riotu-lab / aranizer

Aranizer: A Custom Tokenizer based on SentencePiece and BPE tailored for Arabic Language Modeling

AraNizer

Description

AraNizer is a sophisticated toolkit of custom tokenizers tailored for Arabic language processing. Integrating advanced methodologies such as SentencePiece and Byte Pair Encoding (BPE), these tokenizers are specifically designed for seamless integration with the transformers and sentence_transformers libraries. The AraNizer suite offers a range of tokenizers, each optimized for distinct NLP tasks and accommodating varying vocabulary sizes to cater to a multitude of linguistic applications.

Key Features

• Versatile Tokenization: Supports multiple tokenization strategies (BPE, SentencePiece) for varied NLP tasks.
• Broad Vocabulary Range: Customizable tokenizers with vocabulary sizes ranging from 32k to 86k.
• Seamless Integration: Compatible with popular libraries like transformers and sentence_transformers.
• Optimized for Arabic: Specifically engineered for the intricacies of the Arabic language.

Installation

Install AraNizer effortlessly with pip:

pip install aranizer

Usage

Importing Tokenizers

Import your desired tokenizer from AraNizer. Available tokenizers include: BEP variants: get_bpe with keys bpe32, bpe50, bpe64, bpe86, bpe32T, bpe50T, bpe64T, bpe86T SentencePiece variants: get_sp with keys sp32, sp50, sp64, sp86, sp32T, sp50T, sp64T, sp86T

from aranizer import get_bpe, get_sp  # Import functions to retrieve tokenizers

# Example for importing a BPE tokenizer
bpe_tokenizer = get_bpe("bpe32")  # Replace with your chosen tokenizer key

# Example for importing a SentencePiece tokenizer
sp_tokenizer = get_sp("sp32")  # Replace with your chosen tokenizer key

Tokenizing Text

Tokenize Arabic text using the selected tokenizer:

text = "مثال على النص العربي"  # Example Arabic text

# Using BPE tokenizer
bpe_tokens = bpe_tokenizer.tokenize(text)
print(bpe_tokens)

# Using SentencePiece tokenizer
sp_tokens = sp_tokenizer.tokenize(text)
print(sp_tokens)

Encoding and Decoding

Encode text into token ids and decode back to text.

Encoding: To encode text, use the encode method.

text = "مثال على النص العربي"  # Example Arabic text

# Using BPE tokenizer
encoded_bpe_output = bpe_tokenizer.encode(text, add_special_tokens=True)
print(encoded_bpe_output)

# Using SentencePiece tokenizer
encoded_sp_output = sp_tokenizer.encode(text, add_special_tokens=True)
print(encoded_sp_output)

Decoding: To convert token ids back to text, use the decode method:

# Using BPE tokenizer
decoded_bpe_text = bpe_tokenizer.decode(encoded_bpe_output)
print(decoded_bpe_text)

# Using SentencePiece tokenizer
decoded_sp_text = sp_tokenizer.decode(encoded_sp_output)
print(decoded_sp_text)

Available Tokenizers

Available Tokenizers
get_bpe("bpe32"): Based on BPE Tokenizer with Vocab Size of 32k
get_bpe("bpe50"): Based on BPE Tokenizer with Vocab Size of 50k
get_bpe("bpe64"): Based on BPE Tokenizer with Vocab Size of 64k
get_bpe("bpe86"): Based on BPE Tokenizer with Vocab Size of 86k
get_bpe("bpe32T"): Based on BPE Tokenizer with Vocab Size of 32k (with Tashkeel (diacritics))
get_bpe("bpe50T"): Based on BPE Tokenizer with Vocab Size of 50k (with Tashkeel (diacritics))
get_bpe("bpe64T"): Based on BPE Tokenizer with Vocab Size of 64k (with Tashkeel (diacritics))
get_bpe("bpe86T"): Based on BPE Tokenizer with Vocab Size of 86k (with Tashkeel (diacritics))
get_sp("sp32"): Based on SentencePiece Tokenizer with Vocab Size of 32k
get_sp("sp50"): Based on SentencePiece Tokenizer with Vocab Size of 50k
get_sp("sp64"): Based on SentencePiece Tokenizer with Vocab Size of 64k
get_sp("sp86"): Based on SentencePiece Tokenizer with Vocab Size of 86k
get_sp("sp32T"): Based on SentencePiece Tokenizer with Vocab Size of 32k (with Tashkeel (diacritics))
get_sp("sp50T"): Based on SentencePiece Tokenizer with Vocab Size of 50k (with Tashkeel (diacritics))
get_sp("sp64T"): Based on SentencePiece Tokenizer with Vocab Size of 64k (with Tashkeel (diacritics))
get_sp("sp86T"): Based on SentencePiece Tokenizer with Vocab Size of 86k (with Tashkeel (diacritics))

System Requirements

• Python 3.x
• transformers library

Contact:

For queries or assistance, please contact [email protected].

Acknowledgments:

This work is maintained by the Robotics and Internet-of-Things Lab at Prince Sultan University.

Team:

• Prof. Anis Koubaa (Lab Leader)
• Dr. Lahouari Ghouti (NLP Team Leader)
• Eng. Omar Najjar (AI Research Assistant)
• Eng. Serry Sebai (NLP Research Engineer)

Version:

0.2.3

Citations:

Coming soon

aranizer's Issues

Output is messy code when using the tokenizer of aranizer_bpexxx

When choosing the tokenizer of BPE in aranizer like aranizer_bpe50k, aranizer_bpe64k, the tokenized result will be messy code.
The input text is arabic text.