wechsel

韋希塞爾

Wechsel的代碼：在NAACL2022上發表的單語言模型的跨語言轉移的子詞嵌入的有效初始化。

論文：https：//aclanthology.org/2022.naacl-main.293/

紙上的模型可在Huggingface Hub上使用：

• roberta-base-wechsel-french
• roberta-base-wechsel-german
• roberta-base-wechsel-chinese
• roberta-base-wechsel-swahili
• gpt2-wechsel-french
• gpt2-wechsel-german
• gpt2-wechsel-chinese
• gpt2-wechsel-swahili

安裝

我們通過PYPI分發Python包：

 pip install wechsel

另外，請克隆存儲庫，安裝requirements.txt並在wechsel/中運行代碼。

示例用法

將英語roberta-base轉移到斯瓦希里語：

 import torch
from transformers import AutoModel , AutoTokenizer
from datasets import load_dataset
from wechsel import WECHSEL , load_embeddings

source_tokenizer = AutoTokenizer . from_pretrained ( "roberta-base" )
model = AutoModel . from_pretrained ( "roberta-base" )

target_tokenizer = source_tokenizer . train_new_from_iterator (
    load_dataset ( "oscar" , "unshuffled_deduplicated_sw" , split = "train" )[ "text" ],
    vocab_size = len ( source_tokenizer )
)

wechsel = WECHSEL (
    load_embeddings ( "en" ),
    load_embeddings ( "sw" ),
    bilingual_dictionary = "swahili"
)

target_embeddings , info = wechsel . apply (
    source_tokenizer ,
    target_tokenizer ,
    model . get_input_embeddings (). weight . detach (). numpy (),
)

model . get_input_embeddings (). weight . data = torch . from_numpy ( target_embeddings )
model . config . vocab_size = len ( target_embeddings )

# if the model has separate output embeddings, also copy those
if not model . config . tie_word_embeddings :
    target_out_embeddings , info = wechsel . apply (
        source_tokenizer ,
        target_tokenizer ,
        model . get_output_embeddings (). weight . detach (). numpy (),
    )
    model . get_output_embeddings (). weight . data = torch . from_numpy ( target_out_embeddings )

# use `model` and `target_tokenizer` to continue training in Swahili! 

雙語詞典

我們將3276個雙語詞典從英語分發到其他語言，以與dicts/一起使用。

引用

請引用韋切塞爾

 @inproceedings{minixhofer-etal-2022-wechsel,
    title = "{WECHSEL}: Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models",
    author = "Minixhofer, Benjamin  and
      Paischer, Fabian  and
      Rekabsaz, Navid",
    booktitle = "Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jul,
    year = "2022",
    address = "Seattle, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.naacl-main.293",
    pages = "3992--4006",
    abstract = "Large pretrained language models (LMs) have become the central building block of many NLP applications. Training these models requires ever more computational resources and most of the existing models are trained on English text only. It is exceedingly expensive to train these models in other languages. To alleviate this problem, we introduce a novel method {--} called WECHSEL {--} to efficiently and effectively transfer pretrained LMs to new languages. WECHSEL can be applied to any model which uses subword-based tokenization and learns an embedding for each subword. The tokenizer of the source model (in English) is replaced with a tokenizer in the target language and token embeddings are initialized such that they are semantically similar to the English tokens by utilizing multilingual static word embeddings covering English and the target language. We use WECHSEL to transfer the English RoBERTa and GPT-2 models to four languages (French, German, Chinese and Swahili). We also study the benefits of our method on very low-resource languages. WECHSEL improves over proposed methods for cross-lingual parameter transfer and outperforms models of comparable size trained from scratch with up to 64x less training effort. Our method makes training large language models for new languages more accessible and less damaging to the environment. We make our code and models publicly available.",
}

致謝

Google TPU研究雲（TRC）的Cloud TPU支持研究。我們感謝Andy Koh和Artus Krohn-Grimberghe提供了其他計算資源。 Ellis Unit Linz，LIT AI實驗室，機器學習研究所，由聯邦州立Upper Austria支持。我們感謝該項目Incontrol-RL（FFG-881064）。