Bertweetro: Pre-Trained Language Models for Romanian Social Media Content

The introduction of Transformers, like BERT or RoBERTa, have revolutionized NLP due to their ability to better “understand” the meaning of texts. These models are created (pre-trained) in a self-supervised manner on large scale data to predict words in a sentence but can be adjusted (fine-tuned) for other specific NLP applications. Initially, these models were created using literary texts but very quickly the need to process social media content emerged. Social media texts have some problematic characteristics (they are short, informal, filled with typos, etc.) which means that a traditional BERT model will have problems when dealing with this type of input. For this reason, dedicated models need to be pre-trained on microblogging content and many such models have been developed in popular languages like English or Spanish. For under-represented languages, like Romanian, this is more difficult to achieve due to the lack of open-source resources. In this paper we present our efforts in pre-training from scratch 8 BERTweetRO models, based on RoBERTa architecture, with the help of a Romanian tweets corpus. To evaluate our models, we fine-tune them on 2 down-stream tasks, Sentiment Analysis (with 3 classes) and Topic Classification (with 26 classes), and compare them against Multilingual BERT plus a number of other popular classic and deep learning models. We include a commercial solution in this comparison and show that some BERTweetRO variants and almost all models trained on the translated data have a better accuracy than the commercial solution. Our best performing BERTweetRO variants place second after Multilingual BERT in most of our experiments, which is a good result considering that our Romanian corpus used for pre-training is relatively small, containing around 51,000 texts.