While our experiments show promising results, they indicate that the task is extremely challenging with significant scope for improvement. We make both the datasets (synthetic and gold standard) and the associated code publicly available to the research community.