SentencePiece

Note: We're actively refactoring the codebase to simplify support for open source releases and Google internal needs. During this time, builds and tests will be unstable and not all features will be supported.

SentencePiece is an unsupervised text tokenizer and detokenizer mainly for Neural Network-based text generation systems where the vocabulary size is predetermined prior to the neural model training. SentencePiece implements three approaches:

subword units (or byte-pair-encoding (BPE) Sennrich et al.)
unigram language model Kudo.]

It supports direct training from raw sentences. SentencePiece allows us to make a purely end-to-end system that does not depend on language-specific pre/postprocessing.

NOTE: This is not an official Google product.

Technical highlights

Purely data driven: SentencePiece trains tokenization and detokenization models from sentences. Pre-tokenization (Moses tokenizer/MeCab/KyTea) is not always required.
Language independent: SentencePiece treats the sentences just as sequences of Unicode characters. There is no language-dependent logic.
Multiple subword algorithms: BPE and unigram language model are supported.
Subword regularization: SentencePiece implements subword sampling for subword regularization which helps to improve the robustness and accuracy of NMT models.
Fast and lightweight: Segmentation speed is around 50k sentences/sec, and memory footprint is around 6MB.
Self-contained: The same tokenization/detokenization is obtained as long as the same model file is used.
Direct vocabulary id generation: SentencePiece manages vocabulary to id mapping and can directly generate vocabulary id sequences from raw sentences.
NFKC-based normalization: SentencePiece performs NFKC-based text normalization.

Comparisons with other implementations

Feature	SentencePiece	subword-nmt	WordPiece
Supported algorithm	BPE, unigram, char, word	BPE	BPE
OSS?	Yes	Yes	Google internal
Subword regularization	Yes	No	No
Python Library (pip)	Yes	No	N/A
C++ Library	Yes	No	N/A
Pre-segmentation required?	No	Yes	Yes
Customizable normalization (e.g., NFKC)	Yes	No	N/A
Direct id generation	Yes	No	N/A

Note that BPE algorithm used in WordPiece is slightly different from the original BPE.

Overview

What is SentencePiece?

SentencePiece is a re-implementation of sub-word units, an effective way to alleviate the open vocabulary problems in neural machine translation. SentencePiece supports two segmentation algorithms, byte-pair-encoding (BPE) Sennrich et al. and unigram language model Kudo.. Here are the high level differences from other implementations.

The number of unique tokens is predetermined

Neural Machine Translation models typically operate with a fixed vocabulary. Unlike most unsupervised word segmentation algorithms, which assume an infinite vocabulary, SentencePiece trains the segmentation model such that the final vocabulary size is fixed, e.g., 8k, 16k, or 32k.

Note that SentencePiece specifies the final vocabulary size for training, which is different from subword-nmt that uses the number of merge operations. The number of merge operations is a BPE-specific parameter and not applicable to other segmentation algorithms, including unigram, word and character.

Trains from raw sentences

Previous sub-word implementations assume that the input sentences are pre-tokenized. This constraint was required for efficient training, but makes the preprocessing complicated as we have to run language dependent tokenizers in advance. The implementation of SentencePiece is fast enough to train the model from raw sentences. This is useful for training the tokenizer and detokenizer for Chinese, Japanese and Korean where no explicit spaces exist between words.

Whitespace is treated as a basic symbol

The first step of Natural Language processing is text tokenization. For example, a standard English tokenizer would segment the text "Hello world." into the following three tokens.

[Hello] [World] [.]

One observation is that the original input and tokenized sequence are NOT reversibly convertible. For instance, the information that is no space between “World” and “.” is dropped from the tokenized sequence, since e.g., Tokenize(“World.”) == Tokenize(“World .”)

SentencePiece treats the input text just as a sequence of Unicode characters. Whitespace is also handled as a normal symbol. To handle the whitespace as a basic token explicitly, SentencePiece first escapes the whitespace with a meta symbol "▁" (U+2581) as follows.

Hello▁World.

Then, this text is segmented into small pieces, for example:

[Hello] [▁Wor] [ld] [.]

Since the whitespace is preserved in the segmented text, we can detokenize the text without any ambiguities.

  detokenized = ''.join(pieces).replace('_', ' ')

This feature makes it possible to perform detokenization without relying on language-specific resources.

Note that we cannot apply the same lossless conversions when splitting the sentence with standard word segmenters, since they treat the whitespace as a special symbol. Tokenized sequences do not preserve the necessary information to restore the original sentence.

(en) Hello world. → [Hello] [World] [.] (A space between Hello and World)
(ja) こんにちは世界。 → [こんにちは] [世界] [。] (No space between こんにちは and 世界)

Subword regularization

Subword regularization Kudo.] is a simple regularization method that virtually augments training data with on-the-fly subword sampling, which helps to improve the accuracy as well as robustness of NMT models.

To enable subword regularization, you would like to integrate SentencePiece library (C++/Python) into the NMT system to sample one segmentation for each parameter update, which is different from the standard off-line data preparations. Here's the example of Python library. You can find that 'New York' is segmented differently on each SampleEncode call. The details of sampling parameters are found in sentencepiece_processor.h.

>>> import sentencepiece as spm
>>> s = spm.SentencePieceProcessor()
>>> s.Load('spm.model')
>>> for n in range(5):
...     s.SampleEncodeAsPieces('New York', -1, 0.1)
...
['▁', 'N', 'e', 'w', '▁York']
['▁', 'New', '▁York']
['▁', 'New', '▁Y', 'o', 'r', 'k']
['▁', 'New', '▁York']
['▁', 'New', '▁York']

Installation

Pre-requisites

SentencePiece requires the following before setup:

Bazel

C++

Building the underlying C++ libraries is as easy as

bazel build src:all

Testing can be done with

bazel test src:all

Python module

We provide python API access for the SentencePieceProcessor to support encoding and decoding text.

You can install the Python API with:

% pip install sentencepiece

For more detail, see Python module

TensorFlow module

We will be no longer supported direct integration with Tensorflow. Tensorflow users are suggested to adopt the new Tokzenization ops published as part of TF.Text. Those ops will soon support running pre-trained SentencePiece models.

API Examples

Train a model

Train a sample model with full character coverage and a 8000 vocabulary size:

spm_train \
  --input=<input> \
  --model_prefix=<model_name> \
  --vocab_size=8000 \
  --character_coverage=1.0 \
  --model_type=<type>

Key flags:

input: one-sentence-per-line raw corpus file. No need to run tokenizer, normalizer or preprocessor. By default, SentencePiece normalizes the input with Unicode NFKC. You can pass a comma-separated list of files.
model_prefix: output model name prefix. <model_name>.model and <model_name>.vocab are generated.
vocab_size: vocabulary size, e.g., 8000, 16000, or 32000
character_coverage: amount of characters covered by the model, good defaults are: 0.9995 for languages with rich character set like Japanse or Chinese and 1.0 for other languages with small character set.
--model_type: model type. Choose from unigram (default), bpe, char, or word. The input sentence must be pretokenized when using word type.

The help flag will other parameters for training.

Encode raw text into sentence pieces/ids

The spm_encode tool can encode the text into a piece sequence given a pre-trained model:

spm_encode \
  --model=<model_file> \
  --output_format=piece \
  'input text' \
  output

With output_format=id, spm_encode will convert the text to an id sequence:

spm_encode \
  --model=<model_file> \
  --output_format=id \
  'input text' \
  output

The extra_options flag can be changed insert the end of sentence or begin of sentence markers or reverse the input sequence:

spm_encode --extra_options=eos # (add </s> only)
spm_encode --extra_options=bos:eos # (add <s> and </s>)
spm_encode --extra_options=reverse:bos:eos # (reverse input and add <s> and </s>)

The output_format flag can be changed to support nbest segmentation and nbest sampling with the nbest_ and sample_ prefixes:

spm_encode --model=<model_file> --output_format=sample_piece --nbest_size=-1 --alpha=0.5 'input text' output
spm_encode --model=<model_file> --output_format=nbest_id --nbest_size=10 'input_text' output

Decode sentence pieces/ids into raw text

The spm_decode tool can convert segmented content into unsegmented text from either white space separated pieces or an id sequence:

spm_decode --model=<model_file> --input_format=piece 'input pieces' output
% spm_decode --model=<model_file> --input_format=id 'input ids' output

The extra_options can be set to reverse the ordering:

spm_decode --extra_options=reverse 'input' output

End-to-End Example

We can train a simple 1000 vocabulary sized model on the text in data/botchan.txt with the following:

spm_train \
  --input=data/botchan.txt \
  --model_prefix=m \
  --vocab_size=1000

While running, we should see the following text:

unigram_model_trainer.cc(494) LOG(INFO) Starts training with :
input: "../data/botchan.txt"
... <snip>
unigram_model_trainer.cc(529) LOG(INFO) EM sub_iter=1 size=1100 obj=10.4973 num_tokens=37630 num_tokens/piece=34.2091
trainer_interface.cc(272) LOG(INFO) Saving model: m.model
trainer_interface.cc(281) LOG(INFO) Saving vocabs: m.vocab

The resulting model can then be tested with a few command lines:

% echo "I saw a girl with a telescope." | spm_encode --model=m.model
▁I ▁saw ▁a ▁girl ▁with ▁a ▁ te le s c o pe .

% echo "I saw a girl with a telescope." | spm_encode --model=m.model --output_format=id
9 459 11 939 44 11 4 142 82 8 28 21 132 6

% echo "9 459 11 939 44 11 4 142 82 8 28 21 132 6" | spm_decode --model=m.model --input_format=id
I saw a girl with a telescope.

You can find that the original input sentence is restored from the vocabulary id sequence.

Export vocabulary list

The spm_export_vocab tool emits the learned vocabulary and emission probabilities. Vocabulary IDs will correspond to the piece's line number in the vocab file.

spm_export_vocab \
  --model=<model_file> \
  --output=<output file>

Redefine special meta tokens

By default, SentencePiece uses the following meta tokens with default reserved vocabulary IDs:

<unk> with ID 0 for Unknown tokens.
<s> with ID 1 for BOS.
</s> with ID 2 for EOS.
`<pad> with no pre-reserved Id for padding.

We can redefine this mapping in the training phase as follows.

spm_train \
  --bos_id=0 \
  --eos_id=1 \
  --unk_id=5 \
  --input=... \
  --model_prefix=... \
  --character_coverage=...

Each meta token can be disabled by using assigning -1. For example, bos_id=-1 disabled the BOS meta token.

NOTE: This does not work for unk_id. The unknown token must always be in the vocabulary.

The full set of pre-defined meta token flags are:

bos_id: for BOS.
eos_id: for EOS.
unk_id: for Unknown.

If you want to assign another special tokens, please see Use custom symbols.

Vocabulary restriction

spm_encode accepts a --vocabulary and a --vocabulary_threshold option so that spm_encode will only produce symbols which also appear in the vocabulary (with at least some frequency). The background of this feature is described in subword-nmt page.

The usage is basically the same as that of subword-nmt. Assuming that L1 and L2 are the two languages (source/target languages), train the shared spm model, and get resulting vocabulary for each:

cat {train_file}.L1 {train_file}.L2 | shuffle > train
spm_train --input=train --model_prefix=spm --vocab_size=8000 --character_coverage=0.9995
spm_encode --model=spm.model --generate_vocabulary < {train_file}.L1 > {vocab_file}.L1
spm_encode --model=spm.model --generate_vocabulary < {train_file}.L2 > {vocab_file}.L2

shuffle command is used just in case because spm_train loads the first 10M lines of corpus by default.

Then segment train/test corpus with --vocabulary option

spm_encode --model=spm.model --vocabulary={vocab_file}.L1 --vocabulary_threshold=50 < {test_file}.L1 > {test_file}.seg.L1
spm_encode --model=spm.model --vocabulary={vocab_file}.L2 --vocabulary_threshold=50 < {test_file}.L2 > {test_file}.seg.L2

Advanced topics

SentencePiece Experiments
SentencePieceProcessor C++ API
Use custom text normalization rules
Use custom symbols
Python Module
TensorFlow Module
[Segmentation and training algorithms in detail]

huangyue / sentencepiece