google-research/bert

BERT stands for Bidirectional Encoder Representations from Transformers. This repository is Google Research's official release of the TensorFlow code and pre-trained model weights for BERT, a natural language processing model that changed how machines understand text. The problem BERT solved is that earlier text models read sentences either left to right or right to left, missing the full context of a word in relation to everything around it. BERT reads the entire sentence in both directions simultaneously, giving it a much richer understanding of what each word means in context. The way it works is that BERT was pre-trained on a massive amount of text using two tasks: predicting randomly masked words in a sentence (which forces the model to understand context from both sides), and predicting whether one sentence logically follows another. After pre-training, BERT can be fine-tuned on a specific task, such as question answering, sentiment analysis, or text classification, by training it a bit more on a smaller labeled dataset for that task. This fine-tuning approach works remarkably well, letting a single large pre-trained model be adapted to many different language understanding tasks with relatively little additional data. This repository provides the pre-trained BERT-Base and BERT-Large models in both cased and uncased variants, as well as multilingual models, plus the code to fine-tune them on downstream tasks. You would use this repository if you are an NLP researcher or practitioner who wants to fine-tune BERT on your own text classification, question answering, or other language tasks, or if you want to study the original implementation. The tech stack is Python with TensorFlow.