**sub-word units** (also known as **wordpieces** [[Wu et al.](https://arxiv.org/pdf/1609.08144.pdf)]
[[Schuster et al.](https://static.googleusercontent.com/media/research.google.com/ja//pubs/archive/37842.pdf)]
and **byte-pair-encoding (BPE)** [[Sennrich et al.](http://www.aclweb.org/anthology/P16-1162)]) 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](https://github.com/moses-smt/mosesdecoder/blob/master/scripts/tokenizer/tokenizer.perl)/[MeCab](http://taku910.github.io/mecab/)/[KyTea](http://www.phontron.com/kytea/)) is required.
- **Language independent**: SentencePiece treats the sentences just as sequences of Unicode characters. There is no language-dependent logic.
- **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.
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.](https://arxiv.org/pdf/1609.08144.pdf)][[Schuster et al.](https://static.googleusercontent.com/media/research.google.com/ja//pubs/archive/37842.pdf)] and **byte-pair-encoding (BPE)** [[Sennrich et al.](http://www.aclweb.org/anthology/P16-1162)]). Unlike previous sub-word approaches that train tokenizers from pretokenized sentences, SentencePiece directly trains the tokenizer and detokenizer from raw sentences.
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.
*`--model_type`: model type. Choose from `unigram` (default), `bpe`, `char`, or `word`. The input sentence must be pretokenized when using `word` type.
***neologd**: [MeCab with neologd](https://github.com/neologd/mecab-ipadic-neologd) for Japanese.
***(Moses/KyTea)+SentencePiece**: Apply SentencePiece (Unigram) to pre-tokenized sentences. We have several variants with different tokenizers., e.g., **(Moses/MeCab)+SentencePiece**, **(MeCab/Moses)+SentencePiece**.
* **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 is 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.
* Pretokenization can slightly improve the BLEU scores in English to Japanese. In Japanese to English translation, pretokenization doesn't help to improve BLEU.
* **Neologd** shows poor BLEU score. Toeknizing sentences with a large named entity dictionary might not be effective in neural-based text processing.
* **SentencePiece(Unigram)** shows slightly better text compression ratio than **BPE**, but no significant differences in BLEU score.
* The selection of vocabulary size for SentencePiece is sensitive in English to Japanese. This is probably because the vocabulary size will drastically affect the tokenization results in Japanese which has no explicit spaces between words.
***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](https://github.com/moses-smt/mosesdecoder/blob/master/scripts/tokenizer/tokenizer.perl). We used [KyTea](http://www.phontron.com/kytea/) for
Japanese and in-house segmenters for Korean and Chinese respectively.
* NMT parameters: ([Google’s Neural Machine Translation System](https://arxiv.org/pdf/1609.08144.pdf) 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](https://github.com/moses-smt/mosesdecoder/blob/master/scripts/tokenizer/detokenizer.perl) and in-house rule-based detokenizer (CJK) for **MosesPretok**.
* Data sets:
* [KFTT](http://www.phontron.com/kftt/index.html)
* [MultiUN](http://opus.lingfil.uu.se/MultiUN.php) (First 5M and next
5k/5k sentences are used for training and development/testing respectively.)
* [WMT16](http://www.statmt.org/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.](http://www.aclweb.org/anthology/P16-1162)] and [[Wu et al.](https://arxiv.org/pdf/1609.08144.pdf)].
