Journal "Software Engineering"
a journal on theoretical and applied science and technology
ISSN 2220-3397
Issue N10 2019 year
Social networks, microblogging sites, online shops, forums and other resources are now becoming more and more popular. People use Internet as a platform for their self-expression. Everyone can share their experiences and give a product feedback. People read public reviews to choose hotels, restaurants, smartphones, weekend movies, service providers etc. Aspect-based sentiment analysis (ABSA) systems allow extracting user opinions about different product/service aspects and features. ABSA includes four main tasks: Aspect term extraction (ATE), Aspect term polarity (ATP), Aspect category detection (ACD) and Aspect category polarity (ACP). All of these tasks are domain-sensitive. It means that ABSA-system must be built for a particular domain or adapted to it to achieve high quality of classification. Lexicon-based systems are considered as more adaptable then machine-learning or hybrid systems. There are many adaptation techniques: with or without using of core lexicon; based on graphs, semantic networks, domain-sensitive thesauri etc. Some approaches use the concept of different level of domain-sensitivity to reduce the amount of adaptation procedures, but this idea requires a source of common-sense knowledge integrated to the system. This paper is devoted to the problem of domain adaptation in ABSA-tasks in Russian. It represents an approach based on generalized lexicons. Presented approach includes a statistical-based algorithm of sentiment lexicon extraction from short informal text corpuses. Mathematical model of sentiment mark is introduced. Semantic graph is used to extend basic sentiment lexicon. The graph contains the data extracted from common dictionaries, which is domain-independent, reliable and full enough to use presented graph as a source of common-sense knowledge. Graph structure allows adapting the data to a particular domain with keeping important common-sense knowledge in background. Semantic adaptation can also elevate the quality of explicit sentiment calculation based on graph.