Improvement of sentiment analysis via re-evaluation of objective words in SenticNet for hotel reviews

• Published in: Language resources and evaluation Vol. 55; no. 2; pp. 585 - 595
• Main Authors: Hung, Chihli, Wu, Wan-Rong, Chou, Hsien-Ming
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 01.06.2021; Springer Nature B.V
• Subjects: Classification; Computational Linguistics; Computer Science; Data mining; Decision trees; Language and Literature; Lexicon; Linguistics; Multilayer perceptrons; Negation; Neural networks; Orientation; Polarity; Project Notes; Sentences; Sentiment analysis; Social Sciences; Support vector machines; Sentiment lexicon; Sentiment analysis; Objective word; Word of mouth; SenticNet
• ISSN: 1574-020X; 1574-0218
• DOI: 10.1007/s10579-020-09512-6

In order to extract the correct sentiment polarity from word of mouth (WOM), a wide-scale and well-organized sentiment lexicon is generally beneficial. SenticNet is one such lexicon. However, it consists of a high proportion of objective words, which are generally considered to be of little use for sentiment classification due to their ambiguity and lack of sentiments. In the literature, there is a dearth of models that focus on this issue. An objective word appearing more frequently in positive sentences than in negative sentences implies a strong relationship in a positive sentiment orientation, and conversely, an objective word appearing more frequently in negative sentences implies a strong relationship in a negative sentiment orientation. Thus, the ratio of an objective word appearing in positive and negative sentences provides the sentiment orientation. Based on this concept, this paper re-assigns the sentiment values to the objective words in SenticNet and builds a revised SenticNet. Three classification techniques, the J48 decision tree, support vector machine, and multilayer perceptron neural network are used for classification. According to the experiments, the proposed models which extract sentiment values from the revised SenticNet, significantly outperform those models which extract sentiment values from the original non-revised SenticNet.