IMPORTANT: When installing TF Text with pip install, please note the version of TensorFlow you are running, as you should specify the corresponding minor version of TF Text (eg. for tensorflow==2.3.x use tensorflow_text==2.3.x).

• Introduction
• Documentation
• Unicode
• Normalization
• Tokenization
  • Whitespace Tokenizer
  • UnicodeScript Tokenizer
  • Unicode split
  • Offsets
  • TF.Data Example
  • Keras API
• Other Text Ops
  • Wordshape
  • N-grams & Sliding Window
• Installation
  • Install using PIP
  • Build from source steps:

TensorFlow Text provides a collection of text related classes and ops ready to use with TensorFlow 2.0. The library can perform the preprocessing regularly required by text-based models, and includes other features useful for sequence modeling not provided by core TensorFlow.

The benefit of using these ops in your text preprocessing is that they are done in the TensorFlow graph. You do not need to worry about tokenization in training being different than the tokenization at inference, or managing preprocessing scripts.

Please visit http://tensorflow.org/text for all documentation. This site includes API docs, guides for working with TensorFlow Text, as well as tutorials for building specific models.

Most ops expect that the strings are in UTF-8. If you're using a different encoding, you can use the core tensorflow transcode op to transcode into UTF-8. You can also use the same op to coerce your string to structurally valid UTF-8 if your input could be invalid.

docs = tf.constant([u'Everything not saved will be lost.'.encode('UTF-16-BE'),
                    u'Sad☹'.encode('UTF-16-BE')])
utf8_docs = tf.strings.unicode_transcode(docs, input_encoding='UTF-16-BE',
                                         output_encoding='UTF-8')

When dealing with different sources of text, it's important that the same words are recognized to be identical. A common technique for case-insensitive matching in Unicode is case folding (similar to lower-casing). (Note that case folding internally applies NFKC normalization.)

We also provide Unicode normalization ops for transforming strings into a canonical representation of characters, with Normalization Form KC being the default (NFKC).

print(text.case_fold_utf8(['Everything not saved will be lost.']))
print(text.normalize_utf8(['Äffin']))
print(text.normalize_utf8(['Äffin'], 'nfkd'))

tf.Tensor(['everything not saved will be lost.'], shape=(1,), dtype=string)
tf.Tensor(['\xc3\x84ffin'], shape=(1,), dtype=string)
tf.Tensor(['A\xcc\x88ffin'], shape=(1,), dtype=string)

Tokenization is the process of breaking up a string into tokens. Commonly, these tokens are words, numbers, and/or punctuation.

The main interfaces are Tokenizer and TokenizerWithOffsets which each have a single method tokenize and tokenizeWithOffsets respectively. There are multiple implementing tokenizers available now. Each of these implement TokenizerWithOffsets (which extends Tokenizer) which includes an option for getting byte offsets into the original string. This allows the caller to know the bytes in the original string the token was created from.

All of the tokenizers return RaggedTensors with the inner-most dimension of tokens mapping to the original individual strings. As a result, the resulting shape's rank is increased by one. Please review the ragged tensor guide if you are unfamiliar with them. https://www.tensorflow.org/guide/ragged_tensor

This is a basic tokenizer that splits UTF-8 strings on ICU defined whitespace characters (eg. space, tab, new line).

tokenizer = text.WhitespaceTokenizer()
tokens = tokenizer.tokenize(['everything not saved will be lost.', u'Sad☹'.encode('UTF-8')])
print(tokens.to_list())

[['everything', 'not', 'saved', 'will', 'be', 'lost.'], ['Sad\xe2\x98\xb9']]

This tokenizer splits UTF-8 strings based on Unicode script boundaries. The script codes used correspond to International Components for Unicode (ICU) UScriptCode values. See: http://icu-project.org/apiref/icu4c/uscript_8h.html

In practice, this is similar to the WhitespaceTokenizer with the most apparent difference being that it will split punctuation (USCRIPT_COMMON) from language texts (eg. USCRIPT_LATIN, USCRIPT_CYRILLIC, etc) while also separating language texts from each other.

tokenizer = text.UnicodeScriptTokenizer()
tokens = tokenizer.tokenize(['everything not saved will be lost.',
                             u'Sad☹'.encode('UTF-8')])
print(tokens.to_list())

[['everything', 'not', 'saved', 'will', 'be', 'lost', '.'],
 ['Sad', '\xe2\x98\xb9']]

When tokenizing languages without whitespace to segment words, it is common to just split by character, which can be accomplished using the unicode_split op found in core.

tokens = tf.strings.unicode_split([u"仅今年前".encode('UTF-8')], 'UTF-8')
print(tokens.to_list())

[['\xe4\xbb\x85', '\xe4\xbb\x8a', '\xe5\xb9\xb4', '\xe5\x89\x8d']]

When tokenizing strings, it is often desired to know where in the original string the token originated from. For this reason, each tokenizer which implements TokenizerWithOffsets has a tokenize_with_offsets method that will return the byte offsets along with the tokens. The start_offsets lists the bytes in the original string each token starts at (inclusive), and the end_offsets lists the bytes where each token ends at (exclusive, i.e., first byte after the token).

tokenizer = text.UnicodeScriptTokenizer()
(tokens, start_offsets, end_offsets) = tokenizer.tokenize_with_offsets(
    ['everything not saved will be lost.', u'Sad☹'.encode('UTF-8')])
print(tokens.to_list())
print(start_offsets.to_list())
print(end_offsets.to_list())

[['everything', 'not', 'saved', 'will', 'be', 'lost', '.'],
 ['Sad', '\xe2\x98\xb9']]
[[0, 11, 15, 21, 26, 29, 33], [0, 3]]
[[10, 14, 20, 25, 28, 33, 34], [3, 6]]

Tokenizers work as expected with the tf.data API. A simple example is provided below.

docs = tf.data.Dataset.from_tensor_slices([['Never tell me the odds.'],
                                           ["It's a trap!"]])
tokenizer = text.WhitespaceTokenizer()
tokenized_docs = docs.map(lambda x: tokenizer.tokenize(x))
iterator = tokenized_docs.make_one_shot_iterator()
print(iterator.get_next().to_list())
print(iterator.get_next().to_list())

[['Never', 'tell', 'me', 'the', 'odds.']]
[["It's", 'a', 'trap!']]

When you use different tokenizers and ops to preprocess your data, the resulting outputs are Ragged Tensors. The Keras API makes it easy now to train a model using Ragged Tensors without having to worry about padding or masking the data, by either using the ToDense layer which handles all of these for you or relying on Keras built-in layers support for natively working on ragged data.

model = tf.keras.Sequential([
  tf.keras.layers.InputLayer(input_shape=(None,), dtype='int32', ragged=True)
  text.keras.layers.ToDense(pad_value=0, mask=True),
  tf.keras.layers.Embedding(100, 16),
  tf.keras.layers.LSTM(32),
  tf.keras.layers.Dense(32, activation='relu'),
  tf.keras.layers.Dense(1, activation='sigmoid')
])

TF.Text packages other useful preprocessing ops. We will review a couple below.

A common feature used in some natural language understanding models is to see if the text string has a certain property. For example, a sentence breaking model might contain features which check for word capitalization or if a punctuation character is at the end of a string.

Wordshape defines a variety of useful regular expression based helper functions for matching various relevant patterns in your input text. Here are a few examples.

tokenizer = text.WhitespaceTokenizer()
tokens = tokenizer.tokenize(['Everything not saved will be lost.',
                             u'Sad☹'.encode('UTF-8')])

# Is capitalized?
f1 = text.wordshape(tokens, text.WordShape.HAS_TITLE_CASE)
# Are all letters uppercased?
f2 = text.wordshape(tokens, text.WordShape.IS_UPPERCASE)
# Does the token contain punctuation?
f3 = text.wordshape(tokens, text.WordShape.HAS_SOME_PUNCT_OR_SYMBOL)
# Is the token a number?
f4 = text.wordshape(tokens, text.WordShape.IS_NUMERIC_VALUE)

print(f1.to_list())
print(f2.to_list())
print(f3.to_list())
print(f4.to_list())

[[True, False, False, False, False, False], [True]]
[[False, False, False, False, False, False], [False]]
[[False, False, False, False, False, True], [True]]
[[False, False, False, False, False, False], [False]]

N-grams are sequential words given a sliding window size of n. When combining the tokens, there are three reduction mechanisms supported. For text, you would want to use Reduction.STRING_JOIN which appends the strings to each other. The default separator character is a space, but this can be changed with the string_separater argument.

The other two reduction methods are most often used with numerical values, and these are Reduction.SUM and Reduction.MEAN.

tokenizer = text.WhitespaceTokenizer()
tokens = tokenizer.tokenize(['Everything not saved will be lost.',
                             u'Sad☹'.encode('UTF-8')])

# Ngrams, in this case bi-gram (n = 2)
bigrams = text.ngrams(tokens, 2, reduction_type=text.Reduction.STRING_JOIN)

print(bigrams.to_list())

[['Everything not', 'not saved', 'saved will', 'will be', 'be lost.'], []]

When installing TF Text with pip install, please note the version of TensorFlow you are running, as you should specify the corresponding version of TF Text. For example, if you're using TF 2.0, install the 2.0 version of TF Text, and if you're using TF 1.15, install the 1.15 version of TF Text.

pip install -U tensorflow-text==<version>

After version 2.10, we will only be providing pip packages for Linux x86_64 and Intel-based Macs. TensorFlow Text has always leveraged the release infrastructure of the core TensorFlow package to more easily maintain compatible releases with minimal maintenance, allowing the team to focus on TF Text itself and contributions to other parts of the TensorFlow ecosystem.

For other systems like Windows, Aarch64, and Apple Macs, TensorFlow relies on build collaborators, and so we will not be providing packages for them. However, we will continue to accept PRs to make building for these OSs easy for users, and will try to point to community efforts related to them.

Note that TF Text needs to be built in the same environment as TensorFlow. Thus, if you manually build TF Text, it is highly recommended that you also build TensorFlow.

If building on MacOS, you must have coreutils installed. It is probably easiest to do with Homebrew.

1. build and install TensorFlow.
2. Clone the TF Text repo:

   git clone https://github.com/tensorflow/text.git
   cd text

3. Run the build script to create a pip package:

   ./oss_scripts/run_build.sh

   After this step, there should be a *.whl file in current directory. File name similar to tensorflow_text-2.5.0rc0-cp38-cp38-linux_x86_64.whl.
4. Install the package to environment:

   pip install ./tensorflow_text-*-*-*-os_platform.whl

1. Pull image from Tensorflow SIG docker builds.

2. Run a container based with the pulled image and create a bash session. This can be done by running docker run -it {image_name} bash.
   {image_name} can be any name with {tf_verison}-python{python_version} format. An example for python 3.10 and TF version 2.10 :- 2.10-python3.10.

3. Clone the TF-Text Github repository inside container: git clone https://github.com/tensorflow/text.git.
   Once cloned, change to the working directory using cd text/.

4. Run the configuration script(s): ./oss_scripts/configure.sh and ./oss_scripts/prepare_tf_dep.sh.
   This will update bazel and TF dependencies to installed tensorflow in the container.

5. To run the tests, use the bazel command: bazel test --test_output=errors tensorflow_text:all. This will run all the tests declared in the BUILD file.
   To run a specific test, modify the above command replacing :all with the test name (for example :fast_bert_normalizer).

6. Build the pip package/wheel:
   bazel build --config=release_cpu_linux oss_scripts/pip_package:build_pip_package
./bazel-bin/oss_scripts/pip_package/build_pip_package /{wheel_dir}
   

   Once the build is complete, you should see the wheel available under {wheel_dir} directory.