Improving explicit aspects extraction in sentiment analysis using optimized ruleset / Mohammad Ahmad Jomah Tubishat

Mohammad Ahmad, Jomah Tubishat (2019) Improving explicit aspects extraction in sentiment analysis using optimized ruleset / Mohammad Ahmad Jomah Tubishat. PhD thesis, University of Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (207Kb)
Preview	PDF (Thesis PhD) Download (1588Kb) | Preview

Abstract

Aspect extraction, also known as opinion target extraction, is the fine-grained identification of users’ opinion targets, such as the extraction of opinionated product aspects from customer reviews. Aspect extraction is considered as the core task in aspect-based sentiment analysis and other applications. Currently, many studies were conducted using dependency relation rules which give promising results. However, these dependency-based extraction approaches perform better on formal text as its accuracy is based on the dependency parser which gives correct results if the text follow the English rules and grammars. On the other hand, there are also many studies were conducted using sequential syntactic patterns which mimic and follow the ways users expressed their opinion without giving attention to the language rules but give better results on informal text. However, customer reviews normally are a mixed of both types of reviews including formal and informal text. In addition, extraction rules including either pattern-based or dependency-based rules should be selected in a correct way to remove the irrelevant rules and minimize the extraction errors Thus, in this study, to select the most effective extraction rules, an improved version of Whale Optimization Algorithm (IWOA) is developed and applied to a full set of rules. This set of rules includes combination of new created extraction rules with dependency-based rules and pattern-based rules from the previous studies. In addition, the improved WOA is developed by using Cauchy mutation and local search algorithm to solve its local optima problem and improve population diversity. The algorithm was then applied to the full set of 126 rules. Finally, after the aspects list was obtained from the selected rules, a pruning algorithm (PA) is developed to remove the incorrect aspects and retain the correct aspects. Our results from the conducted experiments revealed that the proposed algorithm outperform the state-of-the-art aspect extraction algorithms and optimization algorithms. The IWOA algorithm outperforms other optimization algorithms includes native WOA, PSO, MFO, FFA, GWO, MVO, SSA, and SCA and achieved 86% precision, 94% recall, and 90% F-measure respectively. IWOA superiority resulted because of its ability to escape from local optima and balance between exploitation and exploration. In addition, after application of PA, IWOA+PA outperforms other state-of-the-art aspect extraction works and achieved 92% precision, 93% recall, and 92% F-measure respectively.

Actions (For repository staff only : Login required)