Automatic Detection of Sentiment Towards Explicit Aspect in Russian Publicism Sentences Using Syntactic Structure

Abstract

The paper describes experiments on using the syntactic structure to improve the performance of sentiment detection towards an explicit aspect in Russian sentences. Three methods of syntactic structure usage were evaluated. The first is augmenting sentences from training set by changing their syntactic structure. The second is passing local context subtrees to Interactive Attention Network (IAN). The third is the novel semantic rule-based sentiment detection algorithm based on the previous authors’ work. Sentences augmentation and passing local context did not exceeded the baseline, their macro F-scores were 0.68 and 0.69 respectively. The proposed rule-based algorithm performed almost as well as the baseline method, its macro F-score is 0.70. The ensemble of BERT-SPC and the rule-based algorithm performed with F-score 0.81, which exceeds the baseline by 10 %.

Keywords

• natural language processing
• computational linguistics
• sentiment detection
• aspect-based sentiment analysis
• syntactic structure
• constituency trees