Emergent Mind


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).
Comparison of BERT, OpenAI GPT, and ELMo models highlighting architecture and context processing differences.


  • BERT introduces a novel technique for pre-training deep bidirectional transformers on extensive unlabeled text to enhance language understanding.

  • Utilizes a 'masked language model' pre-training objective, where it predicts randomly masked words based on their context.

  • Incorporates a 'next sentence prediction' task during pre-training to grasp the connections between sentences, improving performance.

  • Allows fine-tuning for specific language tasks with minimal architectural changes, leading to new benchmarks in language processing.

  • BERT's architecture manages to capture rich contextual information, making significant strides in natural language processing.

Introduction to BERT

BERT, which stands for Bidirectional Encoder Representations from Transformers, represents a significant leap in language processing capabilities. As opposed to previous models that primarily focused on unidirectional text understanding or used complex task-specific architectures, BERT excels by pre-training on unlabeled text and jointly conditioning on both left and right contexts.

Pre-training of BERT

BERT is pre-trained on a vast corpus of text which includes the BooksCorpus with 800 million words and English Wikipedia with 2,500 million words. Unlike conventional language models, it uses a "masked language model" (MLM) pre-training objective, inspired by the Cloze task. This means that during training, random words are masked and the model learns to predict the masked word based solely on its context.

Furthermore, BERT introduces a "next sentence prediction" task during pre-training that enables the model to understand the relationships between sentences. Sentence pairs are classified as consecutive or not, with this binary task pre-training BERT on the relationship between subsequent sentences – a critical factor in understanding language.

Fine-tuning BERT for Various Tasks

After pre-training, BERT can be fine-tuned with additional output layers for a wide range of language understanding tasks, from question answering to sentiment analysis. This fine-tuning process adjusts the pre-trained parameters to be more task-specific without the need for extensive architecture modifications. The fine-tuning usually involves performing training again but with much fewer data and iterations.

Benchmark Achievements

BERT has set new records across eleven natural language processing tasks. It achieved significant improvements in the GLUE score and various question answering benchmarks, showcasing the model's state-of-the-art performance.

BERT's Contribution and Impact

BERT's major contribution lies in its approach to handling bidirectional contextual information and its fine-tuning efficiency. It demonstrates that even modest-sized models, when pre-trained on a sufficiently diverse and expansive corpus, can provide substantial benefits across a variety of tasks. BERT's architecture and training approach enable it to develop rich and intricately nuanced language representations that drive the advancements in machine understanding of natural language.

Get summaries of Trending computer science papers delivered straight to your inbox

Unsubscribe anytime.