Article Preview
TopIntroduction
Marketing strategy involves gathering accurate, timely, and sufficient information for supporting marketers to make better decisions. Complete information should cover all necessary customer perspectives, i.e. their feedbacks on the company’s products and services (Amado, Cortez, Rita, & Moro, 2018), (Piao et al., 2019). The rapid growth of advanced technologies such as internet of things, cloud computing, and social network platforms allow the internet users to feedback their opinions on products and services conveniently (Ofusori & Africa, 2021), (Zhou, 2021). Such online user feedbacks reveal to customer satisfaction provided in digital forms including multimedia. Therefore, marketers may use customer feedbacks (CFs) to understand the customers’ opinions on products and services in order to improve their products and services (Alengadan & Khan, 2017), (Alaei et al., 2019).
Even though CFs appear to be helpful in the social marketing, they are mostly represented in unstructured data (Ramachandran & Sciences, 2017), (Wang, Xin, Wang, Huang, & Liu, 2017), (Piao et al., 2019; Thelwall, Buckley, & Paltoglou, 2012) which are for human to read. They cannot immediately be interpreted by the machine (Bello-Orgaz, Jung, & Camacho, 2016), (Sivarajah, Kamal, Irani, & Weerakkody, 2017), (Jung, 2017). It requires practitioners and experts to analyze CFs. This work is time consuming and labor intensive. Therefore, there is a need to automatically extract the information from CFs.
There have been increasing interests in automatic information extraction from CFs (W. Chen, Yu, Xian, Wang, & Wen, 2020). Multi-level sentiment information extraction is a technique that can provide the information more accurately than a single level (Piao et al., 2019), (Cambria, 2016). In the past decade, most of researches proposed the sentiment extraction only on a document level using several techniques (Araque, Corcuera-Platas, Sánchez-Rada, & Iglesias, 2017), (Hu & Liu, 2004), (Diamantini, Mircoli, Potena, & Storti, 2018). T. Chen and et.al (T. Chen et al., 2017) and X. Fang and J. Zhan (Fang & Zhan, 2015) proposed word-level, sentence-level, and document-level sentiment extraction. They proved that the sentiment information from multiple levels improved the extraction performance. However, they do not consider contrast information.
The writing in on-line CFs may not be grammatically correct and can also be complex. They often use contrast words in a feedback (Vilares et al., 2017), (Zaw & Tandayya, 2018). For example, the customer feedback “This product seems cheap but it is good quality,” the contrast word ‘but’ helps decrease the negative sentiment of “cheap” and increase the positive sentiment of ‘good quality.’ In fact, D. Vilares and et.al (Vilares et al., 2017) explicated the overall sentiment of such sentence is ‘Positive.’ Based on the scenario, this paper proposes a new algorithm, called the contrast dictionary that can handle complex sentences when extracting the sentiment polarity from CFs.
Considering only positive and negative polarities of the words and applying just the accumulated sum of all sentiments the CFs may not be correct or sufficiently comprehensive. A CF may be constructed in a complex structure and it may contain several attributes or sub-tasks of different sentiments. Analyzing the text structure and exploiting these attributes or aspects would help for the better evaluation of the CFs.