WWDC Notes

Advances in Natural Language Framework

Show Apple's description.
Natural Language is a framework designed to provide high-performance, on-device APIs for natural language processing tasks across all Apple platforms. Learn about the addition of Sentiment Analysis and Text Catalog support in the framework. Gain a deeper understanding of transfer learning for text-based models and the new support for Word Embeddings which can power great search experiences in your app.

Two UX concerns when thinking about natural language :

  • output of natural language text
  • User input of text (ie. search input)

iOS 12 Natural language added text classification and entity tagging

From iOS 13 there’s a new category of Sentiment analysis:

  • The output is value from -1 to 1 represent a confidence score
  • This analysis is extremely performant (it uses a hardware-accelerated neural network).
  • The sentiment analysis is around the positivity or negativity of the sentence
  • Support for seven languages

Word Tagging

Definition: given a sequence of words (a sentence), aka a sequence of tokens in NPL, we would like to assign a label to every token in this sequence.

  • in iOS 13, Word Tagging has gotten improved support with ‘Catalogs’.
  • New this year we can pass our own dictionary (text catalogs) in word tagging to make NPL even more powerful.
  • This is done via MLGazeteer: MLGazetteer and its associated APIs let you create and “Text Catalogs”.
  • You can leverage tons of data and create a model that’s super compressed. Ex. All the named places and persons in Wikipedia (2.5 million entries) take up on 2.5 MB on disk!
  • It’s like language classification models similar to image classification in the Vision framework but for language
  • Our gazetteer has priority over the default (iOS’s) one. Usage:

Word Embedding

Definition

Vector representation of words.

Basically this is the same concepts behind the Image Similarity above, similar words (sneakers and flip flops for example) will be clustered together.

Mapping of a string in continuous sequence of numbers. The vectors represent a score of association. E.g. thunderstorm -> sky, clouds, cloudy

Word Embedding in iOS 13

We can also have our own Word Embedding: maybe for a particular domain or for a language not supported by default.

This also supports other standards (so we don’t need to create our own) as:

  • Word2vec
  • GloVe
  • fasttext
  • Apple uses this technique combined with Image classification to produce search results in the new iOS 13 photos app: if we search for thunderstorm the photos app will suggest results for sky too
  • API supports the querying the ‘distance’ between two words, nearest neighbors etc.
  • The default models supports 7 languages. expect more in future iOS releases.
  • Custom Word Embedding can be more domain specific and more sophisticated knowledge

Transfer Learning for Text Classification

  • This allows for smaller sets of data for training and improved accuracy For example, when your training data doesn’t include the words used by the user.
  • Transfer of learning notices the meaning of words and the context in which they are used. Using this knowledge, ML infers the meaning for text classification
  • Limited language support (static vs dynamic)

Missing anything? Corrections? Contributions are welcome 😃

Related

Written by

zntfdr

Federico Zanetello

Federico Zanetello is an iOS Engineer with strong passion for Swift, minimalism, and design. When he’s not busy automating things, he can be found writing at fivestars.blog and/or playing with the latest shiny toys.