A comparative study of embedding techniques for aspect-based sentiment analysis

Abstract Sentiment analysis has become increasingly vital for understanding public opinion, customer feedback, and market trends. Traditional sentiment analysis methods often struggle to capture nuanced opinions expressed towards different aspects or features of a product or service. To address this challenge, our study proposes an aspect-based sentiment analysis (ABSA) approach leveraging Natural Language Processing (NLP) techniques and long-short-term memory (LSTM), a powerful recurrent neural network model. LSTM networks have achieved promising results for ABSA tasks. However, training LSTMs from scratch can be time-consuming. This study examines the efficacy of various techniques for word embedding, including Word2Vec and pre-trained GloVe embeddings and a statistical tool TF-IDF, in accelerating LSTM-based ABSA models. The individual performance of these techniques is compared by integrating them into LSTM models on a benchmark ABSA dataset and evaluating their ability to capture contextual nuances and sentiment associations for different aspects. Our findings provide insights into selecting the most suitable feature extraction method for ABSA tasks, demonstrating that pre-trained GloVe embeddings outperform Word2Vec and TF-IDF in capturing aspect-specific sentiments. Furthermore, this study highlights the importance of embedding quality and its impact on the overall performance of sentiment analysis models.