Automobile product ranking based on the singlish comments in social media platforms

Abstract

In today's world, many customers buy or choose products based on online reviews. The internet contains a vast collection of natural language. People share their subjective thoughts and experiences with one another in various social media platforms. Product reviews can be analyzed to determine how people feel about a particular product .In Sri Lanka, people widely use Singlish (Sinhala-English) to comment and give reviews on products, rather than a single pure language .Therefore in this research it has extracted data from social media platforms on various brands in the automobile industry and propose a system to rank the automobile brands in Sri Lanka based on the social media comments which are written on Singlish. When ranking products, it is not practical to rank products based only on the frequency of the products. Because a brand having the highest number of comments does not necessarily indicate that it has good market perception compared to other brands. In order to get an accurate overview, the study have considered both the people's sentiment towards the particular brand and the frequency of comments. When ranking the products research has done several rankings based on different aspects namely market value, country of origin and second hand market, vehicle performance, product features which people pay their most attention in the automobile industry and also an overall ranking considering all these aspects together. With that it is possible to identify which vehicle type or brand has the highest and lowest demand in the market, and the automobile manufacturer can get a good understanding where a particular product stands out comparative to other brands and apply their strategies accordingly. When implementing the ranking system 100000 social media comments were extracted and annotated. Convolutionary neural network was used to develop the main model, and out of the different methods tried to predict the sentiment as the part of the main model, random forest method gave a higher accuracy of 96.7 making it a more sophisticated combination.