1# BERT Question and Answer 2 3Use a TensorFlow Lite model to answer questions based on the content of a given 4passage. 5 6Note: (1) To integrate an existing model, try 7[TensorFlow Lite Task Library](https://www.tensorflow.org/lite/inference_with_metadata/task_library/bert_question_answerer). 8(2) To customize a model, try 9[TensorFlow Lite Model Maker](https://www.tensorflow.org/lite/models/modify/model_maker/question_answer). 10 11## Get started 12 13<img src="images/screenshot.gif" class="attempt-right" style="max-width: 300px"> 14 15If you are new to TensorFlow Lite and are working with Android or iOS, we 16recommend exploring the following example applications that can help you get 17started. 18 19<a class="button button-primary" href="https://github.com/tensorflow/examples/tree/master/lite/examples/bert_qa/android">Android 20example</a> 21<a class="button button-primary" href="https://github.com/tensorflow/examples/tree/master/lite/examples/bert_qa/ios">iOS 22example</a> 23 24If you are using a platform other than Android/iOS, or you are already familiar 25with the 26[TensorFlow Lite APIs](https://www.tensorflow.org/api_docs/python/tf/lite), you 27can download our starter question and answer model. 28 29<a class="button button-primary" href="https://tfhub.dev/tensorflow/lite-model/mobilebert/1/metadata/1?lite-format=tflite">Download 30starter model and vocab</a> 31 32For more information about metadata and associated fields (e.g. `vocab.txt`) see 33<a href="https://www.tensorflow.org/lite/models/convert/metadata#read_the_metadata_from_models">Read 34the metadata from models</a>. 35 36## How it works 37 38The model can be used to build a system that can answer users’ questions in 39natural language. It was created using a pre-trained BERT model fine-tuned on 40SQuAD 1.1 dataset. 41 42[BERT](https://github.com/google-research/bert), or Bidirectional Encoder 43Representations from Transformers, is a method of pre-training language 44representations which obtains state-of-the-art results on a wide array of 45Natural Language Processing tasks. 46 47This app uses a compressed version of BERT, MobileBERT, that runs 4x faster and 48has 4x smaller model size. 49 50[SQuAD](https://rajpurkar.github.io/SQuAD-explorer/), or Stanford Question 51Answering Dataset, is a reading comprehension dataset consisting of articles 52from Wikipedia and a set of question-answer pairs for each article. 53 54The model takes a passage and a question as input, then returns a segment of the 55passage that most likely answers the question. It requires semi-complex 56pre-processing including tokenization and post-processing steps that are 57described in the BERT [paper](https://arxiv.org/abs/1810.04805) and implemented 58in the sample app. 59 60## Performance benchmarks 61 62Performance benchmark numbers are generated with the tool 63[described here](https://www.tensorflow.org/lite/performance/benchmarks). 64 65<table> 66 <thead> 67 <tr> 68 <th>Model Name</th> 69 <th>Model size </th> 70 <th>Device </th> 71 <th>CPU</th> 72 </tr> 73 </thead> 74 <tr> 75 <td rowspan = 3> 76 <a href="https://tfhub.dev/tensorflow/lite-model/mobilebert/1/metadata/1?lite-format=tflite">Mobile Bert</a> 77 </td> 78 <td rowspan = 3> 79 100.5 Mb 80 </td> 81 <td>Pixel 3 (Android 10) </td> 82 <td>123ms*</td> 83 </tr> 84 <tr> 85 <td>Pixel 4 (Android 10) </td> 86 <td>74ms*</td> 87 </tr> 88 <tr> 89 <td>iPhone XS (iOS 12.4.1) </td> 90 <td>257ms** </td> 91 </tr> 92</table> 93 94\* 4 threads used. 95 96\*\* 2 threads used on iPhone for the best performance result. 97 98## Example output 99 100### Passage (Input) 101 102> Google LLC is an American multinational technology company that specializes in 103> Internet-related services and products, which include online advertising 104> technologies, search engine, cloud computing, software, and hardware. It is 105> considered one of the Big Four technology companies, alongside Amazon, Apple, 106> and Facebook. 107> 108> Google was founded in September 1998 by Larry Page and Sergey Brin while they 109> were Ph.D. students at Stanford University in California. Together they own 110> about 14 percent of its shares and control 56 percent of the stockholder 111> voting power through supervoting stock. They incorporated Google as a 112> California privately held company on September 4, 1998, in California. Google 113> was then reincorporated in Delaware on October 22, 2002. An initial public 114> offering (IPO) took place on August 19, 2004, and Google moved to its 115> headquarters in Mountain View, California, nicknamed the Googleplex. In August 116> 2015, Google announced plans to reorganize its various interests as a 117> conglomerate called Alphabet Inc. Google is Alphabet's leading subsidiary and 118> will continue to be the umbrella company for Alphabet's Internet interests. 119> Sundar Pichai was appointed CEO of Google, replacing Larry Page who became the 120> CEO of Alphabet. 121 122### Question (Input) 123 124> Who is the CEO of Google? 125 126### Answer (Output) 127 128> Sundar Pichai 129 130## Read more about BERT 131 132* Academic paper: [BERT: Pre-training of Deep Bidirectional Transformers for 133 Language Understanding](https://arxiv.org/abs/1810.04805) 134* [Open-source implementation of BERT](https://github.com/google-research/bert) 135
