Unsupervised and knowledge-poor approaches to sentiment analysis

摘要: Sentiment analysis focuses upon automatic classiffication of a document's sentiment (and more generally extraction opinion from text). Ways expressing have been shown to be dependent on what document is about (domain-dependency). This complicates supervised methods for which rely extensive use training data or linguistic resources that are usually either domain-specific generic. Both kinds prevent classiffiers performing well across range domains, as this requires appropriate in-domain (domain-specific) data. This thesis presents novel unsupervised, knowledge-poor approach aimed at creating domain-independent and multilingual system. The extracts documents processed, uses them analysis. does not require any corpora, large sets rules generic lexicons, makes it domain- languageindependent but the same time able utilise language-specific information. The describes tests approach, applied diffeerent data, including customer reviews various types products, films books, news items; four languages: Chinese, English, Russian Japanese. The only binary classiffication, also three-way (positive, negative neutral), subjectivity classifiation sentences, holders targets. Experimental results suggest often viable alternative systems, especially when collections.