Article Preview
TopIntroduction
Sentiment analysis or opinion mining is an interesting research topic in recent years (Pang & Lee, 2008). It is widely applied to textual information in various domains, i.e., reviews of products and services by customers (Hu & Liu, 2004; Ding et al., 2008; Jo & Oh, 2011), financial criticism on microblogs (Si et al., 2013), medical records (Denecke & Deng, 2015), and online news (Chen & Li, 2017). Instead of determining the sentiment of the whole text, several research works, e.g., Jo and Oh (2011) and Moghaddam and Ester (2012), addressed sentiments at smaller levels such as parts, components, attributes, or aspects of an entity of interest. This kind of analysis is referred to as aspect-based sentiment analysis. It basically involves two main tasks, i.e., aspect extraction and sentiment classification of extracted aspects (Liu & Zhang, 2012).
For aspect extraction, descriptive statistics and topic modelling techniques have been used; for example, term frequencies and Latent Dirichlet Allocation (LDA) were used in the works of Hu and Liu (2004) and Jo and Oh (2011), respectively. For determining aspect sentiments, two major approaches, machine-learning-based and lexicon-based, have been applied. These two approaches, however, still perform poorly when they are applied to complicated text with rich linguistic structures. Consider, for example, the sentence “The new phone is fine, but its battery still lacks capacity for one-day use”, which should be classified as negative with respect to the “power” aspect. Since the terms “fine” and “lacks” indicate positive and negative sentiments, respectively, this sentence may be misclassified as neutral by using a lexicon-based method. For training a machine-learning classification model, when this sentence is used as a negative instance, terms in the first clause “The new phone is fine” may be incorrectly taken as features for the negative class. The use of linguistic structures makes it possible to consider only the second clause “but its battery still lacks capacity for one-day use” to be relevant to the “power” aspect, and the undesired effects of terms occurring in the first clause can be eliminated.
Recently, attempts to utilise linguistic structures to the sentiment analysis have been reported. Polanyi and van den Berg (2011) discussed the application of the Linguistic Discourse Model to sentiment analysis and observed that a discourse structure affects sentiments at the discourse level; however, no experimental result was reported. Zirn, Niepert, Stuckenschmidt, and Strube (2011) applied the Rhetorical Structure Theory (RST) (Mann & Thompson, 1988) to indicate parts of text that are relevant to a sentiment in product reviews, and an overall sentiment score was calculated from the indicated parts, with the sentiment classification accuracy of 69% being achieved. Sanglerdsinlapachai, Plangprasopchok, and Nantajeewarawat (2016) applied RST to identify text portions relevant to specific aspects in a dataset containing mobile phone reviews, and the sentiment classification accuracy of 72.4% was obtained by averaging the polarity scores of the relevant parts.
In this study, the authors step further by investigating the dependence of the overall sentiment of a given group of related text portions and the polarity scores of its main part (nucleus) and complement parts (satellites) and examining the effects of RST relation types on polarity score aggregation. Sentiment classification rules are induced from RST components, i.e., nuclei, satellites, and/or relation types. The induced rules are evaluated at the level of local aspect segments, each of which consists of related text portions with a well-defined boundary and can be easily annotated with a sentiment. Comparisons with methods that do not employ information about RST components are conducted.