Aspect-based sentiment analysis of Filipino COVID-19 tweets with memory networks

• Main Author: Guzman, John Paul
• Format: text
• Language: English
• Published: Animo Repository 2024
• Subjects: Machine learning; Sentiment analysis; Twitterbots; Mathematics
• Online Access: https://animorepository.dlsu.edu.ph/etdm_math/10; https://animorepository.dlsu.edu.ph/context/etdm_math/article/1011/viewcontent/2024_Guzman_Aspect_Based_Sentiment_Analysis_of_Filipino_COVID_19_Tweets_With_Full_text.pdf
• Institution: De La Salle University

Sentiment analysis is the process of extracting opinions from text. It has a wide variety of applications, such as monitoring social media, processing customer reviews, and improving health communication. Aspect-based sentiment classification (ABSC) is a type of sentiment analysis that determines sentiments toward specified aspects in a given text. Knowledge about the specific target of the opinion allows for a more fine-grained analysis of sentiments. This results in richer insights that can lead to further applications.In this paper, the goal is to model the ABSC task for text written in both English and Filipino. Furthermore, we aim to make this model adaptable to text in various languages and topics. To achieve this, we study the mathematical foundations of machine learning and deep learning. We explore techniques to overcome the challenges associated with multilingual ABSC, such as working with limited data and varying grammar structures. Then, we construct a memory network as a mathematical model for the task of multilingual ABSC. Our model is evaluated on four datasets to demonstrate its capacity to learn ABSC in different languages and topics. Three benchmark datasets in English are used to compare our model with existing ones. Additionally, we created a multilingual dataset referred to as the Filipino Health Twitter (FHT) dataset. This dataset consists of health-related tweets published in the Philippines during the COVID-19 pandemic. The tweets are written in English and Filipino, with possible code-switching. Our model achieved slight improvements over similar models in the benchmark datasets. Meanwhile, its performance in the FHT dataset is comparable to that observed in the benchmark datasets. Finally, we demonstrated potential applications that utilize insights from an ABSC model. This includes using ABSC to analyze the sentiments gathered during the COVID-19 pandemic.