A Hierarchical Emotion Classification Technique for Thai Reviews

Abstract

Emotion classification is an interesting problem in affective computing that can be applied in various tasks, such as speech synthesis, image processing and text processing. With the increasing amount of textual data on the Internet, especially reviews of customers that express opinions and emotions about products. These reviews are important feedback for companies. Emotion classification aims to identify an emotion label for each review. This research investigated three approaches for emotion classification of opinions in the Thai language, written in unstructured format, free form or informal style. Different sets of features were studied in detail and analyzed. The experimental results showed that a hierarchical approach, where the subjectivity of the review is determined first, then the polarity of opinion is identified and finally the emotional label is calculated, yielded the highest performance, with precision, recall and F-measure at 0.691, 0.743 and 0.709, respectively.

Comparing Hierarchical Approaches to Enhance Supervised Emotive Text Classification

“…In the literature, a popular method to address this problem is to divide the class hierarchy into a set of independent classification problems and apply classification at each hierarchical level. For example, Esmin et al [13], Ghazi et al [14], Charoensuk and Sornil [15], and Angiani et al [16] undertook an investigation into a three-level hierarchical classification approach towards classifying emotions expressed in text. The first hierarchical level focused on classifying whether the texts were emotive or not, the second level aimed to classify their sentiment polarity, and, subsequently, the third level aimed to classify whether texts expressed one of Ekman's six basic emotions.…”

A Survey of Automatic Text Classification Based on Thai Social Media Data

“…Kucut is a Python-language library for cutting Thai text via corpus and machine learning. Research includes Charoensuk and Sornil (2018), Sanguansat (2016), Vateekul and Koomsubha (2016), and Viriyavisuthisakul et al (2015).…”

“…Trakultaweekoon and Klaithin (2016) classified texts by sentiment toward a product, service, person, place, or event. Charoensuk and Sornil (2018) classified texts based on the emotion or feeling that the writers wanted to communicate.…”

A New Term Frequency with Gaussian Technique for Text Classification and Sentiment Analysis

“…tfg(𝑡, 𝐺) = log 𝑓 𝑡,𝑑 + 1 * 𝐺 (6) where 𝐺 is the distribution value of term 𝑖 in document 𝑗 of category 𝑐; 𝑛 is the number of documents in category 𝑐; 𝑤 , is the frequency of term 𝑖 in document 𝑗; 𝜇 , is the mean of all terms 𝑖 that are within a document in category 𝑐; 𝜎 , is the variance in term 𝑖 within a document in category 𝑐; 𝑐 is the category of a document; and 𝑓 , is the frequency of term (t) in document (d).…”

“…Previous studies used traditional term weighting and machine learning (ML) techniques, but their accuracy was not high. In addition, these studies utilized datasets from the Internet that did not explicitly identify classes of text [4][5][6][7].…”