Exploring transformers for aspect based sentiment analysis

• Main Author: Poudel, Roshan
• Publication Date: 2023
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10773/41732
• topic: Sentiment analysis; Aspect-based sentiment analysis (ABSA); Transformers; Deep learning; Natural language processing (NLP)

In recent years, sentiment analysis has gained significant attention for its wide range of applications in various domains. Aspect-based sentiment analysis (ABSA) is a challenging task within sentiment analysis that aims to identify the sentiment polarity towards specific aspects or attributes of a target entity in a given text. Transformers, a type of deep neural network architecture, have shown promising results in many natural language processing (NLP) tasks, including sentiment analysis. This dissertation explores the effectiveness of a BERT+BiLSTM+CRF model for ABSA and investigates the impact of model sizes and layer freezing. Several experiments are performed using different ABSA datasets, comparing the results with existing state-of-the-art models. The findings indicate that increasing model size is not necessarily the best approach to improve performance, and freezing a subset of layers can lead to comparable results with reduced computational requirements. The study also highlights the impact of pretraining methods and datasets in downstream tasks. The developed end-to-end system is modular, allowing for the replacement of BERT with any other transformer-based model based on the use case. The research contributes to the understanding of transformer-based models for ABSA and provides insights for future studies in this field.