Subject-Object Aspect-Based Sentiment Analysis Model Based on News Texts

Abstract

News as an important part of the open source intelligence has always played an important role in international relations and national security fields. However, the fine-grained sentiment analysis work such as Aspect-Based Sentiment Analysis (ABSA) tasks focused on the service industry and e-commerce comments and the Target Aspect-Based Sentiment Analysis (T-ABSA) tasks focused on the Twitter datasets, these two types of tasks usually limit the context to the fixed subjects or objects, simplifying the model by limiting several aspects simultaneously. In addition, the sentiment analysis work based on the news texts most focused on the chapters level. Therefore, based on the characteristics of news texts that it contains multiple subjects and objects, this paper proposed the Subject-Object Aspect-Based Sentiment Analysis (SO-ABSA) model, which can do fine-grained emotional element extraction in the context of indefinite subjects, objects and aspects. More emotional elements can be mined through SO-ABSA model. The proposed model can extract the uncertain entities efficiently and improve the accuracy of subjects and objects extraction. Moreover, the uncertain aspects also can be extracted flexibly and the sentiment analysis result can represent the subjects’ emotional tendency towards specific aspects. To evaluate our method, we built a subject-object oriented dataset (SOOD) with data sourced from 30,000 news articles. We propose a subject-object aspect emotion analysis model and evaluate the model on the SOOD dataset. The experimental results show the effectiveness of our model.