README.md
| 1 | --- |
| 2 | language: en |
| 3 | tags: |
| 4 | - exbert |
| 5 | license: apache-2.0 |
| 6 | datasets: |
| 7 | - bookcorpus |
| 8 | - wikipedia |
| 9 | --- |
| 10 | |
| 11 | # BERT base model (cased) |
| 12 | |
| 13 | Pretrained model on English language using a masked language modeling (MLM) objective. It was introduced in |
| 14 | [this paper](https://arxiv.org/abs/1810.04805) and first released in |
| 15 | [this repository](https://github.com/google-research/bert). This model is case-sensitive: it makes a difference between |
| 16 | english and English. |
| 17 | |
| 18 | Disclaimer: The team releasing BERT did not write a model card for this model so this model card has been written by |
| 19 | the Hugging Face team. |
| 20 | |
| 21 | ## Model description |
| 22 | |
| 23 | BERT is a transformers model pretrained on a large corpus of English data in a self-supervised fashion. This means it |
| 24 | was pretrained on the raw texts only, with no humans labelling them in any way (which is why it can use lots of |
| 25 | publicly available data) with an automatic process to generate inputs and labels from those texts. More precisely, it |
| 26 | was pretrained with two objectives: |
| 27 | |
| 28 | - Masked language modeling (MLM): taking a sentence, the model randomly masks 15% of the words in the input then run |
| 29 | the entire masked sentence through the model and has to predict the masked words. This is different from traditional |
| 30 | recurrent neural networks (RNNs) that usually see the words one after the other, or from autoregressive models like |
| 31 | GPT which internally mask the future tokens. It allows the model to learn a bidirectional representation of the |
| 32 | sentence. |
| 33 | - Next sentence prediction (NSP): the models concatenates two masked sentences as inputs during pretraining. Sometimes |
| 34 | they correspond to sentences that were next to each other in the original text, sometimes not. The model then has to |
| 35 | predict if the two sentences were following each other or not. |
| 36 | |
| 37 | This way, the model learns an inner representation of the English language that can then be used to extract features |
| 38 | useful for downstream tasks: if you have a dataset of labeled sentences for instance, you can train a standard |
| 39 | classifier using the features produced by the BERT model as inputs. |
| 40 | |
| 41 | ## Intended uses & limitations |
| 42 | |
| 43 | You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to |
| 44 | be fine-tuned on a downstream task. See the [model hub](https://huggingface.co/models?filter=bert) to look for |
| 45 | fine-tuned versions on a task that interests you. |
| 46 | |
| 47 | Note that this model is primarily aimed at being fine-tuned on tasks that use the whole sentence (potentially masked) |
| 48 | to make decisions, such as sequence classification, token classification or question answering. For tasks such as text |
| 49 | generation you should look at model like GPT2. |
| 50 | |
| 51 | ### How to use |
| 52 | |
| 53 | You can use this model directly with a pipeline for masked language modeling: |
| 54 | |
| 55 | ```python |
| 56 | >>> from transformers import pipeline |
| 57 | >>> unmasker = pipeline('fill-mask', model='bert-base-cased') |
| 58 | >>> unmasker("Hello I'm a [MASK] model.") |
| 59 | |
| 60 | [{'sequence': "[CLS] Hello I'm a fashion model. [SEP]", |
| 61 | 'score': 0.09019174426794052, |
| 62 | 'token': 4633, |
| 63 | 'token_str': 'fashion'}, |
| 64 | {'sequence': "[CLS] Hello I'm a new model. [SEP]", |
| 65 | 'score': 0.06349995732307434, |
| 66 | 'token': 1207, |
| 67 | 'token_str': 'new'}, |
| 68 | {'sequence': "[CLS] Hello I'm a male model. [SEP]", |
| 69 | 'score': 0.06228214129805565, |
| 70 | 'token': 2581, |
| 71 | 'token_str': 'male'}, |
| 72 | {'sequence': "[CLS] Hello I'm a professional model. [SEP]", |
| 73 | 'score': 0.0441727414727211, |
| 74 | 'token': 1848, |
| 75 | 'token_str': 'professional'}, |
| 76 | {'sequence': "[CLS] Hello I'm a super model. [SEP]", |
| 77 | 'score': 0.03326151892542839, |
| 78 | 'token': 7688, |
| 79 | 'token_str': 'super'}] |
| 80 | ``` |
| 81 | |
| 82 | Here is how to use this model to get the features of a given text in PyTorch: |
| 83 | |
| 84 | ```python |
| 85 | from transformers import BertTokenizer, BertModel |
| 86 | tokenizer = BertTokenizer.from_pretrained('bert-base-cased') |
| 87 | model = BertModel.from_pretrained("bert-base-cased") |
| 88 | text = "Replace me by any text you'd like." |
| 89 | encoded_input = tokenizer(text, return_tensors='pt') |
| 90 | output = model(**encoded_input) |
| 91 | ``` |
| 92 | |
| 93 | and in TensorFlow: |
| 94 | |
| 95 | ```python |
| 96 | from transformers import BertTokenizer, TFBertModel |
| 97 | tokenizer = BertTokenizer.from_pretrained('bert-base-cased') |
| 98 | model = TFBertModel.from_pretrained("bert-base-cased") |
| 99 | text = "Replace me by any text you'd like." |
| 100 | encoded_input = tokenizer(text, return_tensors='tf') |
| 101 | output = model(encoded_input) |
| 102 | ``` |
| 103 | |
| 104 | ### Limitations and bias |
| 105 | |
| 106 | Even if the training data used for this model could be characterized as fairly neutral, this model can have biased |
| 107 | predictions: |
| 108 | |
| 109 | ```python |
| 110 | >>> from transformers import pipeline |
| 111 | >>> unmasker = pipeline('fill-mask', model='bert-base-cased') |
| 112 | >>> unmasker("The man worked as a [MASK].") |
| 113 | |
| 114 | [{'sequence': '[CLS] The man worked as a lawyer. [SEP]', |
| 115 | 'score': 0.04804691672325134, |
| 116 | 'token': 4545, |
| 117 | 'token_str': 'lawyer'}, |
| 118 | {'sequence': '[CLS] The man worked as a waiter. [SEP]', |
| 119 | 'score': 0.037494491785764694, |
| 120 | 'token': 17989, |
| 121 | 'token_str': 'waiter'}, |
| 122 | {'sequence': '[CLS] The man worked as a cop. [SEP]', |
| 123 | 'score': 0.035512614995241165, |
| 124 | 'token': 9947, |
| 125 | 'token_str': 'cop'}, |
| 126 | {'sequence': '[CLS] The man worked as a detective. [SEP]', |
| 127 | 'score': 0.031271643936634064, |
| 128 | 'token': 9140, |
| 129 | 'token_str': 'detective'}, |
| 130 | {'sequence': '[CLS] The man worked as a doctor. [SEP]', |
| 131 | 'score': 0.027423162013292313, |
| 132 | 'token': 3995, |
| 133 | 'token_str': 'doctor'}] |
| 134 | |
| 135 | >>> unmasker("The woman worked as a [MASK].") |
| 136 | |
| 137 | [{'sequence': '[CLS] The woman worked as a nurse. [SEP]', |
| 138 | 'score': 0.16927455365657806, |
| 139 | 'token': 7439, |
| 140 | 'token_str': 'nurse'}, |
| 141 | {'sequence': '[CLS] The woman worked as a waitress. [SEP]', |
| 142 | 'score': 0.1501094549894333, |
| 143 | 'token': 15098, |
| 144 | 'token_str': 'waitress'}, |
| 145 | {'sequence': '[CLS] The woman worked as a maid. [SEP]', |
| 146 | 'score': 0.05600163713097572, |
| 147 | 'token': 13487, |
| 148 | 'token_str': 'maid'}, |
| 149 | {'sequence': '[CLS] The woman worked as a housekeeper. [SEP]', |
| 150 | 'score': 0.04838843643665314, |
| 151 | 'token': 26458, |
| 152 | 'token_str': 'housekeeper'}, |
| 153 | {'sequence': '[CLS] The woman worked as a cook. [SEP]', |
| 154 | 'score': 0.029980547726154327, |
| 155 | 'token': 9834, |
| 156 | 'token_str': 'cook'}] |
| 157 | ``` |
| 158 | |
| 159 | This bias will also affect all fine-tuned versions of this model. |
| 160 | |
| 161 | ## Training data |
| 162 | |
| 163 | The BERT model was pretrained on [BookCorpus](https://yknzhu.wixsite.com/mbweb), a dataset consisting of 11,038 |
| 164 | unpublished books and [English Wikipedia](https://en.wikipedia.org/wiki/English_Wikipedia) (excluding lists, tables and |
| 165 | headers). |
| 166 | |
| 167 | ## Training procedure |
| 168 | |
| 169 | ### Preprocessing |
| 170 | |
| 171 | The texts are tokenized using WordPiece and a vocabulary size of 30,000. The inputs of the model are then of the form: |
| 172 | |
| 173 | ``` |
| 174 | [CLS] Sentence A [SEP] Sentence B [SEP] |
| 175 | ``` |
| 176 | |
| 177 | With probability 0.5, sentence A and sentence B correspond to two consecutive sentences in the original corpus and in |
| 178 | the other cases, it's another random sentence in the corpus. Note that what is considered a sentence here is a |
| 179 | consecutive span of text usually longer than a single sentence. The only constrain is that the result with the two |
| 180 | "sentences" has a combined length of less than 512 tokens. |
| 181 | |
| 182 | The details of the masking procedure for each sentence are the following: |
| 183 | - 15% of the tokens are masked. |
| 184 | - In 80% of the cases, the masked tokens are replaced by `[MASK]`. |
| 185 | - In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace. |
| 186 | - In the 10% remaining cases, the masked tokens are left as is. |
| 187 | |
| 188 | ### Pretraining |
| 189 | |
| 190 | The model was trained on 4 cloud TPUs in Pod configuration (16 TPU chips total) for one million steps with a batch size |
| 191 | of 256. The sequence length was limited to 128 tokens for 90% of the steps and 512 for the remaining 10%. The optimizer |
| 192 | used is Adam with a learning rate of 1e-4, \\(\beta_{1} = 0.9\\) and \\(\beta_{2} = 0.999\\), a weight decay of 0.01, |
| 193 | learning rate warmup for 10,000 steps and linear decay of the learning rate after. |
| 194 | |
| 195 | ## Evaluation results |
| 196 | |
| 197 | When fine-tuned on downstream tasks, this model achieves the following results: |
| 198 | |
| 199 | Glue test results: |
| 200 | |
| 201 | | Task | MNLI-(m/mm) | QQP | QNLI | SST-2 | CoLA | STS-B | MRPC | RTE | Average | |
| 202 | |:----:|:-----------:|:----:|:----:|:-----:|:----:|:-----:|:----:|:----:|:-------:| |
| 203 | | | 84.6/83.4 | 71.2 | 90.5 | 93.5 | 52.1 | 85.8 | 88.9 | 66.4 | 79.6 | |
| 204 | |
| 205 | |
| 206 | ### BibTeX entry and citation info |
| 207 | |
| 208 | ```bibtex |
| 209 | @article{DBLP:journals/corr/abs-1810-04805, |
| 210 | author = {Jacob Devlin and |
| 211 | Ming{-}Wei Chang and |
| 212 | Kenton Lee and |
| 213 | Kristina Toutanova}, |
| 214 | title = {{BERT:} Pre-training of Deep Bidirectional Transformers for Language |
| 215 | Understanding}, |
| 216 | journal = {CoRR}, |
| 217 | volume = {abs/1810.04805}, |
| 218 | year = {2018}, |
| 219 | url = {http://arxiv.org/abs/1810.04805}, |
| 220 | archivePrefix = {arXiv}, |
| 221 | eprint = {1810.04805}, |
| 222 | timestamp = {Tue, 30 Oct 2018 20:39:56 +0100}, |
| 223 | biburl = {https://dblp.org/rec/journals/corr/abs-1810-04805.bib}, |
| 224 | bibsource = {dblp computer science bibliography, https://dblp.org} |
| 225 | } |
| 226 | ``` |
| 227 | |
| 228 | <a href="https://huggingface.co/exbert/?model=bert-base-cased"> |
| 229 | <img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png"> |
| 230 | </a> |
| 231 | |