README.md

---
license: cc-by-nc-sa-4.0
datasets:
- wikipedia
- cc100
language:
- ja
library_name: transformers
pipeline_tag: fill-mask
---

BERT-base (Juman++ + BPE)
===

## How to load the tokenizer
Please download the dictionary file for Juman++ + BPE from [our GitHub repository](https://github.com/hitachi-nlp/compare-ja-tokenizer/blob/public/data/dict/jumanpp_bpe.json).
Then you can load the tokenizer by specifying the path of the dictionary file to `dict_path`.

```python
from typing import Optional

from tokenizers import Tokenizer, NormalizedString, PreTokenizedString
from tokenizers.processors import BertProcessing
from tokenizers.pre_tokenizers import PreTokenizer
from transformers import PreTrainedTokenizerFast

from pyknp import Juman
import mojimoji
import textspan

class JumanPreTokenizer:
    def __init__(self):
        self.juman = Juman("jumanpp", multithreading=True)
    
    def tokenize(self, sequence: str) -> list[str]:
        text = mojimoji.han_to_zen(sequence).rstrip()
        try:
            result = self.juman.analysis(text)
        except:
            traceback.print_exc()
            text = ""
            result = self.juman.analysis(text)
        return [mrph.midasi for mrph in result.mrph_list()]
    
    def custom_split(self, i: int, normalized_string: NormalizedString) -> list[NormalizedString]:
        text = str(normalized_string)
        tokens = self.tokenize(text)
        tokens_spans = textspan.get_original_spans(tokens, text)
        return [normalized_string[st:ed] for cahr_spans in tokens_spans for st,ed in cahr_spans]
    
    def pre_tokenize(self, pretok: PreTokenizedString):
        pretok.split(self.custom_split)

# load a pre-tokenizer
pre_tokenizer = JumanPreTokenizer()

# load a tokenizer
dict_path = /path/to/jumanpp_bpe.json
tokenizer = Tokenizer.from_file(dict_path)
tokenizer.post_processor = BertProcessing(
    cls=("[CLS]", tokenizer.token_to_id('[CLS]')),
    sep=("[SEP]", tokenizer.token_to_id('[SEP]'))
)

# convert to PreTrainedTokenizerFast
tokenizer = PreTrainedTokenizerFast(
    tokenizer_object=tokenizer,
    unk_token='[UNK]',
    cls_token='[CLS]',
    sep_token='[SEP]',
    pad_token='[PAD]',
    mask_token='[MASK]'
)

# set a pre-tokenizer
tokenizer._tokenizer.pre_tokenizer = PreTokenizer.custom(pre_tokenizer)
```

```python
# Test
test_str = "こんにちは。私は形態素解析器について研究をしています。"
tokenizer.convert_ids_to_tokens(tokenizer(test_str).input_ids)
# -> ['[CLS]','こ','んに','ち','は','。','私','は','形態','素','解析','器','に','ついて','研究','を','して','い','ます','。','[SEP]']
```

## How to load the model
```python
from transformers import AutoModelForMaskedLM
model = AutoModelForMaskedLM.from_pretrained("hitachi-nlp/bert-base_jumanpp-bpe")
```


**See [our repository](https://github.com/hitachi-nlp/compare-ja-tokenizer) for more details!**