Unveiling Sentiments in Vietnamese Education Texts: Could Large Language Model GPT-3.5-turbo Beat PhoBERT?

Text classification in general, and sentiment analysis specifically is a major branch of Natural Language Processing with various uses and challenges. Classifying sentiments from Vietnamese text presents numerous challenges as Vietnamese is inherently and linguistically dissimilar to English. Therefore, there urges measures to recognize subtleties in Vietnamese texts and accurately categorize them into suitable labels. In this work, we leverage the power of the pre-trained model PhoBERT to classify a collected dataset of education-related texts into three separate classes, namely positive, negative, and neutral. It is expected that the model will have positive social implications by bringing people updated information on the current education system and ensuring their comprehension of educational policies, thus accordingly adapting itself towards more progressive growth. The model is first trained on two datasets VLSP 2019 and ViHSD. We then use the fine-tuned model to train on our data gathered from Vietnam online news sites and social media platforms. The input is either a sentence or a paragraph and the expected output is among the three aforementioned labels. Through testing and fine-tuning different models including Large Language Models (LLMs) with one-shot and few-shot approaches, our model consistently outperformed them in sentiment classification tasks with an accuracy of up to 94%.