Dynamic Noise-based Semantic Related Distance Weights with Attention Network for Aspect Sentiment Analysis

Aspect Based Sentiment Analysis (ABSA), is a method of sentiment analysis that focuses on identifying aspects/features in text and classifying their sentiment. Recently, many deep learning-based methods have achieved some success. However, they often focus solely on the relationship between aspect entities and sentiment words, overlooking the inherent fuzziness and dynamics of natural language. To tackle this issue, we propose a method that calculates semantic relevance weight based on dynamic noise, and combine it with an attention network to form our model (DNRA). Experimental results on restaurants dataset show that by our approach, accuracy is improved by 0.70%, reaching 83.83% compared to the baseline model. Meanwhile, we obtained the highest F1 scores compared to the baseline models on three datasets: laptops, restaurants, and Twitter, achieving scores of 73.33%, 76.59%, and 75.22%, respectively. In addition, we conducted a series of experiments to explore how different levels of noise affect performance on various datasets.