Fajri Koto, Jey Han Lau, and Timothy Baldwin. IndoBERTweet: A Pretrained Language Model for Indonesian Twitter with Effective Domain-Specific Vocabulary Initialization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP 2021), Dominican Republic (virtual).
IndoBERTweet is the first large-scale pretrained model for Indonesian Twitter that is trained by extending a monolingually trained Indonesian BERT model with additive domain-specific vocabulary.
In this paper, we show that initializing domain-specific vocabulary with average-pooling of BERT subword embeddings is more efficient than pretraining from scratch, and more effective than initializing based on word2vec projections.
3. Pretraining Data
We crawl Indonesian tweets over a 1-year period using the official Twitter API, from December 2019 to December 2020, with 60 keywords covering 4 main topics: economy, health, education, and government. We obtain in total of 409M word tokens, two times larger than the training data used to pretrain IndoBERT. Due to Twitter policy, this pretraining data will not be released to public.
4. How to use
Load model and tokenizer (tested with transformers==3.5.1)
from transformers import AutoTokenizer, AutoModel tokenizer = AutoTokenizer.from_pretrained("indolem/indobertweet-base-uncased") model = AutoModel.from_pretrained("indolem/indobertweet-base-uncased")
- lower-case all words
- converting user mentions and URLs into @USER and HTTPURL, respectively
- translating emoticons into text using the .
5. Results over 7 Indonesian Twitter Datasets
|IndoBERT (Willie, et al., 2020)||84.1||88.7||73.3||86.8||80.4||86.3||84.3||83.4|
|IndoBERT (Koto, et al., 2020)||84.1||87.9||71.0||86.4||79.3||88.0||86.9||83.4|
|IndoBERTweet (1M steps from scratch)||86.2||90.4||76.0||88.8||87.5||88.1||85.4||86.1|
|IndoBERT + Voc adaptation + 200k steps||86.6||92.7||79.0||88.4||84.0||87.7||86.9||86.5|