Thanks for tuning in to Google I/O. View all sessions on demandWatch on demand

Text and natural language processing with TensorFlow

Before you can train a model on text data, you'll typically need to process (or preprocess) the text. In many cases, text needs to be tokenized and vectorized before it can be fed to a model, and in some cases the text requires additional preprocessing steps such as normalization and feature selection.

After text is processed into a suitable format, you can use it in natural language processing (NLP) workflows such as text classification, text generation, summarization, and translation.

TensorFlow provides two libraries for text and natural language processing: KerasNLP (GitHub) and TensorFlow Text (GitHub).

KerasNLP is a high-level NLP modeling library that includes all the latest transformer-based models as well as lower-level tokenization utilities. It's the recommended solution for most NLP use cases. Built on TensorFlow Text, KerasNLP abstracts low-level text processing operations into an API that's designed for ease of use. But if you prefer not to work with the Keras API, or you need access to the lower-level text processing ops, you can use TensorFlow Text directly.

Keras NLP

The easiest way to get started processing text in TensorFlow is to use KerasNLP. KerasNLP is a natural language processing library that supports workflows built from modular components that have state-of-the-art preset weights and architectures. You can use KerasNLP components with their out-of-the-box configuration. If you need more control, you can easily customize components. KerasNLP provides in-graph computation for all workflows so you can expect easy productionization using the TensorFlow ecosystem.

KerasNLP contains end-to-end implementations of popular model architectures like BERT and FNet. Using KerasNLP models, layers, and tokenizers, you can complete many state-of-the-art NLP workflows, including machine translation, text generation, text classification, and transformer model training.

KerasNLP is an extension of the core Keras API, and every high-level KerasNLP module is a Layer or Model. If you're familiar with Keras, you already understand most of KerasNLP.

TensorFlow Text

KerasNLP provides high-level text processing modules that are available as layers or models. If you need access to lower-level tools, you can use TensorFlow Text. TensorFlow Text provides operations and libraries to help you work with raw text strings and documents. TensorFlow Text can perform the preprocessing regularly required by text-based models, and it also includes other features useful for sequence modeling.

Using TensorFlow Text, you can do the following:

  • Apply feature-rich tokenizers that can split strings on whitespace, separate words and punctuation, and return byte offsets with tokens, so that you know where a string can be found in the source text.
  • Check if a token matches a specified string pattern. You can check for capitalization, punctuation, numerical data, and other token features.
  • Combine tokens into n-grams.
  • Process text within the TensorFlow graph, so that tokenization during training matches tokenization at inference.

Where to start

The following resources will help you get started with TensorFlow text processing: