Aspect Category Detection with Attention Mechanism
Deena Nath, Sanjay Kumar Dwivedi · 2024
Aspect Category Detection (ACD), a sub-task of Aspect-Based Sentiment Analysis (ABSA), has become prominent in recent years due to the rise in customer reviews. Unlike conventional ABSA research that deals with domains such as laptops and restaurants, our research take a novel approach by analyzing three datasets associated to Indian government policies, Indian Budget and Parliamentary debates. We explore various methods for aspect category detection, dividing them into supervised and unsupervised learning categories, ultimately identifying the most suitable approach for our research. We began by presenting the new Topic-Attention Network Model (TANM) that employs the Swish activation function to increase the precision and reliability of aspect category recognition. The proposed model adapts contextual word embedding with topic distribution, which uses Bidirectional LSTM (BiLSTM) to infer aspect categories from the text. The topic-aware attention mechanism is another facet of our approach since it focus on those words which could be important for each aspect. The outcome of the context vectors is then passed through a fully connected layer for accurate aspect category prediction. Thus the method significantly contributes positive improvement in aspect-based sentiment analysis by showing enhanced capability in pinpointing subtle and implicit facets.