Undoubtedly that the Bidirectional Encoder representations from Transformers is the most powerful technique in making Natural Language Processing tasks such as Named Entity Recognition, Question & Answers or Sentiment Analysis, however, the use of traditional techniques remains a major potential for the improvement of recent models, in particular word tokenization techniques and embeddings, but also the improvement of neural network architectures which are now the core of each architecture. recent. In this paper, we conduct a comparative study between Fine-Tuning the Bidirectional Encoder Representations from Transformers and a method of concatenating two embeddings to boost the performance of a stacked Bidirectional Long Short-Term Memory-Bidirectional Gated Recurrent Units model; these two approaches are applied in the context of sentiment analysis of shopping places in Morocco. A search for the best learning rate was made at the level of the two approaches, and a comparison of the best optimizers was made for each sentence embedding combination with regard to the second approach.
View on arXiv