Building iOS Question Answering App with BERT

Discover how words become numbers through word embeddings, turning text into vectors for machine processing and deep learning.

Explore how one hot encoding converts words into binary vectors using a vocabulary-sized dimension, illustrated with a five-word corpus and sparse representations in large vocabularies.

Explore what a deep neural network is and how multiple layers enable learning from data, with forward and backpropagation guiding activation functions and predictions.

Learn how words are represented as vectors through word embeddings, using continuous bag of words and skip-gram models to predict context words with neural networks.

Explore glove embeddings that derive dense word vectors from co-occurrence data, enabling word relations and analogies, and enabling sentiment analysis, movie reviews, and news classification.

Explore how BERT enables transfer learning by using pre-trained embeddings in a bidirectional transformer with encoder–decoder self-attention and masked language modeling.

Build an iOS question answering app from scratch by creating the UI, implementing behind-the-scenes code, integrating the ML model, and enabling speech to text and text to speech.

Create a single view iOS app in Xcode, organize code into model, view, and controller, and build a UI with labels, text views, and speech-to-text to find and highlight answers.

Explore how a BERT-based model, trained on SQuAD data from Wikipedia, uses tokenization and start-end offsets to answer questions from a document in an iOS app.

Explore how a BERT vocabulary dictionary of word tokens is loaded and used for tokenization and encoding in a Swift iOS app, including bundle access and token lookups.

Create a tokenized string class that converts sentences to word tokens using natural language processing library, handles unknown words with subword tokens prefixed by double hash, and returns token ids.

Demonstrate tokenization logic in a BERT-like tokenizer, converting word tokens to token ids, handling unknown words by splitting into subwords with a '##' prefix, and assembling the final token sequence.

Learn to build BERT input for an iOS question-answering app by assembling CLS, question, separator, and document tokens, with token type indicators and token limits.

Explore building a Swift board output class to process BERT model outputs, compute top indices from logits, and connect predictions to a view controller in iOS.

Learn how a BERT facade wraps a model to find answers by processing a question through a document and extracting the answer.

Explore converting speech to text in an iOS question answering app by requesting microphone access, enabling a safe speech recognizer, and managing transcription with an audio session.

Convert the document answer into spoken audio in an iOS question answering app, using a text-to-speech converter and the iOS speech synthesis library to read the answer aloud.

Learn to find an answer to a user question in the UI by wiring a controller action, running a background search with Berglas, and highlighting the answer in the document.

Learn to build the iOS app, view build progress and errors, fix issues, and run on the iPhone simulator or a connected real device.

Demonstrates a mobile iOS question answering app that uses on-device machine learning to answer questions from a document, via speech input, highlighting, and spoken results.

Extend your bert-based iOS question-answering app by testing larger passages (up to 384 tokens), exploring multilingual support (Spanish, Japanese, Italian, Hebrew), and preparing for App Store release with user feedback.