Methods of implicit aspect detection in Russian publicism sentences

Abstract

The paper compares performance of various methods of automatic implicit aspect detection in publicism sentences in Russian. The task of implicit aspect detection is an auxiliary task in the aspect-oriented sentiment analysis. The experiments were conducted on a corpus of sentences extracted from political campaign materials. The best results, with F1-measure reaching 0.84, were obtained using the Navec embeddings and classifiers based on the support vector machine method. Fairly high results, with F1-measure reaching 0.77, were obtained using the bag-of-words model and the naive Bayesian classifier. Other methods showed lower performance. It was also revealed during the experiments that the detection quality can differ significantly between the aspects. The detection quality is the highest for the aspects associated with characteristic marker words, for example, “health car” and “holding elections”. More general aspects, such as “quality of governance”, are detected with the worst quality.