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 sub-word units (also known as wordpieces [Wu et al.] [Schuster et al.] and byte-pair-encoding (BPE) [Sennrich et al.]) with the extension of 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.
This is not an official Google product.
- Purely data driven: SentencePiece trains tokenization and detokenization models from only raw sentences. No pre-tokenization (Moses tokenizer/MeCab/KyTea) is required.
- Language independent: SentencePiece treats the sentences just as sequences of Unicode characters. There is no language-dependent logic.
- 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.
What is SentencePiece?
SentencePiece is an unsupervised text tokenizer and detokenizer designed mainly for Neural Network-based text generation, for example Neural Network Machine Translation. SentencePiece is a re-implementation of sub-word units (also known as wordpieces [Wu et al.][Schuster et al.] and byte-pair-encoding (BPE) [Sennrich et al.]). Unlike previous sub-word approaches that train tokenizers from pretokenized sentences, SentencePiece directly trains the tokenizer and detokenizer from raw sentences. SentencePiece might seem like a sort of unsupervised word segmentation, but there are several differences and constraints in SentencePiece.
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.
Whitespace is considered as 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.
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)
The following tools and libraries are required to build SentencePiece:
- GNU autotools (autoconf automake libtool)
- C++11 compiler
On Ubuntu, autotools and libprotobuf can be install with apt-get:
% sudo apt-get install autoconf automake libtool libprotobuf-c++ protobuf-compiler libprotobuf-dev
Build and Install SentencePiece
% cd /path/to/sentencepiece % ./autogen.sh % ./configure % make % make check % sudo make install
Train SentencePiece Model
% spm_train --input=<input> --model_prefix=<model_name> --vocab_size=8000 --model_type=<type>
--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.
--vocab_size: vocabulary size, e.g., 8000, 16000, or 32000
--model_type: model type. Choose from
word. The input sentence must be pretokenized when using
spm_train loads only the first
--input_sentence_size sentences (default value is 10M).
--help flag to display all parameters for training.
Encode raw text into sentence pieces/ids
% spm_encode --model=<model_file> --output_format=piece < input > output % spm_encode --model=<model_file> --output_format=id < input > output
--extra_options flag to insert the BOS/EOS 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>)
Decode sentence pieces/ids into raw text
% spm_decode --model=<model_file> --input_format=piece < input > output % spm_decode --model=<model_file> --input_format=id < input > output
--extra_options flag to decode the text in reverse order.
% spm_decode --extra_options=reverse < input > output
% spm_train --input=data/botchan.txt --model_prefix=m --vocab_size=1000 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 % 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
% spm_export_vocab --model=<model_file> --output=<output file>
<output file> stores a list of vocabulary and emission log probabilities. The vocabulary id corresponds to the line number in this file.
We have evaluated SentencePiece segmentation with the following configurations.
- BPE (Byte Pair
Encoding) [Sennrich et al.]] (
- Unigram. Language-model based segmentation. (
- BPE (Byte Pair Encoding) [Sennrich et al.]] (
- NoPretok: No pretokenization. We train SentencePiece directly from
raw sentences (
- WsPretok: Trains SentencePiece model from the sentences tokenized by
--split_by_whitespace=true). When handling CJK, this setting is almost equivalent to NoPretok.
- MosesPretok: Trains SentencePiece model from sentences tokenized by Moses tokenizer. We used KyTea for Japanese and in-house segmenters for Korean and Chinese respectively.
- NoPretok: No pretokenization. We train SentencePiece directly from raw sentences (
NMT parameters: (Google’s Neural Machine Translation System is applied for all experiments.)
- 16k shared vocabulary (Shares the same vocabulary for source and target. We train single SentencePiece model by concatenating raw source and target sentences.)
- Dropout prob: 0.2
- num nodes: 512
- num lstms: 8
- Case-sensitive BLEU on detokenized text with NIST scorer.
- For CJK, the same word segmenters are applied prior to NIST scorer.
- No detokenizer is applied for NoPretok and WsPretok, which can directly emit detokenized sentences.
- Applied Moses detokenizer and in-house rule-based detokenizer (CJK) for MosesPretok.
Results (BLEU scores)
|In house en-ko||0.178||0.1851||0.1893||0.1846||0.1872||0.1890|
|In house ko-en||0.1786||0.1954||0.1994||0.1845||0.1956||0.2015|
- MosesPretok does not always improve BLEU scores. Comparable accuracy can be obtained without using language-dependent resources in many language pairs.
- Whitespace pretokenization is a reasonable choice. It does not use language-specific resources.
- NoPretok shows poor BLEU scores. Unigrams are more robust than BPE when no pretokenizer is applied.