On social media platforms, the increase in the number of users and the resulting increase in thedata produced by users have accelerated the prominence of some technologies. The most well-known of these areas is Natural Language Processing (NLP), where sentiment analysis of usercontent is performed. Sentiment analysis is mostly applied to short texts. In this study,experimental comparisons of large language models (LLMs), known deep learning and machinelearning algorithms are made using a dataset containing content from Twitter (X) users. Inaddition to this comparison, the LoRA (Low-Rank Adaption) strategy was applied to theDistilBERT model, one of the transformer-based large language models that forms the basis ofthe study, and the success of this fine-tuning method, which is lower cost and parameterefficient, was demonstrated. 8 different algorithms were included in the study: MachineLearning algorithms (Logistic Regression, Support Vector Machines (SVM), Random Forest,XGBoost), Deep Learning Algorithms (Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM)) and Transformer-Based Algorithms (DistilBERT and DistilBERTfine-tuned with LoRA). According to the results obtained, the LoRA-supported DistilBERTalgorithm was the highest-performing algorithm with 82.38% accuracy. The results show thattransformer-based architectures are much more efficient than classical deep learning andmachine learning algorithms, especially in textual classification processes, when supported bynew methods and strategies that increase efficiency in terms of fine-tuning, such as LoRA. It isanticipated that this study will be an important guide in applications such as sentiment analysis,where model selection is important.
Tunahan TİMUÇİN (Fri,) studied this question.