Fine-tune BERT based on Machine Learning Models For Sentiment Analysis
Nadia Smairi, Houda Abadlia, Hajer Ben Brahim, Wided Lejouad Chaari · Procedia Computer Science · 2024
Sentiments Analysis is a technique applied to examine emotions expressed by the users of social networks where they massively convey their opinions and permanently share their feelings. Therefore, its main objective is to determine the emotional attitude of a discourse by classifying it into different categories as positive, negative or neutral. However, in some research works, the authors failed to accurately specify the sentiments from the context due to the varying text lengths and the presence of equivocal emotional cues. This paper sets forth the deployment and assessment of the capabilities of applying machine learning sentiment analysis techniques using a publicly available IMDB dataset. Notably, this dataset encompasses numerous instances of irony and sarcasm. Lately, the Bidirectional Encoder Representations from Transformers model (BERT) has showcased its efficacy in the domain of text sentiment analysis classification. However, there remains a scope for enhancing the accuracy of SA. In the current study, BERT, and Word2vec are utilized to extract contextual sentences embeddings, while Support Vector Machine (SVM) is used to classify the emotions of the social networks users. To enhance SVM accuracy, the genetic algorithm is integrated to optimize its hyper-parameters. The comparative study revealed that the developed technique applied on IMDB dataset outperforms other baseline machine learning techniques.