Transparent Aspect-Level Sentiment Analysis Based on Dependency Syntax Analysis and Its Application on COVID-19

ABSTRACT

Aspect-level sentiment analysis identifies fine-grained emotion for target words. There are three major issues in current models of aspect-level sentiment analysis. First, few models consider the natural language semantic characteristics of the texts. Second, many models consider the location characteristics of the target words, but ignore the relationships among the target words and among the overall sentences. Third, many models lack transparency in data collection, data processing, and results generating in sentiment analysis. In order to resolve these issues, we propose an aspect-level sentiment analysis model that combines a bidirectional Long Short-Term Memory (LSTM) network and a Graph Convolutional Network (GCN) based on Dependency syntax analysis (Bi-LSTM-DGCN). Our model integrates the dependency syntax analysis of the texts, and explicitly considers the natural language semantic characteristics of the texts. It further fuses the target words and overall sentences. Extensive experiments are conducted on four benchmark datasets, i.e., Restaurantl4, Laptop, Restaurantl6, and Twitter. The experimental results demonstrate that our model outperforms other models like Target-Dependent LSTM (TD-LSTM), Attention-based LSTM with Aspect Embedding (ATAE-LSTM), LSTM+SynATT+TarRep and Convolution over a Dependency Tree (CDT). Our model is further applied to aspect-level sentiment analysis on "government" and "lockdown" of 1,658,250 tweets about "#COVID-19" that we collected from March 1, 2020 to July 1, 2020. The experimental results show that Twitter users' positive and negative sentiments fluctuated over time. Through the transparency analysis in data collection, data processing, and results generating, we discuss the reasons for the evolution of users' emotions over time based on the tweets and on our models.