BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement).

Three Important Things

1. Masked Language Modeling

A challenge of training the deep bidirectional model is that it cannot be done either left-to-right or right-to-left, since the bidirectional nature of the Transformer will allow it to see the tokens in advance.

To resolve this, the authors introduced “masked LM” (MLM), where tokens are masked at random and the task is to predict the masked tokens. This is the first task used in their pre-training phase, the other being next sentence prediction (NSP).

2. Next Sentence Prediction

To ensure that the language model understands the relationships between sentences, it is also pre-trained on next sentence prediction tasks.

3. Robust on Both Fine-tuning and Feature Extraction

Ablation studies shows that BERT performs competitively with both fine-tuning on downstream tasks, and on using frozen feature extractions with a biLSTM layer on top.

Most Glaring Deficiency

In the setup for next sentence prediction, it was unclear to me how the embeddings for IsNext and NotNext were chosen and what the loss function is such that \(C\) has some useful representation.

This is particularly pertinent because of the importance of \(C\) for downstream tasks.

Conclusions for Future Work

To take advantage of bi-directionality in sequential data, we can apply ideas similar to masked language modeling even for data of other modalities.