SentencePiece

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.

Technical highlights

• 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.

Overview

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 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

Required packages

The following tools and libraries are required to build SentencePiece:

• GNU autotools (autoconf automake libtool)
• C++11 compiler
• protobuf library

On Ubuntu, autotools and protobuf library can be install with apt-get:

% sudo apt-get install autoconf automake libtool libprotobuf9v5 protobuf-compiler libprotobuf-dev

(If libprotobuf9v5 is not found, try libprotobuf-c++ instead.)

Build and Install SentencePiece

% cd /path/to/sentencepiece
% ./autogen.sh
% ./configure
% make
% make check
% sudo make install
$ sudo ldconfig -v

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. <model_name>.model and <model_name>.vocab are generated.
• --vocab_size: vocabulary size, e.g., 8000, 16000, or 32000
• --model_type: model type. Choose from unigram (default), bpe, char, or word. The input sentence must be pretokenized when using word type.

Note that spm_train loads only the first --input_sentence_size sentences (default value is 10M).

Use --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

Use --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

Use --extra_options flag to decode the text in reverse order.

% spm_decode --extra_options=reverse < input > output

End-to-End Example

% 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.

Experiments 1 (subword vs word-based model)

Experimental settings

• Segmentation algorithms:

  • Unigram: SentencePiece with a language-model based segmentation. (--model_type=unigram)
  • BPE: SentencePiece with Byte Pair Encoding. [Sennrich et al.]] (--model_type=bpe)
  • Moses: Moses tokenizer for English.
  • KyTea: KyTea for Japanese.
  • MeCab: MeCab for Japanese.
  • neologd: MeCab with neologd for Japanese.

• NMT parameters: (Google’s Neural Machine Translation System is applied for all experiments.)

  • Dropout prob: 0.2
  • num nodes: 512
  • num lstms: 6

• Evaluation metrics:

  • Case-sensitive BLEU on detokenized text with NIST scorer and KyTea segmenter. Used in-house rule-based detokenizer for Moses/KyTea/MeCab/neologd.

• Data sets:

  • KFTT

Results (BLEU scores)

Setting	vocab size	BLEU(dev)	BLEU(test)	src #tokens/sent.	trg #tokens/sent.
enja (Unigram)	8k (shared)	0.2718	0.2922	30.97	25.05
enja (BPE)	8k (shared)	0.2695	0.2919	31.76	25.43
enja (Moses/KyTea)	80k/80k	0.2514	0.2804	21.25	23.21
enja (Moses/MeCab)	80k/80k	0.2436	0.2739	21.25	21.20
enja (Moses/neologd)	80k/80k	0.2102	0.2350	21.25	18.47
jaen (Unigram)	8k (shared)	0.1959	0.2170	25.05	30.97
jaen (BPE)	8k (shared)	0.1975	0.2176	25.43	31.76
jaen (KyTea/Moses)	80k/80k	0.1697	0.1974	23.21	21.25
jaen (MeCab/Moses)	80k/80k	0.1654	0.1870	21.20	21.25
jaen (neologd/Moses)	80k/80k	0.1583	0.1838	18.47	21.25

• SentencePiece (Unigram/BPE) outperforms word-based methods (Moses/KyTea/MeCab/neologd) even with a smaller vocabulary (10% of word-based methods).
• The number of tokens to represent Japanese sentences are almost comparable between SentencePiece (unigram) and KyTea, though the vocabulary of Sentencepice is much smaller. It implies that Sentencepiece can effectively compress the sentences with a smaller vocabulary set.
• Neologd shows poor BLEU score. Tokenizing sentences with a large named entity dictionary might not be effective in neural-based text processing.
• Unigram shows slightly better text compression ratio than BPE, but no significant differences in BLEU score.

Experiments 2 (subwording with various pre-tokenizations)

Experimental settings

We have evaluated SentencePiece segmentation with the following configurations.

• Segmentation algorithms:

  • BPE (Byte Pair Encoding) [Sennrich et al.]] (--model_type=bpe)
  • Unigram. Language-model based segmentation. (--model_type=unigram)

• pretokenization methods:

  • NoPretok: No pretokenization. We train SentencePiece directly from raw sentences (--split_by_whitespace=false).
  • WsPretok: Trains SentencePiece model from the sentences tokenized by whitespaces (--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.

• 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

• Evaluation metrics:

  • 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.

• Data sets:

  • KFTT
  • MultiUN (First 5M and next 5k/5k sentences are used for training and development/testing respectively.)
  • WMT16
  • In-house: (Used 5M parallel sentences for training)

NoPretok and WsPretok do not use any language-dependent resources. BPE+MosePretok is almost the same configuration used in [Sennrich et al.] and [Wu et al.].

Results (BLEU scores)

Language Pair	BPE(NoPretok)	BPE(WsPretok)	BPE(MosesPretok)	Unigram(NoPretok)	Unigram(WsPretok)	Unigram(MosesPretok)
KFTT en-ja	0.2796	0.281	0.286	0.2806	0.280	0.2871
KFTT ja-en	0.1943	0.208	0.1967	0.1985	0.2148	0.198
MultiUN ar-en	0.5268	0.5414	0.5381	0.5317	0.5449	0.5401
MultiUN en-ar	0.4039	0.4147	0.4012	0.4084	0.4172	0.3991
MultiUN en-zh	0.4155	0.4186	0.395	0.4214	0.4165	0.399
MultiUN zh-en	0.46	0.4716	0.4806	0.4644	0.4711	0.4759
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
WMT16 cs-en	0.1987	0.2252	0.2231	0.2164	0.2228	0.2238
WMT16 de-en	0.3194	0.3348	0.3374	0.3261	0.3375	0.3398
WMT16 en-cs	0.1607	0.1827	0.1812	0.1722	0.1778	0.179
WMT16 en-de	0.2847	0.3029	0.3013	0.2946	0.3000	0.3053
WMT16 en-fi	0.1434	0.1528	0.1499	0.1472	0.1568	0.1517
WMT16 en-ru	0.1884	0.1973	0.1989	0.19	0.1982	0.1903
WMT16 fi-en	0.1775	0.1867	0.1877	0.182	0.1882	0.1865
WMT16 ru-en	0.2042	0.2229	0.2194	0.2087	0.2201	0.2155

• 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.

Advanced topics

• SentencePieceProcessor C++ API
• Use custom text normalization rules
• Use custom symbols
• [Segmentation and training algorithms in detail]