Towards Domain-Independent Opinion Target Extraction

摘要

In this paper, we investigate the problem of domain-independent opinion target extraction. The only lexical resource used is domain-independent (general) sentiment dictionary. We begin from investigating syntactic descriptions (rules) using dependency parsing jointly with sentiment dictionary. We conclude that such a solution is not sufficient for opinion target extraction due to low precision. To overcome this difficulty, we propose a well-known supervised machine learning method as the second step, after applying syntactic rules. We find that supervised model without lexical features outperforms by large margin a comparable one with lexical features. The results appear promising and contribute to domain-independent opinion target extraction. All experiments were carried out on a publicly available Polish dependency treebank with manually verified opinion and sentiment annotations, as well as opinion target information.