| Summary: | 碩士 === 國立臺南大學 === 數位學習科技學系數位學習科技碩士在職專班 === 105 === Huge amounts of information is sent to all kinds of network platforms, and the spread of the Internet from the traditional site operators into users. Online customer reviews are helping manufactures to decide marketing strategies, and consumers used to watch online reviews before they buy products. However, the network information is complex. Consumers search keywords in the use of evaluations that from a large number of links in the page one by one, and then read the article carefully to find out the comments, which is a lengthy process for the consumer. This study present a prototype which automated collect and classify consumer reviews based on multi-attribute products. Through the topic model and sentiment analysis, consumers can quickly extract the results and the product evaluation that do not need to read the entire content.
In this study, 2837 reviews were collected from PTT. Through the topic analysis of LDA to build feature lexicon, the reviews would be automatically classified into marketing and non-marketing category, and then the sentiment polarity would be analyzed by marketing 4C category. Experiment result showed that classified as marketing 4C category the average of accuracy was 93.84% and the average of F1-Measure was 93.08%. For the category of sentiment polarity of consumer reviews the average accuracy was 91.74% and F1-Measure was 90.68%. It was proved that the sentiment polarity of this machine classification has good accuracy, so that consumers have reliable and correct reference when searching for articles without having to spend too much time browsing the contents of the article.
|
|---|
