| Sumario: | Sentiment analysis (SA) is a study area focused on obtaining contextual polarity from the text. Currently, deep learning has obtained outstanding results in this task. However, much annotated data are necessary to train these algorithms, and obtaining this data is expensive and difficult. In the context of low-resource scenarios, this problem is even more significant because there are little available data. Transfer learning (TL) can be used to minimize this problem because it is possible to develop some architectures using fewer data. Language models are a way of applying TL in natural language processing (NLP), and they have achieved competitive results. Nevertheless, some models need many hours of training using many computational resources, and in some contexts, people and organizations do not have the resources to do this. In this paper, we explore the models BERT (Pretraining of Deep Bidirectional Transformers for Language Understanding), MultiFiT (Efficient Multilingual Language Model Fine-tuning), ALBERT (A Lite BERT for Self-supervised Learning of Language Representations), and RoBERTa (A Robustly Optimized BERT Pretraining Approach). In all of our experiments, these models obtain better results than CNN (convolutional neural network) and LSTM (Long Short Term Memory) models. To MultiFiT and RoBERTa models, we propose a pretrained language model (PTLM) using Twitter data. Using this approach, we obtained competitive results compared with the models trained in formal language datasets. The main goal is to show the impacts of TL and language models comparing results with other techniques and showing the computational costs of using these approaches.
|
|---|