Automatic construction of target-specific sentiment lexicon. (February 2019)

• Record Type: Journal Article
• Authors: Wu, Sixing; Wu, Fangzhao; Chang, Yue; Wu, Chuhan; Huang, Yongfeng
• Is Part Of: Expert systems with applications. Volume 116(2019)
• Page Start: 285
• Page End: 298
• Publication Date: 2019-02
• Subjects: Opinion mining -- Sentiment lexicon -- Sentiment analysis -- Opinion target extraction; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2018.09.024
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: An approach to construct a target-specific sentiment lexicon is proposed. An unsupervised method to extract opinion targets is proposed. The method achieves both high accuracy and recall in extraction. A framework is proposed to classify sentiment polarities of opinion pairs accurately. The framework extracts and integrates different sources of sentiment information. Abstract: Sentiment lexicon plays an important role in sentiment analysis system. In most existing sentiment lexica, each sentiment word or phrase is given a sentiment label or score. However, a sentiment word may express different sentiment orientations describing different targets. It's beneficial but challenging to incorporate knowledge of opinion targets into sentiment lexicon. In this paper we propose an automatic approach to construct a target-specific sentiment lexicon, in which each term is an opinion pair consisting of an opinion target and an opinion word. The approach solves two principle problems in construction process, namely, opinion target extraction and opinion pair sentiment classification. An unsupervised algorithm is proposed to extract opinion pairs in high quality. Both semantic feature and syntactic feature are incorporated in the algorithm, to extract opinion pairs containing correct opinion targets. A group of opinion pairs are generated and a framework is proposed to classify their sentiment polarities. Knowledge of available resources including general-purpose sentiment lexicon and thesaurus, and context knowledge including syntactic relations and sentiment information in sentences, are extracted and integrated in a unified framework to calculate sentiment scores of opinion pairs. Experimental results on product reviews datasets in different domains prove the effectiveness of our method in target-specific sentiment lexicon construction, which can improve performances of opinion target extraction and opinion pair sentiment classification. In addition, our lexicon also achieves better performance in target-level sentiment classification compared with several general-purpose sentiment lexicons.