sappho192 / BertJapaneseTokenizer

Minimal Tokenizer implementation of BertJapanese(cl-tohoku/bert-base-japanese) in C#

BertJapaneseTokenizer

Minimal Tokenizer implementation of BertJapanese(cl-tohoku/bert-base-japanese) in C#

Quickstart

1. Just build the BertJapaneseTokenizer project and copy the BertJapaneseTokenizer.dll file to your dotnet project.
2. Add the BertJapaneseTokenizer.dll to your project reference.
3. In your project, install [Mecab.DotNet] package from Nuget.
4. Download unidic mecab dictionary unidic-mecab-2.1.2_bin.zip from https://clrd.ninjal.ac.jp/unidic_archive/cwj/2.1.2/ and unzip the archive into somewhere.
5. Download vocab file BertJapanese from Huggingface. For example, vocab.txt of bert-base-japanese-v2 can be accessed from [here].
   (Or you can simply use my extension method GetVocabFromHub(). See the example below.)
6. Check the example code below and you are good to go.

using BertJapaneseTokenizer;

var dicPath = @"D:\DATASET\unidic-mecab-2.1.2_bin";
//var vocabPath = @"D:\DATASET\bert-japanese\bert-base-japanese-v2\vocab.txt";
var vocabPath = await HuggingFace.GetVocabFromHub("cl-tohoku/bert-base-japanese-v2");
var tokenizer = new BertJapaneseTokenizer.BertJapaneseTokenizer(dicPath, vocabPath);

var sentence = "打ち合わせが終わった後にご飯を食べましょう。";
//var sentence = "ご飯を食べましょう。";
//var sentence = "打ち合わせ";

(var tokenIds, var attentionMask) = tokenizer.EncodePlus(sentence);

Console.WriteLine($"Sentence: {sentence}");
Console.WriteLine($"Token IDs: {string.Join(", ", tokenIds)}");

var decoded = tokenizer.Decode(tokenIds);
Console.WriteLine($"Decoded: {decoded}");

To-do List

• Implement Decode() method
• Support BPE-type vocabulary (like cl-tohoku/bert-base-japanese-char)