Article Preview
Top1. Introduction
The task of sentiment analysis was first formulated in the early 2000s, particularly in the work of Dave, Lawrence, and Pennock (2003) as well as that of Nasukawa and Yi (2003). In the past fifteen years, a great deal of research has sought to analyse and evaluate opinions on products and services in media sources (Singh, Sharma, Dey, 2015). As elaborated in Sentiment Analysis and Subjectivity. Handbook of Natural Language Processing, Second Edition (2010), there are three levels of analysis worth considering: (i) document-level, (ii) sentence-level, and (iii) aspect-level. In reference to document level analysis, the research of Sharma, Nigam, and Jain (2014), Sharma, Hoque, and Chandra (2016), Tang, Qin, Liu, and Yang (2015), and Xia, Xu, Yu, and Qi, (2016) need to be taken into consideration. Significant works which discuss sentence-based sentiment analysis include Marcheggiani, Tackstrom, Esuli, and Sebastiani (2014) and Yang and Cardie, (2014). Our core research interest is in aspect-based sentiment analysis, the focus of several important studies, including those of Chinsha and Joseph (2015) and Wang, Jiang, Lan, and Wu, (2017). The challenge of sentiment analysis is that the opinions which users provide on various topics are often complex and multifaceted. The sentiment valence of a word can change depending on its specific context, a phenomenon which is referred to as contextual valence shifting (Polanyi and Zaenen, 2006). Contextual valence shifting has made traditional methods, such as machine learning with bag of words or n-gram models, ineffective, as the latter tend to tackle single words, assigning them a positive or negative polarity based on pre-defined sentiment lexicons. In contrast, the most widely used techniques in sentiment analysis tend to account for contextual valence shifting by paying considerable attention to semantic elements, namely words or phrases which are factors giving rise to polarity shifting (Carrillo de Albornoz, Plaza, & Gervas, 2010; Jia, Yu, & Meng, 2009; Tran & Phan, 2018).
In the present paper we propose an aspect-based sentiment analysis model for Vietnamese reviews, which combines an opinion dictionary and syntactic dependency rules to extract reliable word pairs (sentiment - aspect). Using domain-specific ontologies we grouped these aspects and determined the sentiment polarity of each.
The key innovations presented here are as follows:
- ●
We have provided an effective model for solving the sentiment analysis problem by combining an opinion dictionary, domain-specific ontologies, and syntactic dependency rules.
- ●
We have extracted multi-word sentiment phrases (instead of individual words) for fine-grained (multi-class) classification, including extracting nouns and verbs which bear emotional factors, rather than adjectives alone.
The remainder of this paper is organized as follows: In Section 2 related work is summarised. In Section 3 the proposed sentiment analysis model is outlined. In Section 4 the experiments used to evaluate the model are described. Finally, the results of the paper are summarized and avenues for future work are discussed.